A review of applications of machine learning in quantum dots research

Abstract

Machine learning (ML) is increasingly applied in quantum dot (QD) research to support data analysis, device control, materials optimization, sensing, and theoretical modeling. This review surveys recent ML-based approaches across experimental, applied, and computational QD studies, with emphasis on how data-driven methods are embedded within established physical workflows rather than treated as standalone solutions. ML techniques are examined in the context of automated device tuning, high-throughput characterization, synthesis parameter exploration, chemical and biological sensing, photonic and optoelectronic device analysis, and reduced-order modeling of interacting quantum systems. In most cases, ML serves to replace time-consuming fitting procedures, guide experimental sampling, or approximate computationally intensive simulations. Common methodological patterns include supervised learning on limited datasets, transfer learning across device instances, and hybrid approaches incorporating physical constraints into model design or training objectives. Recurrent limitations are also identified, including dataset bias, restricted cross-laboratory transferability, lack of standardized benchmarks, and limited treatment of uncertainty. Rather than positioning ML as a standalone solution, this work frames it as a complementary tool whose reliability and scientific value depend on integration with physical insight, experimental design, and validation protocols.

Introduction

Motivation

Quantum dots (QDs) are nanoscale semiconductor structures in which charge carriers are confined in all three spatial dimensions, leading to discrete electronic states whose energies depend on size, composition, and confinement geometry [1–3]. This quantum confinement gives rise to characteristic optical and electronic properties, including size-tunable emission spectra, sharp optical transitions, and extended carrier lifetimes. These features have motivated sustained interest in QDs as model systems for studying quantum confinement as well as for applications in photonics, quantum information processing, sensing, and bioimaging [4, 5].

Traditionally, the study of QDs has relied on a combination of theoretical modeling [6, 7], numerical simulation [8, 9], and controlled experimentation [10, 11]. Electronic structure calculations based on effective-mass, tight-binding, or atomistic approaches have been used to predict energy levels and wavefunctions, while optical and transport measurements provide experimental access to confinement, coherence, and coupling effects. At the device level, fabrication and tuning procedures typically involve repeated manual adjustment of growth parameters, gate voltages, or environmental conditions, guided by physical intuition and incremental feedback from measurements. While these approaches have been successful for relatively simple or well-isolated systems, they become increasingly difficult to scale as QD platforms grow in complexity. Modern QD devices often involve many coupled degrees of freedom, strong sensitivity to disorder, and nonlinear responses to control parameters, making exhaustive exploration of parameter space impractical.

Machine learning (ML) refers to a class of computational methods that infer patterns or functional relationships directly from data, rather than relying on explicitly programmed rules or closed-form physical models [12]. In supervised learning, models are trained on example input–output pairs to approximate complex mappings, while unsupervised and probabilistic methods aim to identify structure, correlations, or uncertainty directly from observations. Importantly, ML does not replace physical modeling; instead, it provides a complementary framework for handling high-dimensional datasets, accelerating predictions, and guiding decision-making when analytical solutions or brute-force simulations are computationally prohibitive.

In the context of quantum dot research, ML methods become relevant precisely because many QD-related problems are data-rich but analytically intractable. Large collections of simulated or experimentally measured spectra, transport traces, or device configurations can be used to train models that rapidly predict electronic or optical properties without repeated numerical diagonalization [13–15]. Data-driven optimization techniques, such as Bayesian optimization, enable efficient navigation of multi-parameter synthesis, growth, or gating spaces to identify operating regimes that satisfy target performance criteria with fewer experimental iterations [16]. In parallel, reinforcement learning and related control strategies allow QD devices with many tunable parameters to be adjusted autonomously, stabilizing desired quantum states or operating points through closed-loop interaction with the experiment [17–19].

Viewed more broadly, the incorporation of ML into QD research reflects a shift from purely forward modeling toward hybrid workflows that combine physical insight with adaptive, data-driven inference. Rather than presupposing complete knowledge of the underlying system, these approaches leverage experimental and simulation data to manage complexity, reduce manual intervention, and explore regimes that would otherwise be difficult to access systematically. This perspective provides the foundation for the thematic organization adopted in this review, which categorizes ML-enabled quantum dot research according to modeling, applications, synthesis, control, and materials discovery.

Papers classification and statistics

A structured framework was developed to categorize scientific articles on quantum dots according to both research focus and methodology. The framework utilizes standard bibliometric metadata, including titles, abstracts, and author-provided keywords, to extract information about each article, loaded from Scopus based on keyword "quantum dots" and "ML". These data are processed using a computational language model gpt−4.1-mini via API [20], which interprets the text and identifies the most relevant research areas for each publication. Articles are assigned to six predefined topics that represent major directions in quantum dot research [21–30]. The first topic, theoretical modeling and simulation, includes studies focused on understanding electronic structure, quantum confinement, and computational modeling of quantum dots. The second topic, photonics and quantum computing applications, covers research on optical devices, single-photon sources, and integration of quantum dots into quantum information systems. The third topic, sensors and application-oriented devices, encompasses work on chemical, biological, and environmental sensing applications. The fourth topic, synthesis, fabrication, and optimization, includes studies on chemical synthesis methods, growth processes, surface engineering, and optimization of material properties [31–33]. The fifth topic, automated tuning, control, and closed-loop optimization, captures research implementing feedback-controlled experiments, autonomous adjustment of device parameters, or self-optimizing systems. The sixth topic, materials discovery and data-driven design, covers work using computational screening, ML, or inverse design to identify new quantum dot compositions and properties. In addition to these domain-specific topics, an orthogonal flag indicates whether a study employs physics-informed ML, in which prior knowledge of physical laws is incorporated into computational models [34–37]. The chosen topics reflect both the technical and application-oriented landscape of quantum dot research, and their definition allows systematic mapping of trends, research clusters, and methodological approaches. Figure 1 summarizes the thematic overlap in quantum dot research by combining representative three-topic Venn diagrams with a global UpSet [38] plot that captures intersection patterns across all identified research areas.

Fig. 1.

Topic overlap across quantum dot research areas. Top row: pairwise and triple overlaps between selected thematic groups. Bottom: global intersection structure across all topics using an UpSet representation. Papers were found from Scopus based on keyword "quantum dots" and "ML" (articles only)

Figure 2a highlights clear regional patterns in publications linking quantum dots and ML. China is the largest contributor with 130 papers (23% of the corpus), followed by the United States with 100 (17%) and India with 57 (10%). Several European states also contribute meaningfully, notably the United Kingdom (35), Germany (23), and Switzerland (18); other active participants include Australia, Iran, and South Korea.

Fig. 2.

a County distribution of publications on QDs and ML research (all types of papers, including conference and articles). b Chronological evolution of ML applications in QD research

The leading positions of China and the United States likely reflect concentrated national investment in nanotechnology, artificial intelligence, and materials science, supported by large-scale funding programs, strategic initiatives, and extensive university–industry networks. India’s standing appears driven by rapid expansion in both AI and nanomaterials research under national technology efforts.

European output remains steady, underpinned by mature research infrastructures and an emphasis on interdisciplinary collaborations often facilitated by EU frameworks. Meanwhile, growing but smaller contributions from countries such as Saudi Arabia, Singapore, and Russia indicate emerging capacity and interest in QD–ML topics.

In recent years, the role of ML in QD research has expanded considerably, reflecting both the maturation of ML methodologies and the growing complexity of QD applications. Early studies primarily leveraged ML to accelerate material discovery and to optimize synthesis conditions for achieving desired optical and structural properties. This was followed by the adoption of ML models to predict functional parameters such as bandgap, charge transport, and device efficiency, enabling more precise performance modeling. More recent developments emphasize automated device tuning, real-time feedback control, and advanced data interpretation in sensing and imaging contexts. At the same time, hybrid approaches combining ML with physics-based simulations and explainable AI are emerging, aiming for both practical automation and fundamental insight.. Figure 2b summarizes this chronological progression and emphasizes the key thematic transitions that have shaped the field.

Aim of the study

This review provides a synthesis the integration of ML into quantum dot research as a methodological development driven by increasing system complexity and data availability. Rather than treating ML as a standalone toolkit, the review frames it as a set of data-driven inference and control strategies that complement established theoretical and experimental approaches. Section 2 describes the experimental platforms, measurement modalities, and data-generation pipelines that underpin ML-based analysis in quantum dot systems, highlighting how instrumentation choices shape the structure and quality of accessible data. Section 3 analyzes how different classes of ML methods—ranging from supervised regression models to probabilistic optimization and reinforcement-based control—are employed to address modeling, tuning, and decision-making tasks across diverse quantum dot platforms. Section 4 synthesizes recurring patterns, methodological constraints, and open challenges that limit generalization and transferability, while also identifying directions in which data-driven approaches appear most compatible with physical modeling. Section 5 concludes by situating ML-assisted quantum dot research within the broader context of computational materials science and experimental automation, and outlines questions that remain unresolved.

Experimental data acquisition and instrumentation

ML approaches targeted at QD materials and devices rely on extensive, high-quality experimental datasets. These data streams are generated by a set of complementary instruments that probe optical, chemical, structural and dynamical properties across multiple length- and time-scales. Careful control of acquisition parameters and comprehensive metadata capture are essential to ensure model reproducibility, enable domain adaptation, and support uncertainty quantification. The following layout alternates descriptive text blocks with representative instrument schematics, providing contextual detail on why each measurement modality is important for ML workflows.

Optical spectroscopy forms the backbone of many QD characterization campaigns [39–41]. Steady-state absorption and photoluminescence (PL) spectra provide direct measures of band-edge energies, Stokes shifts [42], spectral linewidths [43] and ensemble heterogeneity [44]; these features often correlate strongly with device-relevant metrics such as quantum yield and color purity. Time-resolved techniques—ranging from time-correlated single-photon counting (TCSPC) [45] for nanosecond–picosecond lifetimes to streak-camera [46] or transient-absorption methods for ultrafast dynamics—reveal carrier relaxation pathways [47], nonradiative recombination [48] channels and exciton dynamics [49] that are otherwise hidden in steady-state traces. For ML, these optical observables become primary descriptors: spectral fingerprints are treated as high-dimensional vectors for representation learning [50], while lifetimes and dynamics are used as target variables [51] or labels in supervised tasks. Acquisition settings (excitation wavelength and power, monochromator bandwidth, detector gain and integration time) should be recorded as metadata because they influence signal-to-noise and model generalization. A fluorimeter setup [52] shown in Fig. 3.

Fig. 3.

Schematic of a fluorimeter configured for both steady-state and time-resolved photoluminescence measurements. Systems typically combine pulsed and continuous excitation sources, dispersive elements or filters, and time-resolved detection to generate spectral and kinetic datasets for ML applications

Chemical- and surface-sensitive spectroscopies complement optical data by revealing ligand composition [53, 54], binding modes and surface passivation quality—factors that strongly affect optical stability and charge trapping. Fourier-transform infrared spectroscopy (FT–IR) and Raman spectroscopy capture vibrational signatures of surface ligands and matrix interactions, while X-ray photoelectron spectroscopy (XPS) quantifies elemental composition and oxidation states. These measurements are frequently used to engineer descriptor vectors that encode surface chemistry: peak positions, relative intensities, and deconvolved component areas are fed into models that predict stability, trap-state density or emission shifts. Combined with optical features, these chemical descriptors improve model interpretability and enable inverse-design loops where synthesis conditions are adjusted to optimize target photophysical outcomes. The FT-IR spectrometer setup is shown in Fig. 4.

Fig. 4.

FT–IR spectrometer layout for probing vibrational modes of QD ligands and matrices. Spectra inform descriptors related to passivation, ligand exchange, and surface contamination that are important inputs for predictive models

Controlled sample delivery and automated handling underpin reproducible high-throughput experiments. Microfluidic reactors and flow cells enable fine control of mixing, reaction time and concentration gradients, producing systematic datasets for synthesis optimization and active-learning experiments. Precision pumps (syringe, peristaltic or pressure-driven controllers) maintain stable flow rates over wide dynamic ranges, while autosamplers and robotic platforms permit combinatorial screening of ligand ratios, precursor concentrations or temperature. From an ML perspective, such automation reduces human-induced variance, increases dataset size and supports closed-loop optimization: models recommend next experiments, automated hardware executes them, and new measurements update the model iteratively. Accurate logging of flow conditions, residence times, and mixing geometries is therefore a critical part of the training corpus. Figure 5 illustrates the use of a peristaltic pump [55].

Fig. 5.

Peristaltic pump for controlled delivery of colloidal suspensions in flow-based synthesis and measurement systems. Stable fluid handling is key to reproducibility in high-throughput ML-driven workflows

Low-temperature and structural probes add orthogonal information that is often decisive for device engineering. Cryogenic PL and magneto-optical measurements resolve fine-structure splitting, spin dynamics and coherence properties; access to temperatures from a few kelvin to ambient allows separation of phonon-limited processes from intrinsic radiative behavior. A cryostat setup [56, 57], shown in Fig. 6.

Fig. 6.

Cryogenic measurement setup for temperature-dependent optical characterization. Variable-temperature studies enable the extraction of coherence-related metrics and temperature-dependent emission behavior used in ML models

Structural techniques such as powder and grazing-incidence X-ray diffraction (XRD), complemented by electron microscopy (TEM/SEM), determine crystalline phase, average particle size and lattice strain—attributes that map to optical inhomogeneity and carrier localization. For ML-driven materials design, incorporating these orthogonal data types into multimodal models improves robustness and enables richer inverse-design objectives, for example optimizing synthesis to target a specific crystal phase with narrow PL linewidths at low temperature.

X-ray diffraction (XRD) provides indispensable structural metrics for quantum-dot (QD) materials and is therefore a key input modality for machine-learning workflows in materials characterization and design. Diffraction experiments—performed in powder, thin-film, and grazing-incidence geometries—deliver peak positions, integrated intensities and profile shapes that map directly to lattice parameters, phase composition, average coherent domain size (Scherrer analysis), and microstrain (Williamson–Hall or profile-fitting approaches). High-resolution scans using Cu K (or alternative sealed-tube/monochromated sources and synchrotron radiation) with fine 2 step increments and long counting times improve peak centroiding and enable accurate lattice-parameter extraction and strain analysis; reciprocal-space mapping and asymmetric scans further resolve texture and anisotropic strain in films.

For ML applications, properly preprocessed diffraction data yield compact, physically meaningful features: peak positions and shifts (unit-cell changes), full-width-at-half-maximum (FWHM) and integrated area (size/strain proxies), relative phase fractions from Rietveld or Le Bail fits, and higher-order descriptors such as peak asymmetry, background-corrected intensity ratios, or parameters from pair-distribution-function (PDF) analysis for disordered samples. Instrumental contributions (wavelength, divergence, slit settings, goniometer geometry, and instrumental broadening) must be recorded as metadata because they influence profile shape and thus model transferability. Simulated patterns—generated from ab initio or empirical structures—are also valuable for augmenting training sets and for supervised tasks that predict synthesis-to-structure mappings.

In practice, XRD complements real-space imaging (TEM/SEM) and spectroscopic probes to form a multimodal dataset: structural labels derived from diffraction enable supervised regression and classification targets (e.g., crystalline phase, mean size, strain state), while continuous-valued descriptors support inverse-design and optimization loops that adjust synthesis parameters to reach an intended structural signature. Robust preprocessing (background subtraction, smoothing, peak deconvolution, and normalization) and explicit uncertainty estimates for fitted parameters improve model reliability and enable uncertainty-aware algorithms such as Gaussian processes or Bayesian neural networks to guide experimental decision-making (Fig. 7).

Fig. 7.

Schematic of an X-ray diffractometer for structural analysis of QD powders and thin films. Typical measurements include powder scans, grazing-incidence XRD for films, and reciprocal-space mapping for texture and anisotropic strain characterization; extracted features—peak positions, widths and intensities—are widely used as inputs or labels for ML models

ML applications in QD

Scientific landscape

Recent research at the intersection of quantum dot (QD) science and ML spans a broad range of physical platforms, modeling paradigms, and application domains. A recurring theme across this literature is the use of data-driven methods to complement or accelerate established theoretical, numerical, and experimental workflows, particularly in regimes characterized by high-dimensional parameter spaces, nonlinear responses, or limited analytical tractability.

A first class of studies focuses on theoretical modeling and property prediction, where ML is employed to approximate quantum-mechanical quantities that are traditionally obtained through computationally intensive simulations. Neural-network-assisted variational Monte Carlo approaches have been shown to improve ground-state energy estimation and nodal structure optimization in interacting fermionic systems, including multi-electron quantum dots, reducing variational bias relative to fixed-node diffusion Monte Carlo methods [58]. Similarly, transfer-learning strategies combined with atomistic tight-binding calculations enable accurate prediction of single-particle energy levels in double nanowire quantum dots using only sparse training data [59]. In optical modeling, ML regressors such as random forests and decision trees have been demonstrated to reproduce field- and temperature-dependent nonlinear optical rectification coefficients in core–shell and tetrapod quantum dots with high fidelity when trained on numerically generated datasets [60–62]. These works collectively illustrate how ML models can act as efficient surrogates for repeated quantum-mechanical calculations while preserving sensitivity to structural and environmental parameters.

A closely related direction concerns materials discovery and design rule extraction, where high-throughput simulations are combined with statistical learning to identify structure–property relationships. High-throughput TDDFT studies of graphene quantum dots, coupled with supervised learning, have revealed systematic trends linking dopant type, concentration, and geometry to emission wavelength tuning. In silicon carbide and zinc oxide–based quantum dots, correlational analysis and ML-assisted descriptor screening have been used to relate molecular connectivity and adsorption configurations to photovoltaic or sensing-relevant observables [36]. Similar strategies have been adopted for catalytic applications, where Bayesian genetic algorithms guide the selection of transition-metal-incorporated carbon quantum dots for hydrogen evolution, with experimental validation confirming ML-guided predictions [63]. These studies emphasize the role of ML not only as a predictive tool, but also as a means of extracting interpretable design principles from large parameter spaces.

Another substantial body of work addresses device-level optimization and sensing applications, where ML models are integrated with experimental measurements or numerical device simulations. In optoelectronic devices, ensemble learning and gradient-boosting models have been used to optimize organic photodetectors enhanced with quantum dots by identifying critical thickness and transport-layer parameters [64]. For photovoltaic systems based on quantum dots and nanowires, artificial neural networks have been employed to refine analytical Lambert-function models and predict current–voltage characteristics under varying operating conditions. Machine-learning-enhanced sensors incorporating perovskite or carbon quantum dots have also been demonstrated for real-time monitoring tasks, including lithium-ion battery state prediction and chemical pollutant removal, where recurrent or multitask learning models improve temporal forecasting accuracy. These works highlight how ML can bridge physical signal generation and system-level decision-making.

A distinct but increasingly relevant research line explores automated control, optimization, and learning in quantum systems, extending beyond classical ML toward quantum-aware or hybrid paradigms. Reinforcement-learning and evolutionary algorithms have been applied to explore quantum control landscapes, revealing differences in solution-space structure and algorithmic performance depending on reward design and dimensionality reduction strategies [65]. At a more fundamental level, learning-theoretic studies analyze the sample and computational complexity of inferring properties of parametrized quantum circuits, proposing kernel-based methods to balance accuracy and tractability [66]. Related efforts in quantum ML introduce training schemes rooted in physical principles, such as equilibrium propagation derived from Onsager reciprocity, enabling gradient extraction directly from quantum systems [67]. Noise-aware learning frameworks tailored to spin-based quantum dot platforms further demonstrate how pulse-level modeling and ML training can be co-designed to reflect realistic hardware constraints.

Physics-informed and interpretable ML are frequently discussed directions in quantum dot research, yet most existing studies rely on black-box architectures focused on predictive accuracy. Deep neural networks are widely applied to approximate electronic spectra [68, 69], emission energies, and transport characteristics, but the resulting representations are seldom examined in relation to effective Hamiltonians, confinement potentials, or material-dependent scaling relations. Consequently, model outputs are often weakly connected to physically meaningful quantities such as effective masses, Coulomb interaction parameters, or symmetry-driven selection rules.

Physics-informed formulations incorporate constraints such as conservation laws, boundary conditions, or Schrödinger–Poisson consistency into architectures or loss functions, for example by enforcing size-dependent energy scaling or embedding k·p and tight-binding priors. Interpretable and hybrid approaches, including sparse models and physically constrained neural networks, allow analysis of geometric and electrostatic control parameters that govern level structure, coherence decay, and charge stability across different quantum dot platforms. Similar combinations of first-principles modeling and ML regression appear in studies of contact resistance engineering, where symbolic regression links learned models to physically meaningful quantities such as Schottky barrier height and tunneling resistivity. These approaches illustrate a shift away from purely black-box models toward architectures that encode domain knowledge and facilitate generalization across systems. Neural network potentials combined with density functional theory have enabled first-principles prediction of luminescence spectra in real-scale core–shell quantum dots containing thousands of atoms, accurately capturing exciton–phonon coupling effects and reproducing experimental line shapes [70]. In photonic systems, Cui et al. proposed a MAP–MLP hybrid framework for spectral reconstruction that couples Bayesian physical denoising with lightweight neural refinement, achieving robustness against noise and device-to-device variability and demonstrating extensibility to quantum-dot-based spectrometers [71]. Similar principles underpin physics-aware ML approaches for quantum device characterization, where differentiable master equation solvers enable gradient-based parameter estimation and Bayesian inference for single and double quantum dot transport models [72]. These hybrid strategies emphasize interpretability and generalization by explicitly encoding physical structure into the learning process.

Existing literature demonstrates that ML in quantum dot research serves multiple roles: surrogate modeling of quantum-mechanical properties, extraction of structure–property relationships, optimization of synthesis and device parameters, and automation of control and decision-making processes. While methodological diversity remains high and generalization across platforms is often limited, the convergence toward hybrid, physics-aware learning frameworks suggests a common direction for future work.

QD autotuning and quantum computing integration

The following examples illustrate how ML techniques are being applied across multiple facets of QD research—from real-time autotuning and explainable charge-state classification to large-scale device integration, minimal-sweep tuning, atomistic modeling of surface ligands, and real-time growth control. Each study leverages tailored neural-network architectures, advanced feature-extraction methods, or force-field approaches to automate complex workflows, accelerate parameter optimization, and enhance interpretability, while also identifying remaining challenges such as measurement speed, dataset generalization, and system scalability.

Yon et al. [73] demonstrated online charge-state autotuning of a single QD using a convolutional neural network (CNN) integrated into a closed-loop system. The CNN was trained offline on 33,429 labeled patches drawn from nine annotated stability diagrams and detects charge-transition lines within 1818 voltage patches. In 20 experimental runs, the system located the one-electron regime in 19 out of 20 cases (95% success rate). Each run began at a random starting gate voltage within and ; each exploration step scanned one patch of 324 measurements and took 67 s. With an average of 110 steps per run, the mean tuning duration was 2 h 9 min ( min), with measurement time comprising 96% of total runtime—compared to roughly 7 h for a full 2D stability-diagram scan of the same area. The CNN’s classification was further enhanced via an uncertainty-based confidence score that triggers manual verification when needed. Although the sole failure was due to exploration logic rather than CNN misclassification, the 95% success rate (versus 78% offline) highlights the power of real-time autotuning. To drive this approach toward practical scalability, measurement speed must be optimized—e.g., by adopting compressed-sensing strategies:

1

or sparse, adaptive sampling—and model uncertainty could be quantified more rigorously through Bayesian CNNs or Monte Carlo Dropout. Reinforcement-learning agents might also learn optimal voltage-stepping policies, reducing reliance on fixed exploration rules.

The need for efficient, automated tuning and characterization has spurred work on interpretable analysis of triangle plots—images of current flow that encode critical tuning states. Traditional CNN-based classifiers deliver high accuracy but little transparency. Schug et al. [74] proposed using Gabor-wavelet transforms for feature-extraction, yet a more physically grounded approach employs synthetic triangle modeling. In this method, each plot is approximated by a sum of sigmoid functions,

2

where . Parameters are optimized to minimize a similarity metric combining gradient- and identity-based terms. The resulting feature vectors map directly to physical channel edges, enabling classification with Explainable Boosting Machines (EBMs). This yields accuracy on par with CNNs but with per-feature interpretability via Shapley-value or Integrated-Gradients attribution. Future work could extend this framework to time-resolved triangle data—capturing dynamic instabilities—and incorporate inverse modeling to infer underlying electrostatic potentials.

Thomas et al. [75] report the integration of 1,024 independent silicon QD devices with on-chip digital and analog electronics, all operating below 1 K. A cryogenic high-frequency multiplexer (MUX) addresses a 3232 device array via ten digital address lines and three analog control lines. Full characterization and analysis of all devices completes in under 10 min, achieving a signal-to-noise ratio (SNR) above 75 with a per-point integration time of 3.18 s. The minimum integration time for SNR=1 was measured at 556 ps ( ps), yielding an average SNR bandwidth of 6.4 MHz across nine devices. Automated ML routines—principally a CNN classifier—categorize device behavior into “clear Coulomb blockade,” “no blockade,” or “multiple QDs,” facilitating rapid yield assessment and feedback into device design. To further boost throughput and robustness, graph-based neural networks could exploit spatial correlations across the array, and transfer learning could adapt models to new fabrication runs without wholesale retraining.

Semiconductor QD array scalability is often limited by the extensive manual tuning of potential landscapes. Kashtiban et al. [76] introduced a minimal-sweep approach for a GaAs quadruple-QD device, inspired by the Ray-based Classification framework. Instead of full 2D scans, they perform one-dimensional voltage sweeps at multiple barrier-voltage settings, collecting 500 current traces (200 points each). Of these, 250 traces exhibit Coulomb peaks and 250 do not. An LSTM network is trained (400 traces) and tested (100 traces) to detect single-electron transport features. Successful tuning of all four QDs confirms that a few intelligently chosen 1D sweeps suffice for pinch-off voltage identification. Attention mechanisms could further highlight the most informative sweep segments, and active learning could reduce the number of required traces while maintaining reliability. This method paves the way toward fully automated, scalable QD-array control.

At the atomistic scale, ligand geometries on colloidal QDs dictate electronic structure and exciton dynamics. Zhang et al. [53] developed a machine-learning force field (MLFF) using the DeePMD framework, training on density-functional theory (DFT) and atom-centered density matrix propagation (ADMP) data for small CdSe QDs with various ligand passivations (e.g., butyrate). The total energy is decomposed as

3

where denotes the local atomic environment and is a neural network. Active-learning strategies and dataset expansion enable MD simulations of large QDs (e.g., and ) over 30 ps, tracking transitions among bridging, tilted, chelating, and claw geometries. Markov-state modeling reveals that bridging motifs dominate 100 facets while claw motifs favor 111 facets—information vital for designing ligands that tune surface states and carrier localization. Incorporating equivariant neural networks (e.g., NequIP) could further improve rotational invariance, and experimental validation via spectroscopic signatures could close the loop between simulation and measurement.

Finally, Shen et al. [77] tackled the multidimensional optimization of quantum-dot density during molecular-beam epitaxy (MBE) by analyzing real-time reflection high-energy electron diffraction (RHEED) videos. A 3D ResNet-50 model was trained on 120 videos (30 samples4 repeats) labeled as “zero,” “low,” “middle,” or “high” density, achieving 94.4% accuracy for QD formation and 95.1% for density classification. By correlating streak-to-spot transitions in RHEED patterns with density outcomes, this closed-loop system adjusts growth parameters on the fly, producing densities from   to   in near-real time. To reduce dependence on labeled data, self-supervised pretraining on unlabeled RHEED streams could be introduced, and spatiotemporal generative models (e.g., GANs) might predict morphological evolution, offering preemptive control of dot nucleation.

Table 1 summarizes key studies on ML-driven quantum-dot autotuning and large-scale QD integration. It highlights each work’s focus, algorithmic approach, dataset size, main outcomes, practical applications, and identified limitations or directions for future research.

Table 1.

Summary of Studies on QD Autotuning and Quantum Computing Integration

Study	Focus area	ML method	Dataset	Key results	Application/ outcome	Limitations/ future work
Yon et al. [73]	Real-time charge state autotuning	CNN + feedback loop	33,429 labeled patches from 9 diagrams	95% success in locating 1-electron regime, avg. tuning time 2 h 9min	Autonomous calibration of QDs in real time	Speed bottlenecks due to measurement time
Schug et al. [74]	Explainable classification of triangle plots	Synthetic image vectorization + EBM	Synthetic + experimental triangle plots	Interpretable features matching CNN accuracy	Robust QD classification with physical transparency	Needs further validation on larger experimental datasets
Thomas et al. [75]	Scalable read-out from 1024 QD devices	CNN + automated analysis	Data from 1024 QDs with high-speed readout	Classification of QD states in <10 min, SNR >75	Scalable integration and yield assessment of silicon QDs	Generalizability to other QD architectures not addressed
Kashtiban et al. [76]	Minimal-sweep autotuning of QDs	LSTM (current traces)	500 current traces (200 pts each)	Accurate tuning of 4 QDs using 1D sweeps	Efficient multi-dot tuning in GaAs systems	Needs exploration for more complex QD topologies
Zhang et al. [53]	Ligand geometry modeling on colloidal QDs	ML force fields (DeePMD) + MSM	MD + DFT data on  CdSe QDs	Identified ligand dynamics across geometries	Surface behavior prediction of large-scale QDs	Applicable to more QD types but requires further transferability study
Shen et al. [77]	Real-time density control in MBE QD growth	3D ResNet on RHEED videos	120 RHEED videos (30 QD samples 4)	95.1% density classification accuracy	Closed-loop density control during QD growth	Requires integration with broader MBE systems

These studies indicate that ML approaches based on raw measurement images or time traces, such as CNNs and LSTMs, achieve high performance but remain strongly system-specific, as their learned features depend on device geometry, noise statistics, and acquisition protocols. Methods that operate on physically structured representations—such as synthetic triangle models, minimal-sweep transport features, or graph-based abstractions of gate layouts—exhibit stronger potential for generalization across QD platforms, although they may sacrifice some peak accuracy. Data requirements scale accordingly: high-capacity deep networks demand large, consistently labeled datasets, while physics-informed or feature-engineered models reduce data needs by encoding prior knowledge but apply only within regimes where those assumptions hold. At the atomistic scale, machine-learning force fields generalize within narrow chemical spaces yet require retraining for new ligands or core compositions due to locality of atomic environments. Interpretable models enable direct association between ML outputs and electrostatic, geometric, or chemical parameters, supporting diagnostic analysis and cross-device comparison, whereas black-box models favor throughput and automation. Hybrid frameworks that combine physically meaningful intermediate representations with lightweight neural components offer a balanced compromise between performance, data efficiency, interpretability, and scalability across heterogeneous QD systems.

ML for QD synthesis and property optimization

The synthesis of QDs often involves complex multivariable processes, leading to high uncertainty and extensive trial-and-error. ML offers an efficient approach to guide the synthesis process, optimizing material properties with fewer experiments. Recent studies demonstrate how ML techniques can enhance the quantum yield (QY) [78, 79] and other properties of QDs [79, 80], facilitating their use in applications ranging from environmental sensing to quantum computing.

Synthesis of functional nanostructures with the least number of tests is paramount towards the propelling materials development. However, the synthesis method containing multivariable leads to high uncertainty, exhaustive attempts, and exorbitant manpower costs. ML burgeons and provokes an interest in rationally designing and synthesizing materials. Tang et al. [81] collect the dataset of nano-functional materials carbon dots (CDs) on synthetic parameters and optical properties. ML is applied to assist the synthesis process to enhance photoluminescence quantum yield by building the methodology named active adaptive method (AAM) [82], including the model selection, max points screen, and experimental verification. An interactive iteration strategy is the first time considered in AAM with the constant acquisition of the furnished data by itself to perfect the model. CDs exhibit a strong red emission with QY up to 23.3% and enhancement of around 200% compared with the pristine value obtained through the AAM guidance. Furthermore, the guided CDs are applied as metal ions probes for and , with a concentration range of 0–120 and 0–150 M, and their detection limits are 1.17 and 0.06 M. CDs are also applied for dental diagnosis and treatment using excellent optical ability. It can effectively detect early caries and treat mineralization combined with gel. The study shows that the error of experiment verification gradually decreases and QY improves double with the effective feedback loops by AAM, suggesting the great potential of utilizing ML to guide the synthesis of novel materials. Finally, the code is opensource and provided to be referenced for further investigation on the novel inorganic material prediction.

Carbon QDs (CQDs) are photoluminescent carbon nanomaterials with fast response times and excellent optical sensing properties, widely used in industrial and biomedical research. The main challenge in CQD synthesis is achieving a high QY, which is crucial for high sensitivity and applicability. Kannouma et al. [78] investigates the use of ML models to optimize the production of high-QY CQDs. A dataset of 117 data points, collected from experiments on QY and CQD synthesis, was used to train several ML models. The Multilayer Perceptron (MLP) regression model was selected for its low mean absolute error, root mean square error, and high coefficient of determination. Using this model, the optimal synthesis conditions were determined, resulting in a CQD with a QY of 79%. CQDs were then applied as ultra-sensitive luminescent sensors for ion detection, showing a detection limit of 302 nM and linear response over a concentration range of 1.00–200.00 M. The study highlights the use of ML for controlled synthesis and optimization of CQDs, reducing the resource-intensive trial-and-error approach. The effectiveness of the MLP model in predicting QY demonstrates the potential for faster development of high-performance CQDs. Furthermore, the study successfully demonstrates the application of these CQDs for ion detection in environmental and industrial applications.

Image processing, employs techniques such as filtering, segmentation, and ML to extract information from digital images. Quantum-dot Cellular Automata (QCA) emerges as a nano-scale alternative to CMOS for image-processing circuits, despite challenges like fabrication defects. In response, a novel QCA-based morphological operation circuit was developed by Tang et al. [83], designed for both erosion and dilation using a single-bit input, with six inputs (A1–A5, A) and one output (O). The circuit consists of 107 QCA cells and demonstrates a low delay of approximately 0.75 clock phases. Power analysis using QCADesigner-E shows total and average energy dissipations of eV and eV, respectively. Fault tolerance is enhanced through the integration of a five-input majority gate and redundancy strategies, showing resilience against single-cell defects. Experimental results indicate that even when half the cells are randomly faulted, the circuit maintains a high proportion of correct outputs. Compared to earlier designs, the proposed structure achieves superior results in quantum cost, area efficiency, and latency. Table-based comparisons confirm its advantage, using a performance metric of area to benchmark improvements. The architecture paves the way for highly reliable image processors capable of adaptive morphological processing in complex visual tasks.

PbS colloidal QDs (CQDs) have applications in short-wave infrared (SWIR) detection due to their wide tunable bandgap, low thermoelectric noise, and solution processing capabilities. The exciton peak of QDs determines the response band of the detector, and well-monodispersed QDs often exhibit better optical performance in photodetectors. The detection performance of PbS CQD-based SWIR photodetectors is closely linked to the synthetic properties of QDs in the active layer. ML has accelerated the exploration of QDs synthesis processes, where a neural network (NN) model is developed by Xu et al. [84] to predict the exciton peak and peak/valley ratio based on experimental data. In terms of model performance, the NN model achieved a correlation coefficient of 0.93 for exciton peak prediction, and 0.75 for the peak/valley ratio. The prediction error for the exciton peak was only 3.89%, while the peak/valley ratio error was 7.24%. The synthesized CQDs, with a peak/valley ratio of 3.105, were used in SWIR photoconductive devices, yielding a responsivity of 2.53 A/W, a detectivity of 2.08 10¹² Jones, and a noise current of 7.81 10^{− 13} A/.

Polyethylene terephthalate-derived fluorescent carbon QDs (PET-FCQDs), sized nm, were synthesized by Enyoh [85]via a one-step pyrolysis method from PET waste and effectively removed fluoxetine (FLX; 100–400 ng/L) and ciprofloxacin (CIP; 50–150 g/L) from water. Using Box-Behnken Design and ML models (Artificial Neural Networks, ANN, and Support Vector Machines, SVM), optimal removal efficiencies of 95.19% for FLX and 97.85% for CIP were achieved, with ANN demonstrating superior prediction accuracy (–0.88). The adsorption behavior followed Langmuir () and Freundlich isotherms (), with kinetics consistent with intraparticle diffusion (). Maximum adsorption capacities reached 705 ng/g for CIP and 62.27 ng/g for FLX. ATR-FTIR and molecular dynamics analyses identified – stacking, hydrogen bonding, and electrostatic interactions as primary adsorption mechanisms. Molecular docking revealed high binding affinity of PET-FCQDs to the serotonin transporter (SERT; kcal/mol) and topoisomerase IV ( kcal/mol), suggesting their intrinsic antidepressant and antibiotic potential. ADMET analysis confirmed high gastrointestinal absorption, blood-brain barrier permeability, non-carcinogenicity, and non-biodegradability. However, PET-FCQDs showed high toxicity to ecological organisms such as fish, honeybees, and Tetrahymena pyriformis, raising concerns about environmental safety. The process achieved an 85% yield by heating 1 g of PET at 300 °C for 30 min, highlighting a scalable and sustainable method for converting plastic waste into multifunctional nanomaterials with applications in both environmental remediation and biomedical fields.

Majorana zero modes, promising for fault-tolerant quantum computing, are expected at InAs/Al interfaces, but strong hybridization from direct contact suppresses their emergence. Using DFT+U with Bayesian optimization (BO), Jardine et al. [86] investigate ZnTe and CdSe as tunnel barriers to mitigate this coupling. BO-optimized Hubbard U values of 9.4 eV (Zn) and 8.3 eV (Cd) were applied to reproduce HSE band structures, yielding PBE+U(BO) band gaps of 1.48 eV (ZnTe) and 0.96 eV (CdSe), in better agreement with experimental values (2.26–2.38 eV and 1.75 eV, respectively). The lattice constants used were 6.0584 Å for InAs, 6.101 Å for ZnTe, and 6.077 Å for CdSe. Slab models showed that 40 atomic layers are required to converge the ZnTe/CdSe band gaps, and 16 layers suffice to suppress metal-induced gap states (MIGS) in InAs. Band alignment calculations show that ZnTe provides a conduction-band barrier for electrons, while CdSe acts as a valence-band barrier for holes. Due to Fermi level pinning near the InAs conduction band, ZnTe is the more effective barrier for suppressing unwanted electron tunneling. Optimal interfacial distances were found to be 2.337 Å (InAs/Al), 2.465 Å (ZnTe/Al), and 2.11 Å (InAs/ZnTe), ensuring structural realism. Based on these results, authors recommend fabricating InAs/ZnTe/Al trilayer devices with ZnTe thicknesses between 6 and 18 atomic layers to achieve sufficient isolation for Majorana applications.

Table 2 provides an overview of recent ML applications aimed at guiding the synthesis, optimization, and functional design of QDs and related nanostructures. The studies span a range of materials and objectives—from improving photoluminescence yields to optimizing device-level performance—highlighting how diverse ML strategies can accelerate experimental workflows and enable targeted property tuning.

Table 2.

Summary of ML Applications in QD Synthesis and Optimization

Study	Target material/ system	ML approach	Goal/ optimization	Results/ performance	Application	Highlights
Tang et al. [81]	Carbon Dots (CDs)	Active Adaptive Method (AAM)	Enhance photoluminescence QY	QY up to 23.3% (200% enhancement); reduced experimental error	Metal ion detection (, ); dental diagnostics	Iterative self-learning; first-time application of AAM; open-source code
Kannouma et al. [78]	Carbon QDs (CQDs)	MLP regression	Maximize CQD QY	QY achieved: 79%; detection limit: 302 nM	Environmental/ industrial ion sensors	High model accuracy; reduced trial-and-error synthesis
Tang et al. [83]	QCA-based morphological image processor	ML + QCA design	Efficient erosion/ dilation circuit for image processing	Delay: 0.75 clock phases; avg. energy: eV	Nano-scale morphological processing	High fault tolerance; compact design; benchmarked superiority
Xu et al. [84]	PbS CQDs for SWIR devices	Neural Network (NN)	Predict exciton peak and peak/valley ratio	= 0.93 (peak), 0.75 (ratio); Responsivity: 2.53 A/W	SWIR photodetectors	NN-guided synthesis; high detectivity; low noise current
Enyoh [85]	PET-derived CQDs	ANN, SVM + Box-Behnken Design	Optimize pollutant removal (FLX, CIP)	Removal: 95.19% (FLX), 97.85% (CIP); Adsorption: 705 ng/g (CIP)	Environmental remediation; biomedical applications	Drug-binding insight; adsorption modeling; noted ecotoxicity
Jardine et al. [86]	InAs/ ZnTe/ CdSe heterostructures	Bayesian Optimization with DFT+U	Optimal tunnel barrier for Majorana zero modes	ZnTe band gap: 1.48 eV; barrier thickness: 6–18 layers	Quantum computing (Majorana devices)	ZnTe as best barrier; reduced electron tunneling; realistic interfaces

Across these studies on synthesis, materials design, and functional nanostructures, distinct patterns emerge in how ML generalizes and where it remains system-specific. Data-driven regression and neural-network models used for synthesis optimization (e.g., CQDs, PbS CQDs, PET-derived carbon dots) generalize reasonably within narrowly defined chemical spaces, where precursor types, reaction conditions, and target properties remain consistent, but require retraining or recalibration when extended to new chemistries or synthesis routes. Active and Bayesian optimization frameworks show stronger transferability than static predictors, as they adaptively update models based on feedback and can operate under sparse, noisy experimental data. In contrast, models tightly coupled to specific physical formalisms—such as ML force fields, DFT+U calibration, or QCA circuit architectures—are highly system-specific, reflecting their dependence on atomic species, lattice structure, or device geometry. Trade-offs are evident: high-capacity neural networks offer accurate property prediction but demand curated datasets and provide limited physical insight, whereas physics-guided optimization and surrogate models reduce data requirements and align outputs with interpretable parameters at the cost of narrower applicability. Scalability is most favorable for approaches integrating ML into closed-loop or design-of-experiments pipelines, where model complexity is balanced against experimental throughput.

Sensors, biomedical, and environmental applications

The combination of ML with sensing platforms can improve the speed, sensitivity, and portability of analyte detection. These approaches address the limitations of traditional laboratory methods and open opportunities for practical, on-site applications. The following section highlights selected examples where ML has been successfully integrated into field-deployable sensing systems.

User-friendly in-field sensing protocols are essential for tracing analytes in resource-limited environments. Existing sensing methods, requiring professional technicians and expensive laboratory instruments, are not suitable for point-of-care analyses. To overcome this, an artificial intelligence (AI) handheld sensor was developed by Yan et al. [87] for the direct detection of and EDTA in food samples. The sensor integrates a smartphone with a ML application, a 3D-printed handheld device, and a cellulose paper microfluidic chip stained with ratiometric red-green-emission carbon dots (CDs). The system enables a continuous fluorescence (FL) [88–91] color transition from red to green upon introduction, followed by a return from green to red upon EDTA addition, is quantified by analyzing the intensity ratio:

4

where and are the measured emission intensities. Linear calibration curves were obtained:

5

with detection limits calculated by

6

where is the standard deviation of blank measurements and S is the slope of the calibration curve. Ranges of 0–48 M () and 0–96 M (EDTA) yielded LODs of 0.274 M and 0.624 M, respectively. Recoveries between 98.47% and 106.30% in water, milk, spinach, bread, and shampoo validated real-sample performance.

This sensor provides linear ranges of 0–48 M for and 0–96 M for EDTA with detection limits of 0.274 M and 0.624 M, respectively. The detection protocol enables quick, on-site determination of target analytes in various samples, including water, milk, spinach, bread, and shampoo. The system’s performance was validated with recoveries of and EDTA ranging from 98.47% to 106.30% in real samples, demonstrating the feasibility of the AI-based handheld sensor for user-friendly quantification in field settings. The proposed sensor is a promising tool for environmental monitoring, food safety, and healthcare applications, enabling automated, real-time analysis without the need for specialized equipment or expertise.

Improvement ideas may include incorporate cross-validation with k-fold (e.g., ) to guard against overfitting in the regression model; perform a Design of Experiments (DoE) study to optimize chip geometry and reagent concentrations; integrate on-device calibration via embedded digital twin models to adjust for environmental variability (temperature, humidity).

Serum creatinine (CRT) is a biomarker for diagnosing and monitoring renal diseases. Geethukrishnan et al. [92] demonstrate an electrochemical sensor for CRT detection using copper nanowires (CuNW) and molybdenum disulfide QDs (MSQD) on a glassy carbon electrode (GCE). The sensor’s performance was evaluated through cyclic voltammetry (CV) and differential pulse voltammetry (DPV), with a linear response observed from 1.96 M to 966.0 M CRT. The limit of detection (LOD) was found to be 2.3 M in complex mixtures and 0.001 M in urine samples, with root mean square errors (RMSE) of 0.2 M and 0.017 M using artificial neural networks (ANN) and random forest (RF) ML models, respectively. The electrochemical sensor exhibited high sensitivity, with an effective surface area of 17.2 10⁻³ cm², an improvement over the bare GCE (9.1 10⁻³ cm²). Furthermore, the ML algorithms enhanced the sensor’s ability to handle interfering species, achieving an R² of 0.830 for ANN, 0.929 for RF, and 0.764 for k-nearest neighbors (KNN). This sensor offers potential for point-of-care applications in renal disease management.

With increasing awareness of food safety issues, the prompt detection and control of foodborne pathogens have become an important aspect of public health research. Zhang et al. [93] reported a machine-learning-assisted fluorescent sensing system based on aqueous perovskite quantum dots (PQDs) for identifying and inactivating foodborne pathogens. Changes in relative fluorescence signals (RGB) from the sensor array were analyzed using a Support Vector Machine (SVM) model. The system accurately distinguished five pathogens and their mixtures across concentrations from 1.0 to 1.0 CFU/mL, achieving 100% classification accuracy with low detection limits. It also demonstrated reliable identification in tap water samples and inactivated more than 99% of the pathogens within 30 min after detection. These findings indicate that the developed platform could serve as an effective tool for food safety monitoring.

Bacterial contamination poses serious risks in medicine, environment, food, and agriculture, necessitating a platform for rapid detection and inactivation. A paper-based fluorescence sensor array was developed by Zhu et al. [94] using Ag-, Cu-, and Zn-doped carbon QDs (CQDs) printed on filter paper. The fluorescence of the CQDs decreases in the presence of bacteria due to aggregation-caused quenching, with detection enabled by smartphone imaging and ML analysis. This platform successfully identifies five bacterial species—E. coli, S. aureus, P. aeruginosa, S. typhimurium, and L. monocytogenes—across a concentration range of to CFU/mL. Each CQD type shows unique emission peaks at 509 nm (Ag-CQDs), 443 nm (Cu-CQDs), and 554 nm (Zn-CQDs), under specific excitation wavelengths. Antibacterial efficacy tests demonstrated strong bactericidal effects, with Ag-CQDs killing 99.99% of E. coli and 99.92% of S. aureus within 30 min. Cu-CQDs and Zn-CQDs also showed substantial activity, reducing E. coli by 87.2% and 74.8%, and S. aureus by 86.1% and 68.9%, respectively. Live/dead staining confirmed cell membrane damage post-treatment, with increased red fluorescence indicating bacterial death. The platform operates without additional treatment, combining bacterial recognition and inactivation simultaneously. It offers a low-cost, portable, and highly integrated approach suitable for real-world applications in multiple fields.

AI was employed by Ersöz et al. [95] to predict the temperature-dependent current–voltage (I–V) behavior of a Schottky diode fabricated from lanthanum-doped polyethyleneimine-functionalized graphene QDs (La-doped PEI-GQDs). Experimental I–V measurements were conducted at eight temperatures ranging from 77 K to 400 K and used to train three ML models: K-Nearest Neighbors (KNN), Decision Trees (DT), and Gradient Boosting (GB). Predictions were then made for ten unmeasured temperatures between 100 K and 375 K, with datasets containing 1000 data points at 0.01 V intervals. The GB model outperformed others, achieving a coefficient of determination () of 0.9998, mean squared error (MSE) of 0.0026, and mean absolute error (MAE) of 0.0222. While KNN showed the lowest MSE at lower temperatures (100–250 K), model performance degraded at higher temperatures (above 300 K), especially around 325 K where dropped to 0.984. Experimental validation confirmed the accuracy of the predicted I–V characteristics. The diode was fabricated by spin-coating the nanocomposite onto an n-type Si(100) wafer (1–10 cm, 350 m thick) to create a 30 nm thin film, with gold contacts applied via sputtering. The AI-based modeling approach eliminates the need for extensive experimental measurements, enhancing efficiency, reducing energy consumption, and promoting reproducibility. GB, DT, and KNN all maintained high prediction accuracy, but GB consistently produced the best results across all metrics. This study demonstrates that AI-assisted prediction can significantly accelerate diode development and improve sustainability in electronics research.

The photo-Fenton-like process using CDs- nanocomposites was studied by Wang et al. [37]. It demonstrated excellent potential for degrading low-concentration tetracycline (TC) in wastewater across a wide pH range. CDs-modified showed enhanced peroxydisulfate (PDS) adsorption and activation, with an electron transfer of to , facilitating reactive radical generation. The rate constant k follows pseudo-first-order kinetics:

7

with (8 wt.% CDs) versus 0.0095  for pure .

The optimized catalyst, prepared with 8 wt.% CDs, achieved a maximum TC degradation rate of 86.6% within 60 min under 300 W Xe lamp irradiation, outperforming pure (42%). ML (XGB model, ) was employed to fine-tune synthesis and reaction parameters, with catalyst concentration (0.75 g/L) and PDS amount (0.125 g/L) identified as optimal. The CDs- catalyst maintained high activity from pH 1 to 13 due to the protective functional groups on CDs and the mesoporous structure (10–25 nm pore sizes). Electrochemical measurements showed a twofold increase in photocurrent and a reduction in impedance (Rt from 32,859 to 15,082 ) compared to , indicating superior carrier separation. Electron spin resonance revealed enhanced and ·OH/· radical formation under acidic and alkaline conditions, respectively. DFT and Fukui function analyses identified key reactive TC sites and proposed two degradation pathways involving successive deamination, hydroxylation, and ring-opening steps. The bandgap of was reduced from 2.91 eV to 2.35 eV with CDs, improving solar utilization, while differential charge density analysis confirmed electron transfer from to CDs.

Future work may include introducing multi-objective optimization to balance degradation rate and energy consumption; construct a reaction network model via DFT-guided kinetics to simulate intermediate species; develop in situ spectroelectrochemical probes for real-time monitoring of radical species.

QDs have garnered attention as photocatalysts due to their unique properties. However, their sonophotocatalytic/photocatalytic efficiency is hindered by various limitations that prevent effective operation after multiple cycles. Kohnehsari et al. [96] focuses on developing an efficient and recyclable sonophotocatalyst/photocatalyst using ZnO QDs (ZQDs), CuO, and graphene (G). Different amounts of ZQDs (ZQDs(x)/G, x = 10, 20, 30, 40, and 50 mg) were immobilized on graphene using a hydrothermal method, and varying amounts of CuO were impregnated on ZQDs(40)/G. PXRD patterns and Raman spectra confirmed the wurtzite and monoclinic structures for ZQDs and CuO, respectively. FESEM, AFM, and PXRD analyses revealed that the mean crystal size of ZQDs increased after immobilization on graphene and CuO impregnation. Photoluminescence (PL) and Mott-Schottky analyses indicated that the inclusion of graphene oxide (GO) and CuO in CuO(0.5)/ZQDs(40)/G reduced exciton recombination and formed p-n heterojunctions. The bandgap energy, VB, and CB potentials were determined using DRS and Mott-Schottky analyses. The photocatalytic and sonophotocatalytic degradation rates of tetracycline (TC) were measured, with apparent rate constants (kapp) of 0.030 min⁻¹ for photocatalytic and 0.060 min⁻¹ for sonophotocatalytic degradation using CuO(0.5)/ZQDs(40)/G. The treated effluent exhibited favorable effects on wheat seed growth. The energy consumption was lower for CuO(0.5)/ZQDs(40)/G during sonophotocatalytic/photocatalytic TC degradation, and it was effective over multiple cycles. Additionally, ML models, including Random Forest (RF) and AdaBoost, were developed for predicting the degradation process. The AdaBoost model outperformed the RF model with lower statistical metrics (SAE, MAE, MSE, RMSE), indicating better prediction accuracy.

To provide a concise overview of the recent advances, Table 3 summarizes key studies where ML techniques have been integrated with QD-based systems across sensing, biomedical, and environmental applications. The table highlights the types of systems developed, the ML methods employed, performance metrics achieved, target analytes, and notable features of each work.

Table 3.

Summary of ML-Enhanced QD Applications in Sensors, Biomedical, and Environmental Fields

Study	System/ application	ML technique	Key outcomes/ performance	Analyte/ function	Highlights
Yan et al. [87]	Handheld AI sensor with smartphone + CD microfluidic chip	Custom ML app for FL color detection	LOD: 0.274 M (), 0.624 M (EDTA); recovery: 98.47–106.30%	, EDTA in food/water	Ratiometric FL with red-green CDs; portable, real-time analysis
Geethukrishnan et al. [92]	Electrochemical CRT sensor with CuNW + MSQD on GCE	ANN, RF, KNN	LOD: 0.001 M (urine); RMSE: 0.017 M (RF); : 0.929 (RF)	Creatinine (CRT) in urine	Enhanced surface area; low-cost; ML boosts selectivity
Zhang et al. [47]	PQD-based fluorescence sensor array for pathogens	SVM	Accuracy: 100% for 5 pathogens (1.0103 to 107 CFU/mL); >99% inactivation	Foodborne pathogens	Fast ID + disinfection; works in water; low LODs
Zhu et al. [94]	Paper-based CQD (Ag/Cu/Zn) biosensor	Smartphone + ML classifier	Detection: 103–107 CFU/mL; Ag-CQDs kill 99.99% E. coli	5 bacteria incl. E. coli, S. aureus	Emission: 443–554 nm; live/dead staining confirms efficacy
Ersöz et al. [95]	La-doped PEI-GQD Schottky diode I–V prediction	KNN, DT, Gradient Boosting (GB)	: 0.9998 (GB); MAE: 0.0222; accurate 100–375 K predictions	I–V behavior under varying temperature	AI saves time/energy in semiconductor design
Wang et al. [37]	CDs- nanocomposite for photo-Fenton TC removal	XGB (Extreme Gradient Boosting)	86.6% degradation (60 min); ; wide pH stability (1–13)	Tetracycline (TC) in wastewater	e-transfer (); high radical formation; DFT/Fukui support
Kohnehsari et al. [96]	CuO/ ZQDs/ Graphene sonophotocatalyst	RF, AdaBoost	kapp: 0.060 ; AdaBoost > RF in all metrics	Tetracycline degradation	p-n junctions reduce recombination; reusable; boosts crop health

These sensing- and catalysis-oriented studies illustrate that ML approaches coupled to low-dimensional, physically meaningful features—such as ratiometric fluorescence ratios, electrochemical peak currents, or kinetic rate constants—tend to generalize more reliably across samples and operating conditions than image-heavy or highly device-specific models. Regression models, tree-based ensembles, and shallow neural networks perform well in field-deployable sensing systems because they tolerate limited datasets, support rapid calibration, and allow partial interpretability of contributing variables. In contrast, smartphone-based image analysis [97], SVM classifiers on fluorescence arrays, and temperature-dependent I–V predictors achieve high accuracy but remain sensitive to illumination conditions, sensor fabrication variability, and environmental drift, limiting transferability without recalibration. For photocatalytic and photo-Fenton systems, ensemble ML methods and gradient-boosting models effectively navigate multivariate synthesis and reaction spaces, yet their predictions remain confined to the explored compositional and operational regimes. Data requirements scale with task complexity: point-of-care sensing benefits from compact, task-specific datasets, whereas catalytic optimization and multifunctional materials design require broader experimental coverage. Interpretability is highest when ML augments established physicochemical models, such as adsorption isotherms, reaction kinetics, or band-structure analysis, while black-box predictors primarily serve screening and automation roles. Overall, ML methods in QD-enabled sensing and catalysis are most transferable when embedded within physically constrained measurement protocols, balancing predictive performance with data efficiency and operational scalability.

Photonics, optoelectronics, and quantum light sources

This section surveys recent ML strategies applied to photon-based devices and quantum light sources, including: ML regression of photon-correlation histograms [98] for rapid estimation [99]; spatially resolved neural-network prediction of photovoltaic J–V and trap-state parameters in PbS CQD solar cells [15]; ML-driven optimization of graphene-QD–Au nanoparticle photodetectors for sensing [100]; and autoencoded spectral feature regression for automated suitability scoring of single-photon QD emitters [101]. Each approach replaces time-consuming experiments or opaque fitting procedures with data-efficient, high-accuracy inference.

Single-photon sources (SPSs) are foundational to emerging quantum technologies, yet verifying their emission purity via Hanbury Brown–Twiss (HBT) interferometry is costly and slow. Kedziora et al. [99] explores whether ML can outperform conventional least-squares fitting (LSF) in estimating single-photon purity (via the parameter) using partial emission statistics. Eight datasets, each derived from a single InGaAs/GaAs QD under varying laser powers (1.2–30 W), contain 1954-bin time-delay histograms of photon coincidences, with durations ranging from 1210 to 23,950 s. Ground-truth g values are obtained by fitting a rate-based model to full-time histograms using Powell and Trust Region Reflective optimizations. Five ML regressors (OLS, SGD, PLSR, Random Forest, Gradient Boosting) are trained on seven datasets and tested on the eighth via transfer learning. RMSE metrics show that ML outperforms LSF when trained and tested on the same context; however, transfer learning yields inconsistent gains. Data augmentation allows statistical comparison, with ensemble models like gradient boosting often yielding lower RMSE than LSF, especially at low detection counts. The best performance stems from ML models trained with hyperparameter optimization via the HpBandSter package using 4-fold Monte Carlo cross-validation. Despite some generalisation challenges, results confirm ML’s potential for faster, early-stage assessment of SPS quality, motivating further research into cross-context generalisation via domain adaptation and feature engineering.

Recent advances in ML have enabled the prediction of photovoltaic parameters such as carrier mobility, photoluminescence (PL) intensity, and trap-state density from simple current–voltage (J–V) measurements. In the study by Lee et al. [15], neural networks were trained on over 6400 spatially resolved experimental data points per device, acquired with a custom optoelectronic scanner with 25 m resolution and 80 80 step size. The devices used were PbS colloidal QD (CQD) solar cells with layered architectures including Au, PbS CQD, ZnO, and FTO. Six substrates totaling 96 devices were fabricated and analyzed over a period of one year, with each full scan taking approximately 24 h. The networks achieved mean absolute percentage errors (MAPE) of 8.35% for PL intensity, 7.11% for transient photovoltage, 12.43% for transient photocurrent, 14.25% for trap-state density, and 18.14% for carrier mobility on training data. Validation errors closely matched, with a fivefold cross-validation average of 7.70%. A key innovation is the “neighborhood of J–V curves” method, where each prediction uses data from a 33 grid of adjacent measurements. The training error minimized at a neighborhood spacing of 300 m, indicating a correlation length for material inhomogeneities. Visualization of neural network filters showed that voltage and current data were essential, countering trends in other works that omit voltage inputs. These methods significantly accelerate photovoltaic research by replacing hours-long characterization steps with rapid, spatially aware predictions.

Das et al. [100] present a ML framework to optimize photodetector performance for sensor applications, utilizing a heterostructure of nitrogen-doped graphene QDs (Au@N-GQDs) and gold nanoparticles. Over 20 supervised ML models were trained and tested to identify the most effective algorithm for our application. The ML-based sensor showed exceptional selectivity and sensitivity in detecting ions in Brahmaputra river water, with a sensitivity at the nanomolar (nM) level. The sensor achieved a high alignment between model predictions and experimental outcomes, validated by heatmap analysis and confusion matrices. The ML framework predicts the photodetector’s spectral responsivity (Rs) and external quantum efficiency (EQE), key parameters for performance. Rs and EQE were experimentally calculated, with Rs = I / and EQE = 1.24 10⁻⁷ Rs / (nm). The sensor demonstrated high stability, with a minimal deviation of about 2% between predicted and experimental values for the photocurrent ratio (I/). The optimized ML model had an value of 0.98, with RMSE = 0.02089, MSE = 0.000436, and MAE = 0.01984. The ion detection demonstrated non-linear photocurrent changes as ion concentration increased, consistent with experimental data. Furthermore, the sensor was tested for real-world applications, including river water analysis, with a detection limit of 1.0 nM for ions.

A major challenge in quantum photonics is the efficient, on-demand generation of high-quality single photons and entangled photon pairs. Semiconductor QDs are promising emitters, yet their growth often results in random spatial and spectral distributions. To automate the evaluation of QD suitability as a single photon source, Corcione et al. [101] propose a data-driven ML approach. This method involves extracting a minimal but relevant feature representation from QD emission spectra using an autoencoded convolutional neural network (CNN). A regression model then uses these features to predict a suitability score between 0 and 1, where 1 indicates a perfect single photon source. The model is trained on a large dataset of self-assembled InAs/GaAs QD emission spectra, which is partially labeled by experts. The regression model achieves an R2 score of 96% on training and 95% on testing, with the reconstruction error as the most significant feature (correlation > 0.9 with suitability). The final prediction includes both the suitability score and a confidence measure. The method can be applied across various QD fabrication methods and materials, providing a crucial step toward automated QD evaluation. The neural network uses a fully connected architecture with four layers, optimized over 2000 epochs with a batch size of 64, achieving a mean absolute deviation of 0.05 (5% relative error).

Table 4 summarizes representative studies where ML methods were deployed to enhance device characterization, sensing precision, and emitter screening in photonics and quantum optoelectronics. Each entry outlines the application focus, employed ML techniques, key performance metrics, and the resulting impact on experimental workflows or device performance.

Table 4.

Summary of ML Applications in Photonics, Optoelectronics, and Quantum Light Sources

Study	Target application	ML technique(s)	Key performance metrics	Highlights / impact
Kedziora et al. [99]	Estimating in single-photon sources (SPS)	OLS, SGD, PLSR, RF, GB	RMSE lower than LSF; GB best across conditions; near 1.0 in matched training/testing	ML enables early, efficient SPS quality assessment; transfer learning performance varies
Lee et al. [15]	Predicting photovoltaic parameters from J–V curves	Deep Neural Networks	MAPE: 7.11%–18.14%; Cross-validation error: 7.70%	Enables fast spatially-resolved predictions of carrier mobility, PL intensity, trap density, etc.; supports high-throughput solar cell analysis
Das et al. [100]	ion detection using N-GQD-Au photodetectors	20+ supervised models (best unspecified)	= 0.98, RMSE = 0.02089, MAE = 0.01984	High sensitivity (1.0 nM), real-world river water tested; ML enables accurate EQE/Rs prediction
Corcione et al. [101]	Automated QD emitter suitability prediction	CNN + regression (autoencoder-based)	: 96% (train), 95% (test); MAD = 0.05	Universal suitability scoring for QDs; robust across fabrication types; accelerates quantum photonics material screening

These works demonstrate how ML is increasingly used to replace or shorten characterization loops in photon-based devices and quantum light sources by learning direct mappings from experimentally accessible signals to physically relevant figures of merit. In single-photon source assessment, regression models applied to partial photon-correlation histograms enable early estimation of without full Hanbury Brown–Twiss acquisition, reducing measurement time while maintaining quantitative agreement with rate-equation fits [98, 99]. For photovoltaic quantum-dot devices, spatially resolved neural networks infer trap densities, carrier mobility, and recombination proxies directly from local J–V neighborhoods, revealing correlation lengths associated with material inhomogeneity and bypassing transient optoelectronic measurements [15]. In sensor-oriented photodetectors, supervised ML models efficiently navigate multivariate design spaces to predict responsivity and EQE, enabling selective ion detection consistent with experimentally observed nonlinear photocurrent responses [100]. Automated evaluation of single-photon emitter quality further benefits from representation learning, where autoencoded spectral features correlate with emission purity and stability metrics across heterogeneous QD ensembles [101]. Across these studies, ML primarily functions as an inference and screening layer, translating high-dimensional optical or electrical data into task-specific performance indicators rather than replacing underlying physical models. The results indicate that data efficiency and transferability remain tied to how closely learned features align with established photonic observables and device physics.

Theoretical modeling and quantum system simulation

Generalized quantum impurity models, which involve a few localized strongly correlated degrees of freedom coupled to conduction electrons, are used to describe various systems, such as magnetic moments in metals and nanoelectronics quantum devices, including QDs and single-molecule transistors. These models can also be understood via dynamical mean-field theory as self-consistent impurity models. The challenge of simulating such models arises at low temperatures due to many-body effects, particularly the renormalization of parameters. To simplify these models while preserving their low-energy physics, a nonperturbative, unsupervised ML (UML) method is proposed by Rigo et al. [102]. This method optimizes an effective impurity model to match the low-energy physics of a bare model by minimizing the Kullback–Leibler divergence (KLD), which quantifies the distinguishability between two probability distributions:

8

where and . The parameter update follows gradient descent:

9

with learning rate . Convexity of in ensures global convergence.

Specifically, the method applies to quantum impurity models, where the impurity Hamiltonian is mapped to a simpler effective model by minimizing the KLD using gradient descent. The optimization is efficient due to the convexity of the KLD with respect to the effective parameters, and the computational cost is controlled by the need to compute thermal observables, such as impurity operators, in the bare model only once. For testing the method, the Anderson impurity model [103] is used, where results demonstrate that the method can be applied beyond simple models to more complex systems. The UML approach provides an efficient way to derive effective models for systems that are otherwise beyond the reach of traditional methods such as numerical renormalization group or quantum Monte Carlo methods. In particular, the optimization can yield models that better capture the low-energy behavior of systems with complex interactions. The method has been shown to produce accurate results and is systematically improvable, which makes it a promising tool for future work in the field of strongly correlated electron systems. One can accelerate convergence using second-order methods (e.g., L-BFGS) by estimating the Hessian of via stochastic sampling; further, integrating physics-informed network architectures that respect symmetries of H can reduce parameter count by up to 50%. A future direction involves variational autoencoder encoders to discover emergent low-dimensional manifolds of impurity spectral functions for automated model compression.

The implementation of quantum ML (QML) protocols using Adaptive Boson Sampling (ABS) has been demonstrated through experiments on integrated photonic circuits by Hoch et al. [104]. These circuits, fabricated using femtosecond laser writing, employ m-mode universal photonic devices to implement ABS with increasing complexity. Specifically, the ABS protocol with 6 modes and 2 input photons was demonstrated, with results showing fidelities of , , and , indicating high photon indistinguishability. As the complexity increased, an 8-mode circuit was employed for multi-photon states, where the fidelity averaged at . These setups were further expanded to encode quantum states into qutrits, with fidelity values of and . In a binary classification task with a 1D dataset of 15 labeled points, the kernel obtained from ABS achieved an accuracy of 90%, while the 2D dataset, consisting of 200 points, was classified with an accuracy of 80%. The adaptive operations were configured to estimate the quantum kernel for the SVM model, where 15 15 kernels were reconstructed from experimental data. The kernel estimation for both qubits and qutrits was performed through quantum tomography, yielding high-fidelity results and successful kernel computation. These findings show that adaptive Boson Sampling can be an effective tool for feature mapping in quantum ML applications.

Zieliński [59] presents a method for predicting single-particle energies in InAs/InP double nanowire QDs using a neural network trained on atomistic tight-binding calculations. With only a small training set and the help of transfer learning, the model achieves root-mean-square deviations (RMSE) near 1meV, making it highly efficient. The input features include five morphological parameters (dot heights, radii, and separation), and outputs are electron and hole ground state energies. Calculating the full 3750-case dataset requires over 6000 CPU hours, while training the neural network is comparatively light. Regression methods were tested, with neural networks outperforming others once trained on  50% of the data. Data augmentation using k-nearest neighbors improves performance for smaller training sets. When extended to a larger configuration space of 59,049 QD systems, transfer learning from the smaller dataset enables accurate predictions using as little as 4% of the larger set. Despite increased system complexity and computational costs, RMSE remains around 1meV with transfer learning. Visualization and comparison confirm strong agreement between neural network outputs and tight-binding calculations. This approach significantly reduces computational burden and can aid the inverse design of nanostructures based on spectral properties.

Table 5 summarizes recent ML applications in quantum system simulation. Each entry details the target system, modeling approach, dataset scale, algorithmic specifics, and achieved performance metrics.

Table 5.

Summary of ML in Theoretical Modeling and Quantum System Simulation

Study	Target application	ML technique(s)	Key performance metrics	Highlights / impact
Rigo et al. [102]	Effective modeling of quantum impurity systems	Unsupervised ML; KLD minimization via gradient descent	High accuracy in low-energy behavior capture; Convex optimization	Simplifies many-body models; generalizes beyond simple models; efficient vs. traditional methods (e.g., NRG, QMC)
Hoch et al. [104]	Quantum kernel estimation using Adaptive Boson Sampling (ABS)	Quantum-enhanced ML; SVM classification with ABS kernels	Fidelity: , ; Classification accuracy: 90% (1D), 80% (2D)	Photonic circuits for quantum ML; ABS effective for kernel computation in QML tasks; scalable quantum hardware prototype
Zieliński [59]	Predicting electronic states in nanowire QDs	Neural Networks with transfer learning	RMSE 1meV; strong correlation with tight-binding results	Efficient simulation of large design spaces; significant CPU savings; enables inverse design of nanostructures

These studies illustrate the use of ML across distinct regimes of quantum modeling, spanning effective many-body descriptions, photonic quantum kernels, and atomistic nanostructure simulations. In generalized quantum impurity models, unsupervised optimization based on Kullback–Leibler divergence enables the construction of reduced effective Hamiltonians that preserve low-energy thermodynamic behavior, providing an alternative to numerical renormalization group and quantum Monte Carlo methods in regimes dominated by strong correlations [102, 103]. The convex structure of the optimization problem allows stable convergence, while extensions incorporating second-order optimization or symmetry-aware parameterizations further reduce computational cost and model complexity. In quantum photonics, adaptive Boson Sampling implemented on integrated photonic circuits realizes experimentally accessible quantum kernels for supervised learning tasks, demonstrating classification accuracies up to 90% with qubit and qutrit encodings and high-fidelity kernel reconstruction [104]. At the atomistic scale, neural networks trained on tight-binding calculations accurately predict single-particle energy levels in coupled nanowire quantum dots with meV-scale errors, even under substantial reduction of training data through transfer learning [59]. Together, these examples show how ML functions as a surrogate or compression layer, mapping high-dimensional quantum models or experimental observables onto tractable representations while retaining quantitative agreement with established theoretical or numerical references.

Overview of key findings

To better illustrate the landscape of key challenges and future research directions identified in this work, a structured mind map is presented in Fig. 8. The diagram organizes the main limitations encountered across different studies and highlights promising solution pathways.

Fig. 8.

Mind map summarizing key challenges and future directions in ML–assisted QD research

Methodological guidelines

Although a broad spectrum of ML models has been applied to quantum dot (QD) systems, their selection is often guided implicitly by problem structure rather than articulated methodological criteria. In practice, the choice of learning paradigm reflects the relationship between the physical task, the available data, and the required level of interpretability or control [105, 106].

Forward modeling tasks, such as predicting electronic spectra, wave functions, or charge transport characteristics from structural or electrostatic parameters, are naturally formulated as supervised regression problems. When labeled datasets are large and generated either from first-principles simulations or controlled experiments, deep neural networks and neural network potentials provide flexible approximations capable of capturing nonlinear quantum-mechanical dependencies [107]. In contrast, when training data are limited or physical interpretability is required, kernel-based methods, shallow regressors, or hybrid physics–ML models offer improved stability and generalization [72].

For inverse design and parameter optimization problems, where the objective is to identify QD configurations or control settings that achieve target functionality, direct supervised prediction is often insufficient. Bayesian optimization is well suited to experimental regimes characterized by high measurement cost and low throughput, as it enables efficient exploration of multi-parameter spaces while explicitly modeling uncertainty [108]. Evolutionary strategies and reinforcement-learning approaches become advantageous when the control landscape is nonconvex, discontinuous, or when optimization requires sequential decision-making under partial observability [106].

In device tuning, control, and state recognition, classification-oriented models and reinforcement-learning controllers are typically preferred, as these tasks involve discrete state identification and closed-loop interaction with the experimental system. Convolutional neural networks and ensemble classifiers have demonstrated robust performance when trained on stability diagrams or transport measurements, while physics-informed control frameworks reduce data requirements and improve transferability across device architectures [105, 108].

Finally, physics-informed and hybrid learning approaches are particularly effective when data availability is limited, device-to-device variability is high, or extrapolation beyond the training distribution is required. By embedding governing equations, symmetry constraints, or physical priors into the learning process, such models improve interpretability and robustness compared to purely data-driven alternatives [67, 72]. These considerations indicate that successful application of ML in QD research depends less on model complexity than on alignment between the learning paradigm, the data regime, and the physical structure of the problem.

Challenges

Data-related challenges include measurement speed, scalability, and dataset diversity. Closed-loop autotuning systems, such as those by Yon et al. [73], require tens of seconds per voltage patch, resulting in multi-hour total tuning times, while large-scale readout systems demonstrated by Thomas et al. [75] face constraints in cryogenic multiplexing and data analysis. Dataset size and heterogeneity also remain limited: CNNs and synthetic-plot classifiers trained on small, specialized datasets [73, 74] may not generalize to devices with different materials, noise characteristics, or voltage ranges, and open benchmark datasets for QD stability diagrams are scarce.

Algorithmic challenges encompass model generalization, interpretability, and high-dimensional optimization. Many deep learning architectures, including CNNs and LSTMs, optimize predictive accuracy but lack transparency regarding physical parameters [109, 110]. While hybrid frameworks and explainability techniques such as SHAP or LIME provide insights into feature importance, most models remain black boxes [74]. Multi-dot tuning introduces further complexity, as the dimensionality of gate-voltage spaces grows rapidly with array size, requiring methods such as Bayesian optimization, active learning, or graph neural networks [76, 111].

Physical-modeling challenges relate to device fabrication, materials, and the connection between ML outputs and underlying physics. Real-time feedback control of QD growth has been demonstrated in limited material systems [77], but extending these methods to diverse synthesis techniques and materials remains open. Atomistic simulations of ligand configurations and surface dynamics [53] are constrained by short simulation times and limited transferability across chemistries. Similarly, sensing platforms based on colloidal QDs or perovskites show sensitivity but remain vulnerable to environmental variability, electrode fouling, and long-term stability issues [112–114].

Failure of modes of ML in QD research constrain reliability and generalization. Dataset bias is prevalent because training data are often collected from a single fabrication run or measurement setup, leading to models that overfit process-specific artifacts; for example, SPS purity predictors trained on one excitation regime show degraded accuracy when transferred across laser powers or detector efficiencies, sometimes underperforming conventional fitting [99]. The absence of standardized benchmarks for QD characterization tasks (e.g., emission stability, trap density inference, or spectral suitability scoring) makes cross-study comparison difficult and inflates reported performance metrics. Cross-laboratory transferability remains limited, as models trained on spatially resolved photovoltaic or PL datasets often require manual recalibration or retraining when device architectures, substrates, or environmental conditions change [15]. High experimental cost further restricts dataset diversity: spatial scans, HBT measurements, or tight-binding simulations are expensive, resulting in small or imbalanced datasets that necessitate manual data curation or augmentation [59]. Uncertainty quantification is frequently absent or post hoc, even though predictive confidence is critical when ML outputs guide experimental decisions; without calibrated uncertainty estimates, models may yield overconfident but physically inconsistent predictions. These failure modes indicate that ML in QD research should be viewed as an assistive inference layer rather than a drop-in replacement for physics-based validation, motivating stronger emphasis on benchmark datasets, domain adaptation, and uncertainty-aware modeling.

To systematically review recent trends in the application of ML to colloidal QD research, Table 6 outlines the main characteristics of each study, including the use of closed-loop experimental design, property optimization, model interpretability, physics-informed approaches, experimental validation, and availability of open-source code.

Table 6.

Comparative overview of key studies on ML applications in QDs

Research	Closed- loop	Property optimization	Interpretability	Physics- based	Experimental validation	Open code
Yon et al. [73]
Schug et al. [74]
Thomas et al. [75]
Kashtiban et al. [76]
Zhang et al. [53]
Shen et al. [77]
Corcione et al. [101]
Rigo et al. [102]
Hoch et al. [104]
Zieliński [59]

— feature is implemented in the study;     — not implemented or not reported

Future work

To address these challenges, future work should expand and diversify training datasets via multi-site field trials, federated learning, and physics-informed data augmentation to improve model generalization. Development of standardized calibration protocols and integration of real-time [115, 116] self-calibration routines—perhaps leveraging built-in reference features or ratiometric measurements—would enhance reliability across smartphones and lighting conditions. Material innovations, such as encapsulation strategies for QD films [117–119] or antifouling surface chemistries for electrodes, could extend sensor longevity and reproducibility. For catalytic and photocatalytic systems [96, 120], systematic studies of long-term performance in real wastewater or agricultural run-off will be crucial, along with life-cycle assessments to quantify environmental impact. Embedding on-chip microcontrollers or low-power wireless modules could enable fully autonomous, networked deployments, supporting continuous environmental monitoring and rapid response in resource-limited settings.

To address gaps related to data quality, future efforts should assemble larger, multi-center datasets spanning diverse QD materials, device architectures, and environmental conditions, perhaps via federated learning [121] to preserve proprietary data. Including physics constraints—such as energy-level conservation in models or known J–V curve shapes—into ML architectures could bolster robustness and reduce data demands. Systematic uncertainty quantification [122, 123] (e.g. Bayesian neural nets or conformal prediction) will be crucial for deployment in field instruments. Interpretability techniques (saliency maps, symbolic regression) could illuminate which spectral or spatial features drive predictions, guiding both model trust and new physics insights. Finally, real-time, on-chip implementations of these ML routines—leveraging low-power FPGAs or neuromorphic hardware—would enable closed-loop, autonomous optimization of QD growth, device tuning, and performance monitoring in next-generation quantum and optoelectronic systems.

Conclusion

Machine learning has become a widely used tool across multiple areas of quantum dot research, supporting tasks that range from device tuning and characterization to materials development, sensing, and theoretical modeling. In device-level studies, data-driven control and classification methods have reduced manual intervention and enabled more systematic exploration of large parameter spaces, offering practical routes toward scaling from individual quantum dots to larger arrays. These approaches primarily function as accelerators of existing workflows rather than replacements for established experimental protocols.

In materials synthesis and property optimization, ML-assisted parameter searches and adaptive experiment design have shortened iteration cycles and reduced experimental load when targeting optical or electronic properties. Similar benefits are observed in sensing and environmental applications, where ML-enabled signal interpretation supports rapid, on-site analysis while maintaining compatibility with low-cost and portable hardware platforms. These results highlight the role of ML as an organizing layer that links measurement outputs to actionable parameters.

In photonic and optoelectronic systems, ML methods have reduced reliance on time-intensive fitting and mapping procedures, enabling faster estimation of figures of merit such as photon purity, photovoltaic parameters, and detector responsivity. Feature compression and regression techniques further support automated screening of emitters and devices, although their effectiveness remains tied to the quality and representativeness of training data.

At the modeling and simulation level, ML-based surrogates and unsupervised reduction schemes provide approximate descriptions of complex quantum systems that are otherwise computationally demanding. These methods expand the accessible system size or parameter range but continue to rely on physics-based models for validation and interpretation.

Current evidence suggests that ML is most effective in quantum dot research when used as a complementary inference and optimization tool. Persistent challenges include limited cross-platform generalization, experimental cost of data acquisition, and insufficient treatment of predictive uncertainty. Addressing these issues through standardized benchmarks, uncertainty-aware modeling, and closer integration with physical constraints will be necessary for broader and more reliable deployment of ML-assisted quantum dot technologies.

Author contributions

Conceptualization, I.M., V.T. and D.M.; Data curation, I.M., D.M. Project administration, V.T. and A.B.; Resources, V.N.; Software, I.M., A.B, I.M. and A.G.; Supervision, V.T., A.B., R.V., I.M., and V.N.; Validation, I.M., V.T., A.B. and A.G.; Visualization, I.M., D.M.; Writing—original draft, I.M., D.M., V.T., V.N., A.B. and A.G. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethical approval and consent to participate

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Conflict of interest

The authors declare no Conflict of interest.