Autonomous microfluidic experimentation for exploring reaction inference and synthesizing double perovskite nanoplatelets

Abstract

Self-driving laboratories enable accelerated exploration of chemical and materials spaces by coupling automated experimentation with machine-learning-guided decision making. However, extending autonomous discovery to compositionally complex materials with multiple coupled reaction pathways remains a significant challenge. Here, we introduce PoLARIS, a microfluidic self-driving laboratory designed for time- and material-efficient autonomous synthesis, optimization, and mechanistic interrogation of multi-element nanocrystals. Using PoLARIS, we achieve rapid data-driven optimization of metal halide double perovskite nanoplatelets, comprising up to six distinct elements synthesized via a continuous-flow heat-up reaction. The platform integrates a modular microfluidic reactor architecture with closed-loop experiment selection to efficiently navigate a high-dimensional synthesis parameter space. Beyond autonomous multi-element nanoplatelet synthesis and optimization, PoLARIS utilizes dynamic flow experimentation to enable mechanistic inference of precursor reactivity and reaction pathways governing nanoplatelet formation. This work establishes microfluidic self-driving laboratories as a generalizable approach for unifying autonomous synthesis optimization with mechanistic understanding in compositionally complex colloidal materials systems. PoLARIS framework provides a scalable pathway toward autonomous discovery in other multi-element and high-entropy colloidal nanocrystals beyond double perovskites.

Research into reaction mechanisms in the synthesis of materials with complex compositions remains a challenge. Here, the authors present a microfluidic, self-driving lab approach that enables rapid optimization and provides mechanistic insights into double perovskite nanoplatelets.

Introduction

Colloidal nanocrystals (NCs) have undergone a significant evolution over the past five decades, from elemental nanoparticles to increasingly complex, multicomponent systems, which is mainly driven by their tunable physicochemical properties and broad applicability in areas such as photovoltaics, optoelectronics, catalysis, and biomedicine^1–5. Initial breakthroughs centered on the synthesis of noble metal NCs (e.g., gold⁶, silver⁷), followed by binary II-VI and III-V semiconductor quantum dots⁸, and more recently, lead halide perovskites with the general formula APbX₃ (X = Cl, Br, I; A = Cs⁺, MA⁺, FA⁺)^9,10, which have attracted substantial attention for their exceptional optoelectronic performance.

More recently, the colloidal nanoscience field has shifted toward multi-element NCs with expanded chemical design spaces, such as double perovskites and high-entropy nanoparticles. Among these NCs, halide double perovskites (HDPs), typically expressed as A₂M⁺M³⁺X₆, where A is a monovalent A-site cation (e.g., Cs⁺, Rb⁺, K⁺, MA⁺, FA⁺), M⁺ is a monovalent B-site metal (e.g., Ag⁺, Na⁺, Cu⁺), and M³⁺ is a trivalent B-site metal (e.g., Bi³⁺, Sb³⁺, In³⁺)^11–13, are especially compelling due to their reduced toxicity relative to lead-based analogues and intriguing structure-property relationships, including tunable bandgap character (direct-indirect), defect tolerance, and the emergence of self-trapped excitons. These features render HDPs attractive candidates for optoelectronic devices¹⁴, photocatalysis^15,16, and light emission technologies¹⁷.

Despite growing interest, research on HDP nanoplatelets (NPLs) remains nascent. Several scientific and engineering challenges continue to hinder progress. The rich compositional and structural degrees of freedom of HDP NPLs introduce a vast, high-dimensional synthesis space governed by nonlinear and interdependent parameters such as precursor identity and ratio, ligand chemistry, temperature, reaction time, and solvent environment. Traditional experimental strategies, which are largely based on manual trial-and-error and one-variable-at-a-time (OVAT) exploration methods, cannot efficiently navigate such combinatorial complexity¹⁸. These approaches are often inefficient, lack reproducibility, and risk missing critical interactions among parameters that govern NPL formation and the resulting optical/optoelectronic properties. Thus, the rational design and reproducible synthesis of HDP NPLs with targeted properties remains a fundamental bottleneck in the field.

To address these challenges, self-driving laboratories (SDLs) have emerged as a promising paradigm for data-driven, autonomous experimentation in chemistry and materials science¹⁹. SDLs integrate automated high-throughput experimentation with machine learning (ML) algorithms to form closed-loop systems capable of hypothesis generation, automated experimentation, real-time data acquisition, and iterative optimization^20,21. ML algorithms, often powered by Bayesian optimization (BO) or reinforcement learning, efficiently guide experiment-selection within complex design spaces, while data-driven ML models serve as digital surrogates to predict synthesis outcomes and accelerate convergence to optimal conditions^22,23. These autonomous frameworks have already demonstrated impact across a range of materials systems, including small-molecule synthesis, catalyst discovery, and colloidal quantum dot optimization^24–27. Within the SDL landscape, microfluidic-based SDLs offer unique advantages for the accelerated synthesis and discovery of colloidal NCs. Unlike batch systems, microfluidic platforms enable continuous synthesis under dynamically tunable conditions, allowing for efficient exploration of synthesis parameters and rapid generation of high-quality experimental data^28,29. Microfluidic systems exhibit superior control over reaction kinetics due to enhanced mixing, heat transfer, and reproducibility, while their modular architecture facilitates integration of in-situ characterization tools and real-time feedback loops^30,31. These features are particularly advantageous for studying NCs such as double perovskites, where subtle changes in composition and processing conditions can dramatically alter material properties.

While SDLs have been successfully applied to lead-based perovskites and binary NC systems, there remains a significant gap in the application of autonomous microfluidic platforms for the synthesis and optimization of HDP NPLs. Existing studies often focus on compositional libraries prepared via manual methods or rely on fixed batch conditions, limiting the ability to probe transient regimes, explore broad parameter spaces, or uncover nuanced structure-property relationships in a systematic and reproducible manner³².

In this work, we present PoLARIS (perovskite laboratory for autonomous reaction inference and synthesis), which is a modular microfluidic SDL for the autonomous synthesis and optimization of HDP NPLs, Cs₂Ag_xNa_1-xIn_yBi_1-yCl₆ (0 < x, y < 1). PoLARIS integrates programmable fluidic modules for reagent delivery and mixing, in-situ optical characterization for rapid feedback, and a cloud-connected database for real-time data logging and experiment traceability. A bespoke BO algorithm orchestrates experiment selection, enabling intelligent navigation of the multi-element synthesis landscape and adaptive refinement of HDP NPLs properties. Through this integrated framework, we demonstrate fast-tracked data-driven exploration of the multidimensional parameter space governing HDP NPLs formation and uncover previously inaccessible composition-property trends. This study provides a blueprint for accelerating the discovery and understanding of complex, multicomponent nanomaterials through autonomous microfluidic experimentation.

Results

PoLARIS’s autonomous framework

To accelerate the synthesis and optimization of HDP NPLs, specifically Cs₂Ag_xNa_1-xIn_yBi_1-yCl₆ (0 < x, y < 1) system, where Ag⁺ and Na⁺ are alloyed on the monovalent B-sites, and Bi³⁺ is introduced as a dopant on the trivalent B-sites, we defined a single, measurable objective, that is, maximizing the photoluminescence quantum yield (PLQY) of the synthesized NPLs. This objective informed the architecture of a bespoke closed-loop platform (Fig. 1) centered on seven controllable synthesis variables, including the concentrations of NPL precursors (cesium [Cs], silver [Ag], sodium [Na], indium [In-HCl], and bismuth [Bi]), reaction solvent (1-octadecene, ODE), and the reaction temperature. PoLARIS’s hardware comprises five tightly integrated modules, including microfluidic synthesis, in-situ optical characterization, automated data processing, surrogate-model training, and automated decision-making (see Methods), underpinned by a cloud database that captures complete data provenance and enables remote interrogation.

Fig. 1. Schematic illustration of PoLARIS’s framework.

a Accelerated fundamental and applied studies of HDP NPLs with a large parameter space (red asterisks are the combinations of different parameters) using a microfluidic SDL. b Guided by the scientific objective, a customized closed-loop workflow is designed accordingly, consisting of four interconnected stages: microfluidic synthesis, in-situ characterization and real-time data processing, model training, and experiment selection, with all raw and processed data stored in a cloud database. c PoLARIS autonomously reveals optimal synthesis conditions and the evolution of best performance over iterations.

PoLARIS initiates the synthesis space exploration of HDP NPLs with a Latin hypercube sampling (LHS) design to ensure broad, space-filling coverage of the multivariate domain. The microfluidic SDL’s hardware autonomously executes each NPL synthesis condition and records in-situ UV-Vis absorption and photoluminescence (PL) spectra. The data processing pipeline performs background subtraction and baseline correction, then extracts physically meaningful descriptors (absorbance at the excitation wavelength, emission peak position, peak intensity, and integrated emission area). Both raw and processed data, together with the exact synthesis parameters and timestamps, are archived. These normalized inputs train a surrogate ML model that is then utilized to propose new experiments predicted to improve a validated PLQY proxy, thereby closing the autonomous experimentation loop. Iterative execution and model update concentrate sampling in high-performing regions while preserving targeted exploration of uncertain regimes. To maintain data integrity at scale, each iteration includes pre-defined quality control. Convergence is assessed by stabilization of the best-observed PLQY proxy and diminishing expected improvement of the ML-predicted conditions. This workflow ensures that gains in PLQY are attributable to controllable parameters rather than measurement artifacts.

PoLARIS’s hardware

The microfluidic platform integrates precursor delivery, precursor formulation, droplet generation, rapid heat-up reaction, and in-situ spectroscopy (Fig. 2a and Supplementary Fig. 1) to enable high-throughput automated synthesis of HDP NPLs. Seven computer-controlled syringe pumps (equipped with stainless-steel syringes) continuously deliver the NPL precursor solutions and the reaction solvent at specified volumetric flow rates. Six reactive precursor streams are combined in a multiport (seven-way) junction, after which a T-junction introduces an inert carrier fluid (perfluorinated oil, PFO) to segment the reactive phase stream into monodisperse droplets (5–15 µL). Internal recirculation within each droplet ensures rapid and homogeneous mixing of precursors before entering the heated section of the microfluidic reactor, and the PFO carrier phase also served as a self-lubricating layer, preventing the precipitation of NPLs on the wall of tubing²⁸.

Fig. 2. Schematic illustration and validation of PoLARIS’s physical setup.

a Schematic illustration of the developed microfluidic platform (i.e., PoLARIS’ hardware) for data-driven synthesis of HDP NPLs. b Microfluidic platform’s workflow integrates syringe pumps for precursor delivery, a seven-way junction for mixing, a T-junction for droplet generation, an aluminum-based spiral reactor with PID-controlled heating, and a flow cell coupled to a light source and spectrometer for in-situ optical characterization. c Reproducibility and (d) stability studies of the developed microfluidic platform based on the integrated PL area. Source data are provided as a Source Data file.

Reactive phase droplets, containing the HDP NPL precursors, move through a custom-designed and CNC-machined aluminum heating plate where a fast heat-up NPL synthesis proceeds through precursor conversion, monomer formation, nucleation, and growth^33–35. A PID controller maintains the reaction temperature with high fidelity. The reaction channel is arranged in a spiral configuration to minimize reactive phase droplet deformation and breakup, and a top cover plate reduces convective heat loss. Upon exiting the heated section of the microfluidic reactor, reactive phase droplets rapidly cool in ambient air to terminate further NPL growth, then pass through two customized flow cells equipped with a fiber-coupled UV-Vis light source, an excitation light-emitting diode (LED), and two miniaturized spectrometers for in-situ characterization of the in-flow synthesized NPLs. NPLs products are collected downstream for off-line analysis as required.

We quantified microfluidic SDL’s hardware performance via reproducibility and stability tests using three spectral descriptors (peak intensity, absorbance at the excitation wavelength, and integrated emission area). Reproducibility is assessed by an experimental perturbation study with five random synthesis conditions performed between a constant, control synthesis condition as a reference. And to assess the stability, a continuous synthesis of HDP NPLs under constant conditions is run for more than 1 h. Coefficients of variations (CVs) for reproducibility were 2.2%, 2.4%, and 2.1%, respectively; flow stability CVs were 4.5%, 4.9%, and 4.2% (Fig. 2c, d, and Supplementary Figs. 2–5). These metrics confirm high repeatability and operational robustness of PoLARIS’s hardware, establishing the platform’s suitability for autonomous synthesis science studies of HDP NPLs.

ML pipeline

The digital layer of PoLARIS performs real-time data acquisition, data processing, data-driven modelling, and decision making. Since the carrier fluid (PFO) is non-reactive, it is identified and discarded in real-time data analysis. For each reactive phase droplet, the in-situ acquired UV-Vis absorption and PL spectra provide the HDP NPL properties (via optical descriptors listed above). To define a reliable optimization target, we correlated the above-mentioned optical descriptors with the absolute PLQY of HDP NPLs measured off-line (Edinburgh Instruments, FS5 spectrometer). With the assumption that the absorption coefficient is relatively constant over the entire parameter space, the integrated PL area was normalized by an approximation of absorbed photons. Because this method is empirical, and in-situ monitoring of NPL size and thickness evolution was not feasible during the automated campaign, we acknowledge that this proxy may lose accuracy if drastic morphological changes alter the absorption cross-section independently. However, within our explored parameter space, the normalized PL area measured in-situ tracked absolute PLQY (measured off-line) closely (Supplementary Fig. 6), and thus, we adopted it as a robust PLQY proxy for the closed-loop experiments. All HDP NPL synthesis inputs (normalized precursor volumetric flow rates and reaction temperature) and outputs (HDP NPL properties) were scaled (0-1) prior to ML modeling to balance sensitivity across dimensions.

PoLARIS’s ML pipeline used a Gaussian process regression (GPR) surrogate model because of its sample efficiency and calibrated uncertainty, key attributes when experiments are costly, and batch sizes are small²². PoLARIS’s experiment-selection algorithm utilized a hybrid acquisition strategy that interleaves exploitation and exploration (Supplementary Fig. 7). Each five-candidate group of autonomously selected experiments was comprised of four suggestions from a log-transformed expected improvement (EI) policy to refine promising regions of the experimental space and one suggestion from an upper-confidence-bound (UCB) rule to probe uncertain areas of the synthesis space. Simple box and linear constraints enforced practical bounds (e.g., temperature limits, total flow budget) and preserve meaningful precursor ratios. Additionally, model-proposed conditions violating constraints were repaired before execution.

PoLARIS’s model reliability was assessed continuously. Root mean squared error (RMSE) and R-square (R²) analyses based on the GPR model revealed the reliability and accuracy during autonomous optimization campaigns. This combination of reliable data acquisition and ongoing model validation accelerates convergence while yielding interpretable guidance for subsequent scale-up studies.

Closed-loop campaign

After completing hardware validation and benchmarking, PoLARIS was utilized to autonomously explore the synthesis space and optimize the PLQY of HDP NPLs. The data-driven optimization process was designed to demonstrate the autonomous operation capability of the microfluidic SDL and to identify high-performance synthesis conditions within the experimentally constrained design space. The surrogate model was initialized with 80 LHS experimental conditions (Supplementary Table 1), before initiating the autonomous optimization loop. The parameters selected by the ML model (Supplementary Table 2, Fig. 3b) increasingly converged toward their boundary limits, especially for the Ag, Na, Bi, and ODE precursors, as well as the reaction temperature. This boundary-hitting behavior indicates that the model had already exploited the accessible parameter space, and further improvement was constrained by experimental limits, which explains why the PLQY proxy showed limited improvement after 15 iterations. Such behavior is common in microfluidic synthesis systems; in this study, the Perfluoroalkoxy (PFA) tubing’s heat resistance properties limited the maximum reaction temperature (220 °C), and seven precursor streams imposed a volumetric flow rate limit (25–60 µL min⁻¹) of each stream on the input synthesis parameters. As can be seen in Fig. 3a, the best value of PLQY proxy was improved steadily from 17% to 30% (45% for purified sample) over the data-driven optimization process, while the best PLQY proxy in all LHS conditions was 26.4%, demonstrating PoLARIS’s ability to identify improved NPL synthesis conditions without human intervention. The absolute PLQY of the champion NPLs was 28% for crude solution and 45% for purified sample (measured off-line with an integrating sphere, FS5 spectrometer), which further verified the reliability of PLQY proxy calculated through in-situ spectroscopy. Furthermore, the GP model shows robust performance during the closed-loop BO campaign with a high R² > 0.99, and an RMSE value less than 0.01 (Supplementary Fig. 8). The improvement trend and the reliable ML model with high accuracy indicate that the model effectively exploited informative regions of the parameter space and rapidly converged toward high-performance synthesis conditions. The absorption spectra of the champion NPLs exhibit a peak at about 360 nm and the PL spectra (Fig. 3c) show an emission peak at 610 nm with a full width at half maximum (FWHM) of 216 nm. The broad emission is from the self-trapped exciton mechanism^17,36–38. The second peak around 430 nm is attributed to metal-ligand complexes³⁹.

Fig. 3. Results of closed-loop HDP NPL synthesis campaign by PoLARIS.

a Photoluminescence quantum yield (PLQY) proxy over 40 BO iterations following 80 LHS initialization experiments. b The synthesis conditions of the champion HDP NPL sample, autonomously identified by PoLARIS. c Area under the emission peak for 1 h of continuous manufacturing under the optimal condition. The PL and absorption spectra of HDP NPLs synthesized under optimal conditions (left inset), and the purified sample under 365 nm UV light (right inset, scale bar: 1 cm). Source data are provided as a Source Data file.

One of the important features of PoLARIS is its experimental efficiency, where the entire 120 experiments were completed within 12 h, consuming about 30 mL of each precursor. Overall, this closed-loop experimental campaign validates the ability of the developed microfluidic SDL framework to autonomously navigate multi-element synthesis spaces and operate efficiently within realistic constraints.

Next, the synthesis conditions of the champion HDP NPLs (Fig. 3b), identified by PoLARIS within half a day of autonomous experimentation, were directly transferred to a continuous manufacturing run, which demonstrated knowledge transfer with a CV of ≈ 4.0% over a 1 h continuous operation (Fig. 3c). The continuously manufactured HDP NPLs were collected for further off-line characterizations. To elucidate the structure of the optimized HDP NPLs, we performed X-ray diffraction (XRD), transmission electron microscopy (TEM), and TEM energy-dispersive X-ray spectroscopy (EDS). The XRD pattern of the champion NPLs (Supplementary Fig. 10) shows the characteristic reflections of the double-perovskite phase, with prominent diffraction peaks indexed to the (220) and (400) planes. Importantly, no additional reflections attributable to common secondary phases were observed within the measured 2θ range, indicating that the material remains phase-pure under the optimized synthesis conditions.

TEM-EDS further supports the composition of the champion sample. The measured atomic fractions are approximately 20.4% Cs, 6.3% Ag, 6.3% Na, 7.1% In, 0.3% Bi, and 59.6% Cl, as summarized in Supplementary Table 3. The compositional analysis, XRD data, and optical characterization support assignment of the high-PLQY NPLs to a mixed-cation double-perovskite nanocrystal structure.

PoLARIS’s digital twin

The trained GPR model using LHS experimental data also serves as a data-driven digital twin of the HDP NPL reaction that can reliably predict the NPLs’ PLQY. The ML model (i.e., data-driven digital twin) performance was evaluated using 10-fold cross-validation, and the corresponding parity plot (Fig. 4a) demonstrates predictive capability (R² = 0.71), which indicates that the data-driven digital twin captures the essential nonlinear structure-property relationships of the HDP NPL synthesis space. Deviations from the diagonal trend in the parity plot occur predominantly at higher PLQY values, likely due to the sparsity of samples in this regime and the nonlinear compositional interaction of HDP NPLs. Beyond prediction accuracy, interpretability of the trained ML model is obtained through Shapley additive explanations (SHAP) analysis, shown in the bottom panel of Fig. 4b. SHAP quantifies how each synthesis parameter contributes to the predicted PLQY proxy by decomposing model outputs into additive feature contributions. Among the seven HDP NPL synthesis inputs, Cs precursor, In-HCl precursor, and reaction temperature (T_rxn) emerged as the most influential parameters governing PLQY. High positive SHAP contributions for these variables (red points distributed toward positive SHAP values) indicate that higher Cs loading, increased halide-rich environment (In-HCl), and elevated temperatures have a positive impact on PLQY of HDP NPLs. ODE, Ag, Na, and Bi exhibit lower overall impact on the PLQY, with their SHAP values clustered more tightly around zero. ODE exhibits only a modest influence, supporting the interpretation that it plays a secondary role compared to halide stoichiometry and reaction temperature in determining radiative efficiency in these NPLs. The SHAP analysis also highlights mechanistic trends, where high reaction temperature (deep red points) consistently shifts predictions toward higher PLQY, reflecting enhanced crystallinity and reduced defect densities at elevated temperatures⁴⁰. Similarly, higher In-HCl content improves PLQY by promoting full halide coordination of metal centers and suppressing halide vacancies^40,41. These defects are known to introduce nonradiative recombination pathways^42,43. Increased Cs content, another high-impact synthesis parameter, likely stabilizes A-site occupancy and improves octahedral rigidity, thereby reducing structural disorder that often dominates the formation of impurities^44,45. To further illustrate the predictive capabilities of the trained data-driven digital twin, Fig. 4c presents the model-generated surface response maps across selected subspaces of the seven-dimensional synthesis domain. These surfaces represent the predicted PLQY values as a function of two most experimentally important variables, Cs and In-HCl, and three fixed normalized input values (0, 0.5, and 1) of temperature, ODE, and Bi, as well as constant value of Ag and Na due to the low impact on model output (fixed at 0.5). Such visualizations provide an intuitive understanding of how coupled NPL synthesis parameters jointly modulate HDP NPL quality. The data-driven digital twin captures key physical trends where regions of high Cs and In-HCl concentration predict enhanced PLQY, consistent with the SHAP analysis, which shows a strong dependence on halide supply and A-site stoichiometry. In contrast, domains associated with low reaction temperatures or insufficient precursor concentration display suppressed PLQY, reflecting kinetically limited growth and enhanced defect formation. The surfaces smoothly interpolate across sparsely sampled experimental regions, demonstrating that the GPR-based model generalizes effectively beyond the measured LHS points.

Fig. 4. PoLARIS’s digital twin study.

a Digital twin parity plot from 10-fold cross-validation. b SHAP analysis of PoLARIS’s digital twin. The color scale indicates the feature value of each data point. c Surface plot response of photoluminescence quantum yield (PLQY) proxy (z axis) on the normalized value of X_Cs and X_In-HCl for three values (each) T_rxn, X_ODE, and X_Bi for three values (each), and a constant normalized input of X_Na (0.5) and X_Ag (0.5). The color scale indicates the PLQY from 0 to 0.30 in the parameter space. Source data are provided as a Source Data file.

Dynamic flow-driven mechanistic study

Although SHAP analysis identifies which synthesis parameters mostly influence the PLQY proxy, it does not directly reveal how these parameters modulate NPL properties. To bridge this gap, we employed dynamic flow experiments (DFE)^29,46,47 to obtain continuous, time-resolved characterization during controlled perturbation of individual synthesis parameters. DFE utilizes in-situ optical characterization to capture high-density spectral time-series data as one or more process parameters are continuously varied. This approach has demonstrated rapid mapping of low-dimensional parameter spaces and is generalizable to other colloidal NC systems. More importantly, the nature of continuous in-situ characterization enables direct visualization of how NC optical properties evolve in real time under changing reaction conditions. Here, we focused on two input variables identified by SHAP as having the strongest influence on the PLQY proxy (i.e., Cs and In-HCl concentrations) and performed DFEs to mechanistically dissect their effect.

Figure 5 summarizes the results of dynamically varying the Cs concentration while maintaining the other NPL synthesis conditions the same as the champion NPLs. In this experiment, the Cs precursor volumetric flow rate continuously decreased from 60 to 25 μL min⁻¹ at a constant ramp rate of 5 μL min⁻², while all other parameters were fixed. Before the DFE starts, the platform runs for two residence times until it reaches steady state. Because the reactor introduces a time delay corresponding to the droplet residence time, two additional residence time is executed after the ramp to ensure that all DFE conditions are fully characterized⁴⁷. The evolving PL spectra collected throughout the ramp (Fig. 5a) reveal how Cs availability governs HDP NPLs’ emission behavior. At high Cs precursor volumetric flow rates (i.e., high concentrations), the PL spectra display a strong, stable emission peak with high intensity. As the Cs volumetric flow rate (concentration) is reduced, the emission intensity gradually decreases and disappears at low Cs volumetric flow rates ( < 30 μL min⁻¹), at which point a broad, weak emission related to molecular complexes dominates the PL spectra. The spectral features also undergo systematic shifts. Specifically, the emission peak position initially red-shifts as Cs volumetric flow rate decreases from 60 to 35 μL min⁻¹, followed by a rapid blue-shift at lower Cs levels (Fig. 5e). FWHM exhibits an increase from about 0.73 to 0.85 eV as Cs availability decreases (Fig. 5c), indicative of increasing heterogeneity and defect-related broadening. These observations are consistent with the mechanistic role of Cs in stabilizing the double perovskite lattice^44,45. Reduced Cs precursor supply increases the prevalence of A-site vacancies, CsCl-poor domains, and under-coordinated Cl environments⁴⁵. TEM-EDS analysis of the champion and low-Cs HDP NPLs (Supplementary Table 3) showed Cs atomic fractions of 20.43% and 15.90%, respectively, consistent with the presence of A-site vacancies in both samples and a greater degree of Cs deficiency in the low-Cs material. In agreement with this compositional analysis, the XRD patterns of the champion NPLs and the low-Cs sample remain consistent with the same double-perovskite phase, indicating that reduced Cs incorporation primarily introduces lattice defects rather than inducing a phase transformation. These structural and surface defects introduce nonradiative trap states and weaken the radiative self-trapped exciton emission^48,49. As a result, both PL intensity and PLQY become strongly dependent on Cs concentration. The dynamic trends captured by DFEs directly confirm the causal relationships inferred from SHAP analysis and provide real-time spectral evidence linking synthesis parameters to optical performance.

Fig. 5. Dynamic flow experimentation studies of Cs and In-HCl precursors.

a Evolving spectra of Cs and (b) In-HCl precursors. c Full width at half maximum (FWHM), (d) peak intensity and (e) peak position with different availability of Cs precursor. f FWHM, (g) peak intensity and (h) peak position with different availability of In-HCl precursor. The error bars indicate 95% confidence interval (CI) derived from measurement data. Source data are provided as a Source Data file.

Next, we conducted a DFE targeting the In-HCl precursor. As shown in Fig. 5f–h, the PL intensity exhibits a clear nonmonotonic dependence on In-HCl volumetric flow rate (concentration), increasing rapidly at low flow rates, reaching a maximum at intermediate values (35–40 μL min⁻¹), and decreasing at higher flow rates. The emission peak position follows a correlated trend, initially red-shifting with increasing In-HCl concentration before gradually blue-shifting at higher values, indicating changes in electronic structure and effective emissive domain size. Concurrently, the FWHM reaches a minimum at intermediate In-HCl concentrations and broadens toward both low and high In-HCl concentrations, reflecting increased energetic and structural disorder outside the optimal regime. These observations indicate that moderate In-HCl supply optimizes halide coordination and defect passivation, enhancing radiative recombination⁵⁰. At excessively high In-HCl, the reduced PL intensity and broader FWHM may reflect over-stoichiometric/ionically disordered growth environments (e.g., defect complexes, altered incorporation kinetics, or increased heterogeneity), which is a behavior broadly discussed for halide perovskites and perovskite NCs under strongly shifted chemical potentials^51,52. The variation in In content is further corroborated by TEM-EDS analysis (Supplementary Table 3), which shows an increase in the In atomic fraction with increasing In-HCl volumetric flow rate. Consistent with this trend, samples prepared under high In-HCl conditions exhibit XRD patterns that remain largely consistent with the double-perovskite phase, with no detectable secondary phases and only minor peak shifts, indicative of subtle structural variation rather than phase decomposition. In contrast, samples synthesized under low In-HCl conditions show a substantial fraction of secondary phases, which likely accounts for their inferior performance relative to the champion NPLs. The In-HCl DFE thus provides direct experimental validation of the SHAP-derived importance of halide-rich precursors and establishes a causal link between precursor chemistry and nanocrystal optical performance.

Discussion

In this work, we demonstrated PoLARIS, a modular microfluidic SDL that enabled autonomous synthesis, optimization, and mechanistic interrogation of HDP NPLs. By integrating programmable microfluidic synthesis, in-situ optical characterization, reproducible data provenance, and BO-driven closed-loop experimentation, PoLARIS efficiently navigated a high-dimensional, multi-element synthesis space that is otherwise intractable using conventional experimental approaches. Within a single autonomous campaign completed in less than half a day, the platform identified synthesis conditions yielding substantially improved photoluminescence performance while operating under realistic experimental constraints on precursor chemistry and reaction temperature.

Beyond optimization, PoLARIS established a direct link between autonomous discovery and mechanistic understanding. The trained Gaussian process surrogate model served as a data-driven digital twin of the synthesis process, capturing nonlinear composition-processing-property relationships and enabling quantitative interpretation through SHAP analysis. This analysis revealed that A-site stoichiometry, halide-rich precursor chemistry, and reaction temperature exerted the dominant influence on NPL optical performance, whereas ligand and secondary cation concentrations played more modest roles. Importantly, these statistical insights were not treated as endpoints. Instead, dynamic flow experimentation provided continuous, time-resolved spectroscopic evidence that causally connected key synthesis parameters to NPL formation pathways, defect formation, and emission characteristics. The agreement between SHAP-derived importance rankings and independently observed dynamic spectral trends validated the ability of PoLARIS to move autonomous materials discovery beyond black-box optimization toward mechanism-aware experimentation.

Collectively, these results establish PoLARIS as a general framework for accelerating the discovery and understanding of compositionally complex nanomaterials. Although demonstrated here for HDP NPLs, the underlying strategy, combining data-driven optimization, interpretable digital twins, and dynamic flow-based mechanistic probes, is foreseen to be broadly applicable to other multi-element and chemically rich materials systems, including high-entropy NCs. This work, therefore, illustrated how microfluidic SDLs can unify efficiency, reproducibility, and mechanistic insight, advancing autonomous materials research into regimes of complexity that have remained largely inaccessible to traditional experimentation.

Methods

Materials

Cesium carbonate (Cs₂CO₃, ≥ 99.9% trace metal basis), silver acetate (Ag(ac), ≥ 99.99% trace metal basis), sodium acetate (Na(ac), ≥ 99%), indium acetate (In(ac)₃, ≥ 99.99% trace metal basis), bismuth acetate (Bi(ac)₃, ≥ 99.99% trace metal basis), hydrochloride acid (HCl, 37%), oleic acid (OA, technical grade, 90%), oleylamine (OAm, technical grade, 70%), 1-octadecene (ODE, technical grade, 90%), and anhydrous toluene (99.8%) were purchased from Sigma Aldrich and used as received. Perfluorinated oil (PFO, 97%) was purchased from Kurt J. Lesker Company and used as received. PFO was selected as the carrier phase due to its chemical inertness, immiscibility with the reaction solvent, thermal stability, and reliable droplet segmentation behavior.

Cesium precursor

0.1 mmol Cs₂CO₃, 1 mL OA, and 14 mL ODE were loaded into a 60 mL round bottom three-neck flask (Supplementary Tables 4–6). The mixture was vacuumed for 30 min at 120 °C until the solution was transparent. The solution was purged with nitrogen until it cooled down to room temperature.

Silver precursor

0.1 mmol Ag(ac), 0.15 mL OA, 0.3 mL OAm, and 29.55 mL ODE were loaded into a 60 mL round bottom three-neck flask (Supplementary Tables 4–6). The mixture was vacuumed for 30 min at 120 °C until the solution was transparent. The solution was purged with nitrogen until it cooled down to room temperature.

Sodium precursor

0.1 mmol Na(ac), 2 mL OA, and 58 mL ODE were loaded into a 100 mL round bottom three-neck flask (Supplementary Tables 4–6). The mixture was vacuumed for 30 min at 120 °C until the solution was transparent. The solution was purged with nitrogen until it cooled down to room temperature.

Indium precursor

0.2 mmol In(ac)₃, 1 mL OA, 0.6 mL OAm, and 18.4 mL ODE were loaded into a 60 mL round-bottom three-neck flask (Supplementary Tables 4–6). The mixture was vacuumed for 30 min at 120 °C until the solution was transparent. Then 0.1 mL HCl was added to the solution. The solution was further vacuumed for 30 min at 120 °C until it was transparent. The solution was purged with nitrogen until it cooled down to room temperature. This resulted in a molar ratio of approximately 1:6 (In:Cl). The HCl serves a dual purpose ensuring the complete dissolution of the indium precursor into a transparent complex and providing an additional chloride source.

Bismuth precursor

0.1 mmol Bi(ac)₃, 20 mL OA, and 150 mL ODE were loaded into a 250 mL round bottom three-neck flask (Supplementary Tables 4–6). The mixture was vacuumed for 30 min at 120 °C until the solution was transparent. The solution was purged with nitrogen until it cooled down to room temperature.

HDP NPL purification method

HDP NPL samples were first separated from the crude solution by centrifugation at 6652 × g (8000 rpm) for 5 min. The precipitation was then dispersed in anhydrous toluene, while the supernatant was discarded. Then, the dispersed sample was centrifuged 6652 × g for 5 min again. The precipitation was collected and re-dispersed in anhydrous toluene for further characterization.

Off-line characterization

Crystal structure was characterized by X-ray diffraction using a X’Pert-Pro (Cu K source, 1.54 Å, 40 mA, 45 kV) and Rigaku SmartLab X-Ray Diffractometer (Cu K source, 1.54 Å, 48 mA, 45 kV). TEM images and TEM EDS data were obtained by FEI Talos F200X (200 kV).

Hardware

Automated syringe pumps (Fusion 6000, Chemyx) equipped with stainless steel syringes were used to deliver all fluids, including six 20 mL syringes for precursor solutions and one 100 mL syringe for the PFO carrier phase (Chemyx). The microfluidic platform incorporated a multiport PEEK manifold with a 0.04-inch thru hole (P-150, IDEX) and a low-pressure PEEK T-junction with a 0.04-inch thru hole (P-714, IDEX). The reactor consisted of a CNC-machined aluminum heating plate containing a spiral microchannel with a diameter of 1/16 inch. Heating was provided by five 100 W cartridge heaters (Watlow), and temperature regulation achieved using a PID controller (Watlow). In-situ optical measurements were performed using a custom three-port flow cell coupled to a fiber-coupled broadband light source (BAL 2000, Ocean Insight), a fiber-coupled light-emitting diode (365 nm, Thorlabs), and two miniature UV-Vis spectrometers (HDX and QEpro, Ocean Insight). All hardware components, including syringe pumps, temperature controller, UV-Vis light sources, light-emitting diode, and spectrometers, were centrally controlled and synchronized using LabVIEW software (2024 Q1, National Instruments). In a typical experiment, the LabVIEW control program first imports a condition file in CSV format containing the prescribed volumetric flow rates for each syringe pump together with the target reactor temperature for each experimental condition. The microfluidic platform then enters an equilibration stage, during which the reactor is allowed to reach the specified temperature with minimal control error, as defined by a low mean-squared deviation from the setpoint. Once thermal equilibration is achieved, the system undergoes an additional waiting period, typically corresponding to two residence times, to ensure that the new operating condition has fully propagated through the reactor and that the flow system has reached steady state. Only after this steady-state criterion is satisfied does the platform initiate in-situ data acquisition, thereby ensuring that all recorded measurements correspond to well-defined and stable reaction conditions.

PFO recovery and reuse

To improve the sustainability and reduce the consumption of PFO, the carrier phase was recovered, cleaned, and reused following sample collection. After completion of each HDP NPL synthesis experiment, the collected output was transferred to a clean collection vessel and allowed to rest undisturbed to enable gravity-driven phase separation. Once the two phases (PFO and ODE) were clearly separated, the PFO layer was carefully withdrawn from the collection vessel. The recovered PFO was then purified using a two-pass syringe filtration procedure. Specifically, the isolated PFO was pushed through a syringe filter (Filter 1, Whatman GD/X 13) into a clean receiving container. The collected PFO was then passed through a second fresh syringe filter (Filter 2, Whatman GD/X 13) to further remove any residual particulates. After these two sequential filtration steps, the PFO was considered suitable for reuse in subsequent studies.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

M.A. and J.L. conceived the project. J.L., J.X., F.D.L., and M.A. designed the algorithms. J.L., J.X., S.S., and F.D.L. programmed the algorithms. J.L. built the microfluidic platform and conducted the data analysis and visualization. J.L., H.P., and N.M. conducted the investigations under the advisement of M.A. J.L., Z.L., and O.C. designed the precursor chemistry. M.A. and O.C. acquired funding and directed the project. J.L. and M.A. drafted the manuscript. All authors provided feedback on the manuscript.

Peer review

Peer review information

Nature Communications thanks Eline Hutter, Loreta A. Muscarella, and the other, anonymous, reviewer for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data that support the findings of this study are available from the corresponding author upon request. Source data are provided with this paper.

Code availability

Codes used in this study, including the source code for the data processing, autonomous experimentation and digital twin model, are available from Zenodo⁵³ and from the corresponding author upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-72765-2.