Solvent Screening for Separation Processes Using Machine Learning and High-Throughput Technologies

Abstract

As the chemical industry shifts toward sustainable practices, there is a growing initiative to replace conventional fossil-derived solvents with environmentally friendly alternatives such as ionic liquids (ILs) and deep eutectic solvents (DESs). Artificial intelligence (AI) plays a key role in the discovery and design of novel solvents and the development of green processes. This review explores the latest advancements in AI-assisted solvent screening with a specific focus on machine learning (ML) models for physicochemical property prediction and separation process design. Additionally, this paper highlights recent progress in the development of automated high-throughput (HT) platforms for solvent screening. Finally, this paper discusses the challenges and prospects of ML-driven HT strategies for green solvent design and optimization. To this end, this review provides key insights to advance solvent screening strategies for future chemical and separation processes.

1. Introduction

Separation processes play an integral role in the production of a diverse range of products across the oil and gas, food, pharmaceutical, and chemical industries. Traditionally, these processes utilize large-scale physical and thermal treatments, such as centrifugation, filtration, and distillation, for the removal of various impurities and byproducts. However, such processes are energetically intensive and costly, accounting for approximately 10–15% of global energy consumption.¹ Solvent extraction, otherwise known as liquid–liquid extraction (LLE), is a well-established operation for the effective separation of a wide range of synthetic and bio-based compounds from liquid mixtures. By leveraging the differential affinity of target compounds across immiscible liquid phases, solvent extraction can achieve separation with a high extraction efficiency and selectivity. Additionally, solvent extraction offers a simpler process setup and reduced energy costs, positioning it as an attractive strategy for industrial-scale application.

Despite these advantages, the efficiency of an extraction process significantly hinges on the selection of the appropriate solvent. The optimal extractant solvent should exhibit a high distribution coefficient, low mutual solubility, low toxicity, high chemical stability, and low cost.² Most industrial extraction processes rely on conventional organic solvents, owing to their aqueous immiscibility and relatively low bulk costs. However, the environmental concerns regarding their nonrenewable, fossil-fuel origins, as well as their potentially high toxicity and volatility, have prompted the search for greener and safer alternatives.³ In recent years, ionic liquids (ILs) and deep eutectic solvents (DESs) have emerged as potential “green” solvent alternatives, owing to their tunable properties and proposed environmental benefits.^4,5 Specifically, ILs are a class of salts composed of cations and anions with melting points below 100 °C, typically characterized by low volatility, minimal flammability, and high thermal stability. DESs are a neoteric class of solvents derived from eutectic mixtures of hydrogen bond donors (HBDs) and hydrogen bond acceptors (HBAs), which are reported to offer several key advantages including simpler synthesis, lower cost, reduced toxicity, and enhanced biodegradability.⁶⁻⁸ The structures and properties of both ILs and DESs are highly customizable due to the vast number of possible cation/anions and HBA/HBD combinations.⁹ While this flexibility enables their tailored use in a wide range of applications, it also introduces variability in the understanding of their structure–property relationships, often requiring case-by-case experimentation and validation, especially in more complex processes such as chemical and solvent-based separations.¹⁰ As a result, their effectiveness in industrial-scale processes and their long-term environmental sustainability remain subjects of ongoing debate.^11,12 Given the broad range of possible IL and DES structures, along with the time-consuming and costly nature of conventional experimentation, there is a growing demand for faster, more efficient strategies for green solvent screening and design.

To accelerate the solvent screening process, researchers have turned to the use of in silico methods for property estimation. Several useful methods have been broadly utilized in the past decades, including molecular-scale modeling, macroscale modeling, and empirical/semiempirical methods.¹³ Molecular-scale modeling, such as Density Functional Theory (DFT) and classical molecular dynamics, describes atomic bonding and elucidates the interactions of particles (atoms or electrons) at molecular levels. These methods have offered valuable insights into the fundamental molecular behavior, aiding researchers in the synthesis of new materials and the understanding of solvent–solute interactions.¹³⁻¹⁵ In the context of solvent-based separation processes, a notable example is the study by Sun et al., which employed DFT calculations to analyze the extraction mechanism of metal ions in a tributyl phosphate (TBP)–FeCl₃/brine system for lithium recovery.¹⁶ Their DFT results revealed that the electrostatic interactions between the solvent and central metal ions were the driving force behind the extraction and co-extraction processes. Another recent example is the work by Kaim-Sevalneva et al., which explored the solvent extraction of scandium(III) from rare earth elements using quaternary ammonium-based ILs.¹⁷ By applying DFT methods, they were able to validate the neutral exchange mechanism behind the selective extraction of scandium(III). Similarly, other studies have combined experimental work and DFT or classical molecular dynamics simulations to elucidate the molecular mechanisms underlying various extraction processes.¹⁸⁻²¹ Macroscale modeling, also known as thermodynamic modeling, utilizes mathematical relations based on thermodynamic principles to describe bulk system properties, such as phase transitions and heat transfer. One well-established model is COSMO-RS (Conductor-like Screening Model for Realistic Solvation), widely recognized for its effectiveness in estimating a variety of bulk solvent properties, such as solubilities, partition coefficients, and vapor–liquid equilibria.²² A notable example is the study by Filly et al., which utilized COSMO-RS to estimate the σ-profiles of aroma compounds within eight organic solvents, revealing ethyl acetate and dimethyl carbonate as more efficient alternatives to n-hexane for aroma extraction from caraway seeds.²³ Similarly, Wojeicchowski et al. developed a range of hydrophobic DESs specifically for extracting carnosic acid and carnosols from rosemary by utilizing COSMO-RS to determine the optimal pairings of hydrogen bond acceptors and donors.²⁴ The readers can refer to several review articles^13,14,25 regarding more applications of COSMO-RS methods in solvent screening. Moreover, empirical/semiempirical methods blend theoretical and empirical data to estimate molecular properties. GC-based methodologies effectively bridge molecular-scale and macroscale contexts by employing a semiempirical, additive framework. These approaches decompose molecules into functional groups, assign quantitative parameters to each, and sum these values to estimate the system properties. The UNIFAC (Universal Quasi-Chemical Functional Group Activity Coefficient) and UNIQUAC (Universal Quasi-Chemical Activity Coefficient) methods have been widely employed in estimating the properties of a variety of solvents. Birajdar et al., for instance, applied an optimized UNIFAC model to predict the distribution coefficients of bio-based 2,3-butanediol in various organic solvents from aqueous fermentation broth. Their model successfully identified 1-butanol and 2-secondary butyl phenol as the optimal extractant solvents for 2,3-butanediol recovery.²⁶ In addition, the hybrid method, such as the integration of COSMO-RS and GC into the GC-COSMO method,^27,28 shows higher efficiency and reliability. Researchers have also leveraged the combination of GC-based methods with other traditional property estimation methods for solvent screening. For instance, Mu and Gmehling first proposed a combined GC-COSMO model, integrating GC methods with the COSMO-RS and COSMO-SAC models, to enhance their predictive accuracy.^29,30 A study by Dong et al. employed a combined COSMO-SAC-UNIFAC model to predict phase equilibrium data in IL systems.³¹ Their results indicated that the combined model provided relatively accurate predictions, even in the absence of certain UNIFAC group binary parameters, suggesting its potential utility in data sets with missing parameters. In a subsequent study, Zhu et al. incorporated 648 new vacant parameter pairs for 51 main functional groups into the COSMO-SAC-UNIFAC model to expand the model’s applicability to a broader range of IL systems.³² In doing so, their model produced more accurate predictions of vapor–liquid equilibrium and liquid–liquid equilibrium data for 16 binary and 2 ternary systems, outperforming both the individual COSMO-SAC and UNIFAC models. Additionally, Liu et al. developed a GC-COSMO model to suggest new reaction kinetics for reaction solvents in organic synthesis and validated their model in the selected Menschutkin reaction.³³ Peng et al. built an extended GC-COSMO model to estimate the σ-profile and cavity volume of ionic liquid for separation processes and tested their method in the extraction of benzene from cyclohexane and postcombustion carbon capture.³⁴ Despite their widespread application, these traditional methods for property estimation still face several inherent challenges, including high computational costs and slow processing times.¹³ Furthermore, these methods heavily rely on theoretical frameworks, which may fall short of accurately capturing real-world complexities and nuances observed in actual data. Thus, more advanced, real-world, data-driven models are highly desirable.

Machine learning (ML) offers a promising approach for solvent screening, leveraging mathematical relations to analyze patterns and unveil connections within large and complex data sets. In the context of solvent screening, ML models can be employed to rapidly analyze and predict solvent–solute interactions, optimize recovery yields, and enhance selectivity for a wide range of chemical and separation processes.³⁵⁻³⁸ Additionally, ML models can “learn” from new experimental and/or computational data, drive self-optimization, and as a result potentially surpass traditional property estimation methods in both accuracy and efficiency, for effective solvent screening.^39,40 This review explores the latest advancements in AI-assisted solvent screening with a specific focus on ML models for solvent physicochemical property prediction and separation process parameter design. Additionally, this review highlights recent progress in the development of automated high-throughput (HT) platforms for solvent screening and investigation. Furthermore, the review addresses the challenges and future opportunities for integrating ML-driven HT approaches in the screening, design, and optimization of green solvents (Figure 1).

Figure 1.

Machine-learning-driven framework for high-throughput solvent screening.

2. MACHINE LEARNING MODELING

The integration of ML in chemical engineering endeavors is not a novel concept. As far back as early 1980s, researchers had already crafted an expert system designed to forecast the physical properties of intricate fluid mixtures.⁴¹ Subsequently, the swift adoption of neural network methodologies revealed their intrinsic ability to extract insights from vast data sets, consequently enhancing the precision of modeling for complex systems. Nonetheless, the initial application of ML encountered significant hurdles, including insufficient data, limited data accessibility and inadequate computational resources.⁴¹ These limitations hindered ML from realizing its anticipated revolutionary impact. However, over the past few decades, the introduction of crucial technologies, including more complex deep learning architectures such as convolutional neural nets (CNNs) and recurrent neural networks (RNN), advanced parametrization methods such as reinforcement learning and transfer learning, as well as hardware advances such as GPU computing, has the capability to address more complicated “big data” domain problems.^42,43 Notably, the recent advent of Generative Pretrained Transformer (GPT) models, which develop specialized large-scale language models (LLMs), has offered an accessible platform that enhances text mining and complex data analysis.⁴⁴ This advancement significantly speeds up the traditional research methodology, marking a pivotal shift in how data is utilized and understood in various fields.⁴⁵ In addition, a deep generative model such as stable diffusion and transformer has also emerged as a new powerful family of AI with record-breaking performance in graph-based applications, language models, and beyond.⁴⁶⁻⁴⁹

2.1. ML Solvent Screening Workflow

In the context of solvent screening, the ML process requires a balanced integration of chemistry and chemical engineering knowledge with computer science methodologies. This integrated ML process operates through a systematic workflow, resulting in the development of algorithms designed to identify patterns from data and generalize those patterns to make predictions or decisions based on unseen data. This can be distilled into three key stages: data collection and preprocessing, model development, and model deployment.

In the initial stage, researchers compile relevant empirical and/or computational data, typically obtained from previous literature and/or databases such as QM9,⁵⁰ GDB-17,⁵¹ PubChem,⁵² ChEMBL,⁵³ PDBbind,⁵⁴ and ChemSpider,⁵⁵ to serve as the input data set for a given ML model. Additionally, databases such as NIST Chemistry WebBook,⁵⁶ AqSolDB,⁵⁷ Dortmund Data Bank,⁵⁸ and Cambridge Structural Database⁵⁹ are widely used repositories that provide key molecular data that are directly relevant to solvent screening and solvent-based separations, including solvent properties, aqueous solubilities, phase equilibria, molecular bonding interactions and structural details. The obtained data set must be machine-readable in order to be incorporated into ML models. Researchers can employ various methods to convert and display solvent data for computational analyses. One common approach involves the use of string-based representations, such as International Chemical Identifier (InChI), MDL Molfiles, and Chemical Markup Language (CML), which utilize sequences of letters, numbers, and/or symbols to depict molecular structures in a text-based format suitable for computational processing. Notably, the Simplified Molecular Input Line Entry System (SMILES) notation and its derivatives, such as SMILES Arbitrary Target Specification (SMARTS), are among the most widely adopted string-based methods for the representation of a variety of organic and inorganic compounds. It has enabled the processing of several complex solvent structures using ML models, including regression and deep learning algorithms.⁶⁰⁻⁶² Alternatively, graph-based representations and molecular-fingerprint-based representations have also been employed to represent the molecular structures of numerous solvents and compounds for ML models. Each of these notational methods offers unique advantages in terms of detail, specificity, and compatibility with computational tools and, therefore, should be guided by the specific requirements of the data set in question.⁶³ The readers can refer to previous works for detailed discussions on different fingerprints or molecular features.^64,65

In addition, solvent data may be represented through property-based features. These features can serve as supplemental input data for ML models to further enhance the accuracy of the solvent predictions. For example, researchers can generate specific structural features, known as structural descriptors, from solvent data, such as bond length, bond angle, and molecular weight, typically through numerical, vector, or matrix configurations.^66,67 Notably, a study by Aghaie and Zendehboudi demonstrated the prediction of CO₂ solubility in ILs using four distinct ML models.⁶⁸ These ML models were fed with two separate CO₂ solubility data sets, which were characterized by either thermodynamic properties or structural descriptors. Interestingly, the ML models trained using the structural descriptor data set generally exhibited higher predictive accuracies when compared to those based on the thermodynamic properties, suggesting a strong relationship between anionic and cationic structural features for dictating CO₂ solubility.

Solvent data may be further characterized by their distinct chemical and physical characteristics, which are otherwise known as physicochemical descriptors. These descriptors, which encompass electron charge, electron distribution, and solvation energies, are frequently computed by using thermodynamic and quantum chemistry based property estimation methods. A noteworthy example is the study by Amar et al., which utilized a hybridized ML solvent screening approach to identify optimal solvents for asymmetric catalysis.⁶⁹ In particular, Amar et al. utilized a library of 459 solvents, calculating 12 conventional molecular descriptors, two reaction-specific descriptors, and additional descriptors based on screening charge density. Cross-validation analyses revealed that the combination of both physicochemical descriptors and screening charge density descriptors resulted in enhanced accuracy for predicting the reactive conversion in solvents. Though, it is essential to note that most raw solvent data, along with structural and/or physicochemical descriptors, may still present irregularities such as missing values, outliers, or inconsistent units, which can significantly hinder the overall performance of a constructed ML model. Hence, researchers typically subject raw data sets to preprocessing, which includes data cleaning, scaling, and/or normalization.

Given the inherent complexity of molecular properties, no single molecular descriptor or fingerprint is capable of fully characterizing all aspects of these properties. Thus, the selection of molecular representation is often problem dependent. The scalar descriptors (also referred to as zero-dimensional descriptors) contain very general information on molecules, such as molecular weight and number of hydrogens, but lack structural information in detail. Since they are usually easy and quick to compute, they are often considered together with other descriptors in practice. The one-dimensional descriptors, such as fingerprints, contain a string-like vector with substructural information on a particular pattern (e.g., whether a chemical moiety exists). They are also often easily computed and may be sufficient when the targeting property is highly correlated to a specific substructure (e.g., acidity from acid groups). The two-dimensional descriptors offer more detailed structural information, such as connectivity, symmetry, and shape, and are mostly computationally feasible within modern cheminformatics. Therefore, they have been widely used with machine learning.^70,71 The three-dimensional descriptors are typically derived from the spatial coordinates of atoms within a molecule, providing comprehensive geometric information about its structure. However, it is usually time-consuming to calculate these descriptors due to their inherent complexity. Nonetheless, the three-dimensional representation must be employed in certain cases, such as atomic force evaluation. The reader can refer to other articles that comprehensively reviewed molecular representations.⁷²⁻⁷⁵

In the second stage, researchers construct and train an ML model using the preprocessed input data. Generally, ML models can be classified into three main categories: supervised learning, unsupervised learning, and reinforcement learning.^42,76 Supervised learning involves training an ML algorithm on a labeled input data set with the goal of predicting a labeled output or target variable. This includes algorithms such as Linear Regressions (LR), Logistic Regression (LogReg), Artificial Neural Networks (ANNs), Support Vector Machines (SVMs)/Support Vector Regressions (SVRs), Decision Trees (DTs), and Random Forest (RF). The parametrization of these models typically aims to map the correlation between the solvent descriptors as input and the targeting property (e.g., solubility) as output, enabling a predictive evaluation of unseen solvents. The readers can find more comprehensive explanations of these algorithms in previous publications.⁷⁶⁻⁷⁹ In contrast, unsupervised learning involves training an ML algorithm on data without labeled responses to infer the inherent structure or pattern present within a set of data points. Unsupervised learning algorithms, such as K-Means Clustering, Hierarchical Clustering, Principal Component Analysis (PCA), and Association Rule Learning (ARL)/Association Rule Mining (ARM), aim to explore large data sets and freely identify inherent patterns typically through clustering analyses. These models reveal the underlying patterns of a group of solvents and allow for efficient evaluation of new solvents that exhibit desired properties. Several previous articles have provided more explanations regarding these unsupervised learning methods.^76,80−83 The third category of ML models is reinforcement learning, in which an agent learns to form decisions through feedback within a given system or environment. Examples include Q-Learning, Deep Q-Networks, and Monte Carlo Tree Search, which aim to uncover the optimal behavior or patterns within a data set to obtain the maximum reward or output.

It is important to note that each of these ML algorithms possess unique strengths and limitations, particularly in the context of solvent screening and separation process design.^84,85 For instance, ANNs are among the most popular ML models used in these applications.⁸⁶ ANNs are composed of multiple layers of processing units, or “nodes”, which receive data inputs, compute weighted values based on these inputs, and generate an output value.⁸⁷ These values are then passed as inputs to the subsequent layer of nodes, progressing through all layers in the network until reaching the final output layer, resulting in the final prediction. ANNs are particularly valuable for identifying complex, nonlinear relationships within intricate data sets such as molecular structural data and real-time separation process data.⁸⁶ However, this advantage comes with its trade-offs, as ANNs often require large training data sets, which can lead to increased training time and computational resource demands.⁸⁸ Additionally, ANN models can be susceptible to data overfitting if they are overtrained or have too many layers.⁸⁹ SVMs are another widely used ML model in solvent screening. Specifically, these models represent data points as vectors in space and aim to identify the optimal boundary or hyperplane that maximizes the margin between distinct groups of vectors. By utilizing quadratic functions, SVMs can effectively model both linear and nonlinear interactions, only requiring small training data sets, which can enable quicker analyses and enhance generalization.⁹⁰ This performance of SVMs, however, can be compromised when applied to imbalanced data sets, as these models tend to bias toward the majority class of data points that are present.⁹¹ Besides, LR is relatively simple and interpretable, but it is also sensitive to multicollinearity. PCA is more data-driven with less user-dependent parameters to tune but may show a loss of interpretability with new variables created. Q-learning is adaptable to learn optimal results through trial and error, but it is sensitive to hyperparameters and may also show convergence issues. Readers can refer to earlier articles for further details on ML model types.⁹² To this end, researchers must select the appropriate model (or combination of models) and descriptors based on the specific objectives of their tasks as well as the size and complexity of the respective data set.

The final stage of the ML process involves deployment of the trained ML model and evaluation of its predictive performance. The performance evaluation of a ML model is crucial and is conducted using various metrics and statistical techniques tailored to the specific task at hand, such as regression, classification, or clustering. In the context of solvent screening, where the goal is often to quantify complex structure–property relationships and optimize process parameters, regression models are predominantly used due to their capability to provide quantitative evaluation of a given solvent property. To gauge the performance of such models, researchers rely on key statistical metrics, such as the coefficient of determination (R²) and the root-mean-square error (RMSE), which serve as comparative benchmarks regarding accuracy and precision. More specifically, R² is a statistical measure that represents the proportion of variation in a dependent variable that can be predicted by an independent variable. R² is mathematically expressed as

1

where y_i is the actual value, ŷ_i is the predicted value, and y̅_i is the mean of the actual values. The RMSE measures the average difference between a model’s predicted and actual values. The RMSE is mathematically expressed as

2

where n is the number of data points, y_i is the actual value, and ŷ_i is the predicted value. When assessing the performance of different ML models for solvent screening applications, an R² score close to 1 alongside a low RMSE is greatly desired, indicating a high degree of predictive accuracy and precision. It is also worthy to note that the predictive capability of a trained ML model should be evaluated based on not only its own testing or validation data sets but also unseen external data from other resources. Readers can refer to educational articles for more discussions on developing ML models.⁹³

2.2. Prediction of Solvent Physicochemical Properties via ML Models

A thorough understanding of a solvent’s physicochemical properties is crucial for its effective use in chemical and separation processes. As highlighted earlier, the efficacy of a chemical or separation process is contingent on several factors, such as the selection of a solvent system and the resulting interactions with targeted compounds or reagents. While previous experimental studies have documented numerous solvents and their physicochemical properties, the vast array of untested solvent structures and solvent mixtures, compounded with the immense number of potential applications, renders the trial-and-error methodology of conventional experimentation impractical. To solve this constraint, researchers have utilized ML algorithms to rapidly predict the physicochemical properties of organic and ionic solvents, as summarized in Table 1.

Table 1. Summary of Studies Using ML Models for the Prediction of Solvent Physicochemical Properties.

Solvent	Property	Descriptors	No. of data points (training/test)	ML model	R2	Ref.
Organic	Solubility	Structural descriptors, quantum chemical descriptors	Water = 805/95, Water = 445/115, Ethanol = 553/142, Benzene = 370/94, Acetone = 360/92	MLR	0.64	(94)
				PLS	0.64
				ANN	0.67
				SVM	0.71
				GP	0.70
	Polarity	Quantitative structure, DFT descriptors	378/53	RF	0.90	(60)
				ANN	0.97
	Polarity	Structural, quantum chemical and thermodynamic descriptors	378/43	MLR	0.67	(95)
				PLS	0.66
				KNN	0.81
				SVM	0.86
				ET	0.95
				BR	0.90
				RF	0.93
				NN	0.96
	Polarity	Quantitative structure, PaDEL descriptors, PaDEL fingerprints	378/53	NN	0.97	(96)
	Reactivity (conversion; diastereomeric excess)	Screening charge density, COSMOtherm descriptors	25/9	GP	Accuracy = 0.94	(69)
	Gibbs free energy of solvation	RDKit structural descriptors	513/64 and 4477/559	RF	0.99 (RMSE = 0.5 kcal/mol)	(67)
				GB	0.99 (RMSE = 0.43 kcal/mol)
	Gibbs free energy of solvation	Morgan structural descriptors encoded in fixed-size feature vectors	1996/249 and 514/64	ANN	0.96 (RMSE = 0.24 kcal/mol)	(97)
	Gibbs free energy of solvation	Molecular graph	5615/623	Message passing NN	0.98 (MAE = 0.16 kcal/mol)	(164)
	Gibbs free energy of solvation	Molecular graph	10145 (total)	Message passing NN	No R2 reported (RMSE = 0.44 kcal/mol, MAE = 0.21 kcal/mol)	(165)
	Gibbs free energy of solvation	Molecular graph	Solute = 8366 (total), Solvation free energy = 20253 (total), Solvation enthalpy = 6322 (total)	Message passing NN	0.93 (RMSE = 1.32 kcal/mol, MAE = 0.91 kcal/mol)	(166)
	Gibbs free energy of solvation	Structural descriptors, thermodynamic descriptors	Self-solvation free energy = 211 (total)	SVR	(RMSE = 0.3 kcal/mol, MAE = 0.2 kcal/mol)	(167)
IL-organic mixtures	Surface tension and viscosity	Group contribution	Surface tension = 2430/511; Viscosity = 19872/4282	ANN	0.97; 0.96	(103)
				XGB	0.98; 0.86
				LightGBM	0.92; 0.82
	Heat capacity and density	Group contribution	Heat capacity = 1601/353; Density = 24282/5335	ANN	0.99; 0.99	(104)
				XGB	0.99; 0.98
				LightGBM	0.99; 0.98
IL	CO2 solubility	Thermodynamic properties, structural descriptors	1655 (total)	DT	0.94; 0.96	(68)
				RF	0.96; 0.98
				LSSVM	0.75; 0.76
				MLR	0.55; 0.28
	CO2 solubility	Group contribution, temperature, pressure	8093/2023	ANN	0.98	(105)
				SVM	0.98
	CO2 solubility	Structural (experimental)	10116 (total)	ANN	>0.98 (RMSE < 0.025)	(15)
	Toxicity	Electrostatic potential surface area, charge distribution area	99/20	ELM	0.94	(106)
				MLR	0.87
				SVM	0.89
	Toxicity	Quantitative structure (SMILES), structural descriptors	284/71	ANN	0.89	(99)
				SVM	0.92
	Toxicity	Quantitative structure (SMILES), group contribution, structural descriptors	284/71	MLR	0.61	(100)
				MLR-EN	0.76
				ANN	0.65
	Water solubility	COSMO-RS solubilities, PM3 descriptors	16137 (total)	ARM	N/A	(107)
				DT	Accuracy = 0.93
				Multilayer NN (DL)	0.84; 0.99
	Melting temperature	Quantitative structure (SMILES), structural and physicochemical descriptors	1253 (total)	RNN (DL)	0.90	(101)
	Melting temperature	Thermodynamic and structural descriptors	2212 (total)	RF	0.67 (RMSE = 44, MAE = 14)	(168)
	Viscosity	Group contribution, quantitative structure	15251 (total)	NLR	0.96	(102)
				SVM	0.99
	Conductivity	Structural descriptors	IL-split = 324/40; Data-point split = 3406/425	ANN	0.96 (RMSE = 1.63)	(169)
	Conductivity	Structural descriptors	1190/132	SVR	0.99 (RMSE = 0.15, MAE = 0.06)	(170)
				ANN	0.99 (RMSE = 0.02, MAE = 0.09)
	Conductivity	Structural descriptors	2582/286	MLR	0.853 (RMSE = 0.322, MAE = 0.204)	(171)
				RF	0.96 (RMSE = 0.16, MAE = 0.09)
				XGB	0.99 (RMSE = 0.09, MAE = 0.05)
DES	Surface tension	COSMO-RS descriptors (Sσ-profiles)	1084/487	ANN	0.98	(109)
	Viscosity	Basic properties (molar mass, ratio, temperature); content of H2O in DES (WH2O); Morgan fingerprint of HBA/HBD	795/198	SVR	0.93	(110)
				RF	0.96
				NN	0.98
				XGB	0.99
	Viscosity	COSMO-RS derived descriptors	3464/1484	CatBoost	0.99 (RMSE = 0.21 mPa·s)	(172)
	Density	Group contribution	1058/352	MLPANN	0.99	(111)
				LSSVM	0.99
	Density	Group contribution, temperature, critical pressure and temperature, acentric factor	1053/186	LSSVR	0.99	(117)
				MLP	0.98
				CFF	0.97
				GR	0.95
				RBF	0.89
				RNN	0.89
				ANFIS	0.97
	pH	Quantitative structure (SMILES), COSMO-RS descriptors (Sσ-profiles)	473/175	MLR	0.99; 0.99	(113)
				ANN	0.99
	Eutectic temperature and eutectic composition	Quantitative structure (SMILES), COSMO-RS descriptors	1318/329	GBR	0.68; 0.04	(112)
				RFR	0.43; −0.21
				SVR	0.74; 0.34
				KNN	0.56; 0.03
	CO2, CO, CH4, H2 and N2 gas solubility	Thermodynamic descriptors	Not specified	Hybrid SVM/LSTM	No R2 reported (RMSE ranging from 1.6 × 10–4 to 9.8 × 10–4)	(173)
	CO2 solubility	Thermodynamic and structural descriptors	711/300	RF	0.95 (RMSE = 0.22)	(174)
	CO2 solubility	COSMO-RS descriptors	2055/272	ANN	0.99 (RMSE = 7.08 g/kg)	(175)
	CO2 solubility	COSMO-RS descriptors	1084/889	ANN	0.99 (RMSE = 0.13, MAE = 0.09)	(176)
	CO2 absorption capability	Structural descriptors	DES = 1971 (total); IL = 9615 (total)	RF	0.98 (RMSE = 0.10, MAE = 0.06)	(177)
				ANN	0.97 (RMSE = 0.10, MAE = 0.08)
	SO2 adsorption capability	Thermodynamic descriptors	384/96	ANN	0.98 (RMSE = 0.001)	(178)
	Cobalt solubility	Structural and thermodynamic descriptors, experimental condition	Cathode solubility = 592 (total); Metal oxide solubility = 199 (total)	XGB	0.91 (MSE = 0.004)	(179)

In the context of organic solvent screening, a prominent study by Boobier et al. successfully demonstrated the application of ML for the prediction of organic solvent and water solubilities.⁹⁴ By leveraging a curated set of 14 molecular descriptors, the ExtraTrees (ET) model achieved high predictive accuracy, significantly outperforming conventional in silico prediction tools including AquaSol, EPI Suite, and COSMOtherm. Another notable example is the study by Saini and Kumar, which involved the development of an ANN model with four quantum chemical descriptors, one topological descriptor, and one categorical descriptor to predict the polarity values of various organic solvents.⁹⁵ Despite the high predictive accuracy of their model, the reliance on proprietary software for calculating the quantum chemical descriptors could restrict its broader applicability. To address this limitation, Saini and Singh streamlined their approach in a later study by implementing feature importance and reduction techniques.⁹⁶ Their refined ANN model eliminated the need for categorical descriptors while maintaining high predictive performance, making it a simpler, quicker, and more effective tool for ML-based organic solvent screening.

In addition to polarity, predicting the Gibbs free energy of solvation (ΔG_sol) for organic solvents, an important but challenging task, has also been achieved via the ML approach. For example, Ferraz-Caetano et al. used experimentally measured ΔG_sol along with structural and electronic descriptors, such as chi connectivity indices, topological polar surface area, volume surface area state, molecular weight, molar refractivity, and electrotopological state atom, to parametrize several regression models including RF, gradient boosting (GB), SVM, and ANN.⁶⁷ They identified RF and GB as the top two models that can reach improved performances for an RMSE of 0.42 kcal/mol. Similarly, Lim and Jung developed an ANN model for ΔG_sol of organic solvents using structural descriptors encoded into feature vectors, and they benchmarked the model’s performance against other theories.⁹⁷ They achieved a MAE of 0.24 kcal/mol for nonaqueous systems, and their prediction displays comparable results to several quantum-mechanism-based solvation models. Later, they further improved their modeling based on the group-contribution method by considering pairwise atomic interactions in the atomic feature vectors.⁹⁸ This led to a ML model exhibiting the potential to offer more physicochemical insights on the solvation process. There are also several other studies reporting the successful application of ML in predicting ΔG_sol of organic solvents, but due to the length limit of this article, we only demonstrated them in Table 1 and omitted their detailed discussion in our text. Overall, the prediction of ΔG_sol has been successfully achieved by various ML methods with high accuracy. However, compared to the aqueous phase, the available data of ΔG_sol for organic solvent are relatively much smaller and there is still a lack of data for many organic solute molecules such as 2,3-butanediol in organic solvents. The extrapolation capability of the currently developed ML models requires further examinations.

Several research groups have extended the application of ML models to predict the physicochemical properties of green solvents, such as ILs.⁹⁹⁻¹⁰⁴ For instance, Song et al. constructed two hybridized GC-based models (ANN and SVM) to predict the CO₂ solubility of ILs at various temperature and pressure conditions.¹⁰⁵ Their findings indicated the superior performance of their ANN-GC model compared to its SVM counterpart in CO₂ solubility prediction. Similarly, Cao et al. developed three quantitative structure–property predictive models (Multilinear Regression (MLR), SVM, and ELM) using a data set comprising 119 ILs to predict the toxicity values of ILs.¹⁰⁶ Based on their results, their ELM model exhibited the highest predictive accuracy compared to their MLR and SVM models, revealing a strong relationship between increasing alkyl chain length (dictated by the chosen cation) and increasing toxicity. In a separate study, Can et al. employed a combination of three ML models (ARM, DT, multilayer ANN) with molecular descriptors derived from COSMO-RS and semiempirical estimation methods to predict the water capacities of ILs.¹⁰⁷ Their findings demonstrated high predictive performances of all three models in determining the descriptor effects, heuristic cation–anion pairing rules, and predicted water capacity of ILs. Additional examples of IL are highlighted in Table 1. The readers can also refer to previous articles for a wider range of ML applications in the field of ILs.^86,108 Considering the vast number of possible ILs that can be formed, the size of the current IL database is still relatively small. How to efficiently generate unbiased IL data with comprehensive property characterization in batches remains an interesting question. In this context, data bias may significantly contribute to the wide variation in the accuracy of machine learning models reported in the literature.

ML models have also been applied to predict the physicochemical properties of newer classes of solvents, such as DESs.^109−112 For instance, Lemaoui et al. conducted a study in which they developed and compared two novel quantitative structure–property models based on MLR and ANN algorithms to predict the pH of DESs.¹¹³ They found that their ANN model exhibited greater predictive performance, while their MLR model provided simpler interpretability. Another notable example is the study by Abdollahzadeh et al., which involved the construction and comparison of seven different ML algorithms to predict the density of DESs.¹¹⁷ Their least-squares SVR (LSSVR) exhibited the highest predictive accuracy and evaluated the effects of temperature variation, HBA selection, and HBD selection on the predicted densities of DESs. Other examples of ML applications to the studies of viscosity and gas capture capability in DES are provided in Table 1. Earlier computational data-driven works of DESs are also reviewed in several previous articles.^114−117 Similar to ILs, the limited amount of existing DES data relative to the large number of possible DES compositions remains a bottleneck in generating new, unbiased results. Therefore, a deeper fundamental understanding of the DES structures is still needed.

The application of ML models for green solvent screening introduces distinct challenges when compared to conventional solvent screening, primarily due to the relatively low availability of reported data and a limited understanding of their specific structure–property relationships. Thus, researchers often adopt a combination of feature selection techniques to refine the accuracy of their ML models, especially for more complex solvent systems such as ILs/DESs.¹¹⁸ As an example, the study conducted by Cao et al. employed stepwise regression as a feature selection technique to optimize the predictive accuracy of their IL toxicity models. By utilizing a stepping criterion and combination of forward selection and backward elimination procedures, Cao et al. successfully refined their MLR model with an eight-coefficient descriptor equation.¹⁰⁶ In the study by Aghaie and Zendehboudi, Genetic Algorithm Multilinear Regression (GA-MLR) was employed as a feature selection technique to identify the most important descriptors for the construction of linear ML models for IL CO₂ solubility prediction. Their GA-MLR results highlighted specific geometrical descriptors, including Chi_G/D 3D, HOMO–LUMO fraction for anions, Disps, and SpMax_RG for cations, as the most significant features contributing to the accuracy of their ML model predictions.⁶⁹ By selecting ML model features to account for the nuanced and complex molecular interactions within complex solvent systems, including ILs and DESs, researchers can enhance the accuracy of their screening predictions, driving the discovery and development of novel solvents with customizable properties.

However, achieving a balance between the model complexity and interpretability is crucial. ML models, often perceived as “black boxes”, are commonly subjected to a variety of intricate decision-making and extensive data analysis tasks. This inherent complexity, arising from a multitude of features, parameters, and algorithm layers, can pose challenges in elucidating the specific factors influencing a model’s predictions. In solvent screening, a comprehensive understanding of a solvent’s structure–property relationship and behavior is essential for guiding future research. Therefore, employing methods, such as feature importance analyses, are key to interpreting ML model predictions and gaining mechanistic insights.

Notably, the SHAP (SHapley Additive exPlanations) method has emerged as a popular feature importance technique in ML. In contrast to conventional feature importance methods commonly found in RF, DT, and gradient boosting algorithms, SHAP calculates and averages numerical values, known as Shapley values, for each model feature. This approach considers all possible feature combinations, resulting in a fair and uniformly distributed measurement of a specific feature’s impact on the output of an ML model, allowing an enhanced degree of versatility and applicability across a wide range of ML model types. Within the solvent screening space, Lei et al. employed SHAP to elucidate the importance of their ML model features in predicting the surface tension and viscosity of IL-organic solvent binary systems.¹⁰³ Their findings revealed that the IL mole fraction, followed by the presence of −OH functional groups in organic solvent, had the greatest impact in predicting the surface tensions of IL-organic systems, while the IL mole fraction, followed by the presence of −CH₂ groups in the IL, exhibited the greatest impact on viscosity prediction. Similarly, a study by Liu et al. utilized SHAP to quantify the impact of each structural descriptor and parameter in the density and heat capacity of IL-organic solvent binary systems.¹⁰⁴ The SHAP results indicated that the IL mole fraction exhibited the greatest influence in predicting the densities of IL-organic systems, followed by the presence of an IL anion, Tf₂N, and the presence of −CH₃ groups in the organic solvent. For heat capacity prediction, SHAP revealed the IL mole fraction as the greatest contributing feature, followed by the presence of −CH₃ groups on ring structures. Moreover, the SHAP methodology has been utilized for the synthesis of DESs.¹¹⁰ The investigation by Shi et al. leveraged the SHAP framework to methodologically quantify and rank the influence of HBA and HBD structures on the viscosity properties of DESs (Figure 2). Intriguingly, Shi et al. articulated a comprehensive analysis whereby the SHAP values were correlated with various parameters, including temperature, the molar mass of HBA and HBD, and the molar ratio of HBA and HBD. Through this approach, the study has significantly contributed to the advancement of a more rationalized design and synthesis process for DESs, enhancing the precision in tailoring their physicochemical properties.

Figure 2.

ML workflow with model interpretation via Shapley methodology.

2.3. Extraction Process Design and Optimization via ML Models

The application of ML models can surpass the general prediction of solvents’ physicochemical properties to a more focused investigation of the behavior and interactions of solvents in extraction processes.^119−121 The distribution or partitioning coefficient (K_d), often expressed in its logarithmic form (LogD), is the ratio of mass fractions in a mixture of two immiscible liquid phases at equilibrium. K_d is expressed as

3

where m_extract is the mass fraction of a solute in the extract phase (solvent-rich) and m_raffinate is the mass fraction of a solute in the raffinate phase (aqueous). Related to the distribution behavior is the separation factor (β), also known as the selectivity. β is expressed as

4

where m^extract_A/m^extract_B is the ratio of mass fractions of solutes A and B in the extract phase and m^raffinate_A/m^raffinate_B is the ratio of mass fractions of solutes A and B in the raffinate phase.

In ML-based solvent screening, both the distribution coefficient and separation factor can be predicted to evaluate the efficiency of a solvent extraction process, as shown in Table 2. For instance, Liu et al. constructed a DNN to predict the distribution coefficients of ligand-assisted solvent extraction of rare-earth elements.¹²² Utilizing a combination of molecular physicochemical descriptors, generated from RDKit, and atomic extended-connectivity fingerprints, their DNN model predicted the distribution coefficients of four novel ligand structures for the liquid–liquid extraction of trivalent lanthanides from aqueous solution, which was successfully validated through experimentation. Another notable example is the study by Fajar et al. in which two RF algorithms were constructed using a classification data set of 76 ILs and a regression data set of 110 ILs to predict the selectivity and eco-toxicity values of ILs for metal extraction, respectively.¹²³ By leveraging the feature importance RF algorithm, Fajar et al. successfully identified the important factors determining IL metal selectivity and eco-toxicity: cation/anion pairing selection and cation selection (particularly the hydrocarbon region), respectively. Similarly, a study by Darwish et al. utilized a combination of COSMO-RS and ML modeling to prescreen DES constituents and investigate the effect of σ-profile on the predicted extraction behavior of DESs for furfural extraction.¹²⁴ Specifically, Darwish et al. successfully prescreened 108 DES constituents based on predicted distribution coefficient and selectivity, which resulted in the identification of five optimal constituents. Their DT and MLR algorithms indicated positive correlations between HBD and HBA σ-values, revealing aromatic-based DESs containing phenolic hydroxyl groups as the ideal solvents for the liquid–liquid extraction of furfural.

Table 2. Summary of Studies Using ML Models for Extraction Process Design and Optimization.

Extraction process	Key parameters	No. of data points (training/test)	ML model	R2	Ref.
IL/DES extraction of oil sludge	Conductivity, surface tension, pH, viscosity	Not specified	RR	0.96 (RMSE = 0.95)	(119)
			MLP	0.95 (RMSE = 1.01
			SVR	0.95 (RMSE = 1.03)
DES extraction of boron contaminants from aqueous media	Molar ratio, density, viscosity, water solubility, leachability	500/126	ANN	0.98	(120)
Ultrasonic DES extraction of date fruit sugars	Temperature, extraction time, solvent to date fruit ratio	Not specified	RSM	0.84	(121)
			ANN	0.97
Ligand-assisted IL extraction of lanthanides	Molecular physicochemical descriptors, atomic extended-connectivity fingerprints	1085/117	Multilayer NN (DL)	0.85	(122)
IL extraction of metals	Cation/anion structures dictate metal selectivity, cation structure dictates eco-toxicity	186 (total)	RF	0.76	(123)
Hydrophobic DES extraction of furfural	Aromatic-based solvents containing phenolic hydroxyl groups	Not specified	MLR	0.92	(126)
Solvent extraction of water pollutants using emulsion liquid membranes	Extraction time, volume ratio of organic phase to internal phase, emulsification time, feed concentration	850/150	LR	0.60	(125)
			RF	0.92
			XGB	0.94
			ANN	0.82
Cosolvent-assisted supercritical fluid extraction of date fruit sugars	Cosolvent selection, CO2 flow rate, cosolvent ratio, temperature, pressure, time	Not specified (ratio 85/15)	MNLR	0.87	(126)
			ANN	0.99
Process parameter evaluation and optimization for solvent extraction using pulsed disk and doughnut columns	Sauter mean droplet diameter (d32), axial mixing coefficient continuous phase (Ec), height of a mass transfer unit continuous phase (Hoc), volume fraction holdup of dispersed phase (xd)	d32 = 337 (total); Ec = 557 (total); Hoc = 260 (total); xd = 203 (total)	RF	0.99 (d32), 0.92 (xd), 0.97 (Ec), 0.98 (Hoc)	(127)
			SVM	0.98 (d32), 0.96 (xd), 0.95 (Ec), 0.97 (Hoc)
			ANN	0.98 (d32), 0.93 (xd), 0.97 (Ec), 0.98 (Hoc)
Hydrogen recovery using IL binary mixtures	Temperature, molar fraction, chemical functional groups, density, viscosity, surface tension, heat capacity		ANN	0.99	(128)

Furthermore, some research groups have expanded the application of ML models to predict the effects of external process parameters, such as operating temperature, pressure, and time, on extraction processes. For instance, Li et al. employed LR, RF, XGB, and ANN algorithms using a database comprising preparation condition and operating parameter data to simulate the extraction efficiency of emulsion liquid membrane-based extraction for contaminant removal.¹²⁵ They found that their XGB model showed the highest accuracy for predicting the extraction efficiency of their emulsion liquid membrane process. Furthermore, their models identified extraction time, volume ratio of the organic to internal phase, emulsification time, and feed concentration as significant parameters in tailoring the efficiency of their process. Similarly, a study by AlYammahi et al. employed MNLR and ANN models to explore the relationship between various process conditions and sugar extraction efficiency through supercritical CO₂ and cosolvents.¹²⁶ They found that their ANN model performed better than the MNLR in predicting the maximum total sugar content of the extraction process based on the solvent volume ratio, temperature, pressure, and cosolvent volume ratio. Furthermore, Su et al. developed three ML models (RF, SVM, and ANN) to predict the performance of pulsed disc and doughnut columns for liquid–liquid extraction.¹²⁷ By utilizing a comprehensive database comprising 1357 performance criteria, all three ML models predicted performance parameters, including the dispersed-phase holdup, drop size, axial diffusion coefficient, and height of the mass transfer unit. From their results, they observed that the RF model exhibited the highest accuracy in predicting the drop size, axial diffusion coefficient, and height of the mass transfer unit, while the SVM exhibited the highest accuracy in predicting the dispersed-phase holdup. Additionally, the RF feature importance algorithm was applied, which identified pulse intensity, dispersed phase velocity, continuous phase velocity, and continuous phase properties as the most important input features in determining the drop size, dispersed-phase holdup, axial diffusion coefficient, and mass transfer unit height, respectively. In summary, these examples have shown the extensive applicability of ML-based approaches in the optimization of process parameters in various extraction processes. It is also noted that, in many cases, the available training data set may only allow for simpler ML algorithms rather than neural networks. Nevertheless, these simpler ML models can still achieve satisfactory performance. Therefore, future investigations may benefit from starting with simpler ML algorithms before progressing to more complex deep learning methods, which require significantly larger data sets from experimental measurements under various operational conditions.

Last but not least, we highlight a recent successful workflow for IL–IL binary mixtures: from ML-assisted modeling, IL solvent tailoring, to process design and optimization, then to the application in hydrogen recovery from raw coke oven gas.¹²⁸ Since most published studies have focused on the property prediction of pure single-component IL, the property prediction of IL-containing mixtures is still rare. Chen et al. demonstrated an integrated multistep framework starting from raw data curation to a final optimized process for hydrogen recovery using IL mixtures. It involves the initial database development from public resources, descriptor generation and ML model development, mathematical modeling of IL mixtures with case-dependent constraints, and process design modeling and optimization with case-dependent constraints. The authors not only identified the components of a desired IL binary mixture but also suggested a series of key process parameters such as flow rate, adsorption column size, pressures, etc., for hydrogen recovery. Although the current framework lacks flexibility and transferability, it provides a valuable starting point for the solvent screening community to integrate the fundamental development of solvent compounds with the applied design of the solvent-based process in the early stages of research.

3. HIGH-THROUGHPUT SCREENING

Considering the potential inconsistency and low repeatability of reported solvent data, as well as the issue of low-quality training data sets for ML models for ILs/DESs, HT methods can effectively provide high quality data and reduce experimental costs.^129−131 The maturation of AI/ML, in recent years, has presented a unique opportunity to seamlessly integrate intelligent decision-making with HT workflows, thereby creating a synergistic platform for rapid and efficient data collection, data analysis, and process optimization in a wide range of materials, chemistry and biochemical research.^132−134 This integrated approach holds tremendous promise for driving innovation in separations, especially in solvent screening for extraction processes. Examples of HT studies for solvent-based chemical and separation applications are summarized in Table 3.

Table 3. Summary of Studies Using High-Throughput Screening for Chemical and Separation Applications.

Property	Application	HT method	Analytical method	Ref.
Solubility	Automated generation of empirical solubility database of various solute/organic solvent combinations	Unchained Laboratories robotic system	HPLC	(135)
	Automated determination of caffeine solubility in different organic solvents	Customized robotic system with web-camera/algorithm	N/A	(136)
	Automated salt screening of synthetic intermediates and ranking of counterion/solvent combinations for drug development	Automated birefringence imaging station	HPLC	(137)
	Automated determination of aqueous and nonaqueous solubility of redox active materials for battery applications	Big Kahuna robotic system; Unchained Laboratories robotic system	NMR	(138)
			UV–vis
Phase boundary	Automated interface detection of organic solvent and aqueous mixtures for drug discovery	Abbott liquid–liquid extraction station with refractive index detection	HPLC	(139)
	Automated partition coefficient measurement of organic solvent and aqueous mixtures	Unchained Laboratories Freeslate CM3 system; camera and MATLAB image analysis algorithm	HPLC	(140)
			NMR
			KF
	Automated phase separation dynamics of organic solvent and aqueous mixtures	Basler acA1300-39 μm scan monochrome camera for image analysis	N/A	(141)
	Automated phase detection or organic solvent and aqueous mixtures for downstream purification processes	Tecan Freedom EVO200 robotic system with TubeEyeX camera and software	LC	(142)

3.1. Solubility Determination

Thermodynamic solubility data are key to the design and optimization of separation processes. In solvent extraction, the solubility of a target compound plays a significant role in determining the rate and yield of its transfer across liquid phases. Traditionally, researchers have relied on hands-on experimentation using titration, gravimetric analysis, conductivity measurements, and shaking/stir-flask techniques to determine the empirical solubility of various solute–solvent systems. However, such methods are often time-consuming and resource intensive. In recent years, there has been great interest in leveraging HT workflows to accelerate the solubility screening process, specifically for chemical and separation applications. Notably, Qiu and Albrecht established a comprehensive data set comprising 1125 solubility screening panels of 905 solutes that were empirically collected using an Unchained Laboratories automated, HT workflow.¹³⁵ Their automated system contained a 96-well filter assembly with liquid transfer, mixing, and temperature control, which allowed for the preparation of various solute, temperature, and organic solvent combinations for high-performance liquid chromatography (HPLC) analysis. In a separate study, Shiri et al. developed a modular robotic platform integrated with computer vision as a proof-of-concept system for fully autonomous solubility screening.¹³⁶ Specifically, this approach utilized a robotic solid and liquid dosing system for automated solubility sample preparation as well as a webcam and corresponding algorithm to measure the turbidity of various solutions. Similarly, a study by Qiu et al. utilized a HT methodology involving an automated birefringence imaging station for salt screening on synthetic intermediates.¹³⁷ By combining their HT workflow with Z-score statistical analysis, Qiu et al. were able to systematically rank different acids, solvents, and acid/solvent combinations based on their respective solubility. The accuracy and precision of these HT solubility experiments, however, may vary based on various experimental parameters, determined by the specific researcher. To address this limitation, Liang et al. proposed a standardized HT screening workflow for the large-scale determination of redox-active molecule solubilities.¹³⁸ Specifically, Liang et al. evaluated the effects of the mixing time, solid–liquid phase separation process selection, analytical method selection, and volume transfer calibration on their solubility results. Using indol-3-carboxylic acid as a model compound, their optimized system successfully predicted its solubility under various conditions with high accuracy and precision, as validated through experimentation.

3.2. Phase Detection

The phase behavior of a solvent system is vital in defining the overall distribution and partitioning interactions within an extraction process. Traditional evaluation of a multiphase solvent system’s phase behavior and distribution coefficient has typically involved the manual preparation of liquid–liquid extraction samples, phase volume estimation, and manual extraction of aliquots from each phase for chromatographic analysis. To accelerate this process, various automated HT technologies focused on phase detection have been developed. One early example is the Abbott Liquid–Liquid Extraction Station, which featured a cylindrical robot capable of managing 80 sample vials paired with a refractive index phase detector.¹³⁹ In its initial 18 months of operation, the Abbott Station successfully processed over 6000 compounds, significantly improving workflow efficiency. More recently, researchers and engineers have integrated a variety of algorithms to advance these technologies. Duffield et al., for instance, customized their robotic system with a camera and MATLAB image analysis algorithm to visually quantify the phase volumes of organic–aqueous biphasic samples in equilibrium.¹⁴⁰ Notably, their automated image analysis HT workflow achieved highly accurate and robust partition coefficient measurements while using less than 1% of the typical reagent amounts and providing up to 94% of the time savings from previous manual experimental studies. A study by Daglish et al. expanded upon this workflow by incorporating an automated image analysis algorithm to measure the separation rate of biphasic samples over time.¹⁴¹ In addition, a study by Sun et al. introduced a custom Python algorithm combined with their automated Tecan image analysis system for more rapid liquid chromatography data visualization and partition coefficient calculation.¹⁴² In both solubility measurement and phase boundary identification, liquid chromatography and NMR have been employed in most cases, although potential issues, such as measurement efficiency and compatibility, still remain. We believe that further enhancement of these two methods and implementation of new characterization methods will attract more researchers in the near future.

3.3. Self-Driving Laboratories

While numerous automated HT solvent screening platforms have been successfully developed and commercialized, many of these systems still rely on manual data analysis and interpretation by researchers. This reliance has prompted growing interest in the advancement of AI/ML-powered laboratories that can autonomously design, validate, and optimize processes. Readers can refer to several comprehensive review articles for AI-assisted self-driving laboratories in the field of chemistry, separation, and materials science.^39,143−145 For instance, Clayton et al. introduced a self-driven platform for optimizing multistep continuous reaction processes in pharmaceutical development.¹⁴⁶ By integrating multiobjective ML models like TSEMO with automated continuous flow reactor systems, Clayton et al. achieved efficient process optimization of a Sonogashira reaction within only 13 h. Moreover, their system markedly reduced the quantity of catalytic reagents typically required in manual process design and experimentation, saving both time and resources. Another noteworthy example is the study conducted by Pomberger et al., which integrated ML models with a liquid handling robotic platform for the autonomous preparation and adjustment of buffered polyprotic solutions with varied pH levels for formulation chemistry applications.¹⁴⁷ They found that their GP model, featured with chemical input data, achieved the highest accuracy in predicting pH, thereby enabling precise pH adjustment of solutions through automated robotic acid/base transfers. In the context of solvent screening, this study showcases an exciting opportunity to optimize separation processes in real-time, facilitating an accurate and robust closed-loop workflow that minimizes manual intervention and the use of experimental resource use.

In addition to a physical self-driving lab, the digital lab demonstrates another important type of research approach that is usually supported by cloud facilities.¹⁴⁸ One of the mostly studied types of digital laboratories is digital twin.^149−151 A digital twin is a virtual counterpart of a physical entity, designed for simulation, analysis, and evaluation of a real-world object.¹⁵² It can bring numerous advantages, such as cost reduction, performance optimization, enhanced decision-making, and accelerated innovation. For instance, Wu et al. developed a digital twin model to characterize CO₂ capture by single atom solutions.¹⁵⁰ The model successfully revealed the thermoelectrical effect and CO₂ capture performance, and it suggested conditions that increased CO₂ desorption efficiency by 4.2% and reduced energy consumption by 35.4%. Several two-dimensional single atom solutions such as Cu/C, Cu/N, or Cu/Si were proposed as potential advanced materials for CO₂ capture. Pallavicini et al. applied digital twin modeling to validate a demo-scale plant of biogas scrubbing technology.¹⁵³ Through evaluating the H₂S absorption process in a biogas production plant, the model simulated operating conditions different from those of the demo-scale plant to identify the controlling variables for H₂S removal efficiency. The authors found soda flow rate and concentration as the top two influencing parameters, compared with H₂S concentration, temperature, and freshwater flow rate. Örs et al. also show that the AI-assisted operational digital twin can be promising and beneficial in the chemical process engineering.¹⁵¹ Moreover, Lo et al. proposed the concept of a “frugal twin” and reviewed its application to low-cost self-driving laboratories for chemistry and materials science.¹⁵⁴ We believe that the digital twin approach, although not widely implemented in solvent screening and design yet, will attract more attention from scientists and engineers globally in the near future.

4. LIMITATIONS AND PROSPECTS

AI/ML has proven to be a powerful tool for screening, designing, and optimizing novel solvents, chemical reactions, and separation processes. However, it is crucial to acknowledge that the ease and accuracy of ML-based solvent screening strategies can vary based on several factors. The quantity and quality of an input data set, for instance, can significantly impact the accuracy and applicability of an ML model for different solvents, solutes, and processes. A consistent trend among organic solvent ML studies is the abundance of existing data readily accessible through open-source databases. Still, it is worth noting that the amount of data for organic solvents is far below those for aqueous solvation. Therefore, reliable data for organic solvent mixtures remain in short supply. Nonetheless, the well-established knowledge of organic solvents and their properties from the previous literature can provide valuable guidance for ML model development, deployment, and interpretation. On the contrary, newer types of solvents such as ILs and DESs may face challenges in ML applications due to the limited data and relatively modest understanding of their molecular interactions and properties. Moreover, the diverse array of cation/anion or HBA/HBD combinations necessitates the use of advanced ML techniques with additional experimental validation, potentially requiring extra time and effort from researchers.

Addressing these challenges, the synergy of AI/ML with HT technologies appears to be a promising solution. This combination excels at automating experimental tasks and swiftly generating comprehensive empirical data sets, offering a synergistic platform for quickly and efficiently identifying optimal solvent structures, properties, and conditions for chemical and separation processes. Recent research has verified the feasibility of automation laboratories built on this concept. Furthermore, the rapid advancement and adoption of large language models, such as ChatGPT, herald a new era in which these LLM tools serve as invaluable chemistry assistants for solvent screening. These innovations enable researchers, including those without programming skills, to perform complex text mining, standardize data formats, and execute detailed calculations based on published experimental data. However, it is important to note that general LLM models such as ChatGPT often fail in domain sciences such as chemistry or materials science, where its answers require careful examination by a domain expert.^155,156 A further refinement by domain knowledge is necessary to improve the reliability. By integrating real experimental outcomes with established empirical or semiempirical models, the accuracy of predictions can be significantly improved. Such pioneering efforts have been documented in fields like catalysis research,¹⁵⁷ material design,¹⁵⁸ drug discovery,¹⁵⁹ and water harvesting.¹⁶⁰ Although reports on solvent selection and extraction separation are currently sparse, it is anticipated that research in this domain will soon gain significant attraction.

Finally, we would like to address a powerful tool for inverse design, e.g., starting from the desired property to suggesting a possible chemical structure or material component. We have witnessed successes of Gen-AI achieved in numerous fields such as chemical science,¹⁶¹ materials engineering,¹⁶² and food design.¹⁶³ We believe the time has already come to move beyond showcasing the viability of Gen-AI and toward utilizing Gen-AI in practical scientific research. However, applying Gen-AI to separation science such as green solvent screening remains a less explored territory. Integrating self-driving experiments with Gen-AI for solvent selection or process optimization will undoubtedly accelerate extraction research in the near future.

5. Conclusion

The pivotal role of separations across industries, compounded with the drive toward sustainable practices, necessitates the evolution of more efficient and environmentally friendly solvent extraction processes. The integration of AI/ML and automated high-throughput (HT) technologies presents an exciting opportunity to expedite the discovery and development of novel solvents and process designs with enhanced specificity. These integrated methodologies exhibit substantial potential for overcoming bottlenecks associated with traditional solvent selection and experimentation. By enabling a more efficient, data-driven paradigm, the integration of AI/ML and HT technologies offers a promising path toward sustainable and economically viable extraction approaches, reshaping the landscape of industrial separation processes with unprecedented precision and effectiveness.

Glossary

Abbreviations

AI

Artificial Intelligence

ANN

Artificial Neural Network

ARL

Association Rule Learning

ARM

Association Rule Mining

CML

Chemical Markup Language

CNN

Convolutional Neural Net

COSMO-RS

Conductor-like Screening Model for Realistic Solvation

COSMO-SAC

Conductor-like Screening Model for Segment Activity Coefficient

DNN

Deep Neural Network

DES

Deep Eutectic Solvent

DT

Decision Tree

ELM

Extreme Learning Machine

ET

ExtraTrees

GB

Gradient Boosting

GA-MLR

Genetic Algorithm Multilinear Regression

GC

Group Contribution

ΔG_sol

Gibbs Free Energy of Solvation

GPU

Graphics Processing Unit

HBA

Hydrogen Bond Acceptor

HBD

Hydrogen Bond Donor

HT

High-Throughput

HPLC

High-Performance Liquid Chromatography

IL

Ionic Liquid

InChI

International Chemical Identifier

K_d

Distribution Coefficient

KF

Karl Fischer Titration

LC

Liquid Chromatography

LLE

Liquid–Liquid Extraction

LLM

Large-Scale Language Model

LogReg

Logistic Regression

LR

Linear Regression

MAE

Mean Absolute Error

ML

Machine Learning

MLR

Multilinear Regression

MNLR

Multiple Nonlinear Regression

NMR

Nuclear Magnetic Resonance

PCA

Principal Component Analysis

R²

Coefficient of Determination

RF

Random Forest

RMSE

Root Mean Square Error

RNN

Recurrent Neural Network

SVM

Support Vector Machine

SVR

Support Vector Regression

SHAP

SHapley Additive exPlanations

SMARTS

SMILES Arbitrary Target Specification

SMILES

Simplified Molecular Input Line Entry System

TSEMO

Thompson Sampling Efficient Multi-Objective Optimization

UV–vis

Ultraviolet–Visible Spectroscopy

UNIQUAC

Universal Quasi-Chemical Activity Coefficient

UNIFAC

Universal Quasi-Chemical Functional Group Activity Coefficient

XGB

eXtreme Gradient Boosting

Author Contributions

DZ and DL conceptualized the topic and determined the scope. JPE, DZ, and DL conducted the literature search and drafted the manuscript. NS and V-AG provided critical insights into the manuscript content and provided substantial revisions. All authors reviewed and approved the final manuscript.

The authors declare no competing financial interest.