Machine learning-based structure‒property modeling for ionic liquids design and screening: A state-of-the-art review

Abstract

With the growing emphasis on sustainable development, the demand for environmentally friendly solvents in green chemical processes and carbon dioxide capture is increasing. Ionic liquids (ILs), as promising green solvents, offer significant potential but face considerable challenges, particularly in solvent selection. To overcome the limitations of traditional screening methods, machine learning (ML) techniques have recently been applied, offering a more efficient and data-driven approach. This review provides an overview of key ML methods used in solvent screening and compares them with traditional experimental and theoretical techniques. It examines the role of descriptor selection in structure‒property-based methods, such as quantitative structure-activity relationships (QSAR) and quantitative structure‒property relationships (QSPR), which are critical for predicting IL properties. The review also explores the application of these methods to screen IL properties, including toxicity, viscosity, density, and CO₂ solubility. Additionally, it discusses challenges in selecting appropriate models based on data scale and task complexity, integrating physical information for model interpretability, and achieving multi-objective optimization to balance key properties in ionic liquid (IL) design. Finally, it summarizes the achievements, limitations, and prospects of ML applications in ILs research, offering insights into how these methods can advance the development of sustainable ILs.

Graphical abstract

Keywords

machine learning (ML) / ionic liquid (IL) / structure‒property / molecular descriptors / physical property

1 Introduction

As environmental issues become increasingly severe, achieving sustainable development in the chemical industry has emerged as a critical global challenge [1]. Although the widespread application of petroleum-based organic solvents has driven technological progress, their high volatility and toxicity pose significant threats to both the environment and human health [2]. Developing alternative solvents that are both efficient and environmentally friendly has become a pressing concern for scientists and engineers. Among various alternatives, ILs have stood out as a new class of solvents due to their low volatility, broad liquid temperature range, and excellent thermal stability [3]. These properties make ILs an ideal choice for green chemistry and sustainable technologies.

Unlike traditional solvents, ILs consist of organic cations and inorganic or organic anions, giving them a unique molecular structure and the characteristic of “designability” [4]. By selecting different ion combinations or introducing functional groups, researchers can tailor the physicochemical properties of ILs to meet specific requirements. For instance, in carbon capture, ILs based on imidazolium cations have been designed to enhance CO₂ absorption capabilities [5], while amino-functionalized ILs exhibit higher selectivity and solubility [6]. In catalytic reactions, ILs systems based on phosphate anions have significantly improved the efficiency of acid-catalyzed reactions [7]. Furthermore, due to their high conductivity and low volatility, ILs are employed in the design of battery electrolytes, greatly enhancing the stability and safety of electrochemical devices [8]. However, the immense design freedom of ILs presents a vast chemical space: theoretically, the combination of cations and anions can yield up to 1018 potential ILs [9], posing significant challenges designing and screening.

Reviewing the development of ILs, molecular design approaches have evolved from experimental trial-and-error to theory-assisted methods. In previous studies, ILs design primarily relied on extensive laboratory experimentation, exploring combinations of cations and anions to investigate their properties [10,11]. However, this empirical approach was costly, inefficient, and insufficient for addressing the complex and diverse demands of practical applications. With the advent of theoretical computational tools, researchers have employed methods such as molecular dynamics (MD) simulations [12], quantum chemical calculations [13], and thermodynamic predictions [14] to gain deeper insights into the relationship between the structure and properties of ILs at the molecular level. While these computational methods have significantly improved design precision, they remain constrained by computational complexity and high costs, particularly for large-scale screening.

To effectively study the physicochemical properties of ILs, researchers have developed the structure‒property relationship (SPR) approach, which summarizes experimental findings and empirical knowledge [15]. This method analyzes the qualitative relationship between molecular structures (e.g., functional groups and molecular configurations) and physicochemical properties, providing an intuitive approach for preliminary IL design and screening. However, SPR is limited to identifying trends and guiding experimental design. It lacks the capability for quantitatively analysis. To address this, Hansch and Fujita [16] proposed the quantitative structure‒property relationship (QSPR) model, which transforms molecular structural features into molecular descriptors (such as molecular weight, topological indices, and polarity) and employs regression algorithms to quantify relationships between these descriptors and properties. This development not only expanded the scope of IL design but also laid the foundation for data-driven methods.

In recent years, emerging technologies such as ML have further advanced QSPR methodologies [17]. ML enables the construction of complex nonlinear models that can accurately predict IL properties, handle high-dimensional data, and uncover hidden relationships. By integrating experimental data with theoretical computations, ML has significantly enhanced IL design efficiency and optimization, offering a novel pathway for developing multifunctional and high-performance ILs. However, current research still primarily focuses on optimizing single functionalities, while systematic summaries of design principles for multifunctional ILs remain scarce. Furthermore, the organic integration of experimental data, theoretical computations, and ML to achieve cross-scale, multi-objective collaborative optimization presents an ongoing challenge.

Therefore, this review begins with an overview of various ML methods applied in QSPR modeling (see Fig.1), followed by a discussion on molecular design strategies, highlighting the evolution and advancements of IL design methods and their applications. It particularly emphasizes the application of ML-based QSPR methods in solvent design and screening. Additionally, it compares the strengths and limitations of different ML methods across varying data scales and task complexities. It also proposes an interpretative approach that integrates MD and density functional theory (DFT) with ML models to provide a more comprehensive and in-depth physicochemical explanation. Finally, it discusses multi-property prediction for practical engineering applications, offering forward-looking insights for the precise design of novel ILs. This review aims to establish a systematic framework and provide insights for advancing IL-based solvent design and screening across various applications.

2 Machine learning methods

At present, ML methodologies have been extensively employed across numerous conventional disciplines, providing novel insights that have facilitated innovation and advancement within these traditional domains. In particular, the integration of ML techniques—ranging from the prediction of physicochemical properties to process optimization—has demonstrated distinct advantages in the field of solvent design and screening. This section presents a comprehensive overview of the application of various ML techniques in solvent research. A search of the Web of Science database using the keywords “machine learning” and “ionic liquid” over the past years returned hundreds of relevant papers (as illustrated by recent publication trends in Fig.2). The most frequently employed ML methods are neural network algorithms, including artificial neural networks (ANNs), deep neural networks (DNNs), and convolutional neural networks (CNNs), collectively cited in 55 articles. These are followed by decision tree-based algorithms, such as decision trees (DTs) and random forests (RFs), which appear in 32 articles. Support vector machine (SVM) algorithms are referenced in 24 articles, while Gaussian process regression (GPR) is mentioned in 8 articles.

2.1 Neural network methods

Neural network algorithms (as show in Fig.3), the most popular ML algorithms, are extensively employed in prediction of complex properties of various solvents due to their exemplary nonlinear fitting capabilities. ANNs typically consist of one input layer, one output layer, and a few hidden layers [18], making them well-suited for problems with limited number of parameters. However, their predictive accuracy may decline when addressing more complex problems. To overcome this limitation, learning algorithms have emerged. DNN algorithms, which feature multiple hidden layers, can efficiently extract features from complex raw data and are particularly suited to addressing high-dimensional and nonlinear issues.

For example, Ardeshiri and Rashidi [20] modeled the CO₂ removal efficiency of dilute ethanolamine solvents using an ANN approach. They optimized the number of neurons and tested different transfer functions in multilayer perceptron (MLP) and radial basis function (RBF) networks to identify the most effective ANN model. By using experimental parameters for training, they effectively captured the relationship between input variables (e.g., solvent concentration, temperature, and flow rate) and output performance.

Zhu et al. [21] employed a five-enhanced ANN approaches to estimate and compare the solubility of sulfur dioxide (SO₂) in various deep eutectic solvents (DESs). The models employed molecular weight, solvent water content, temperature, pressure, and SO₂ absorption data as inputs. The comparison revealed that the single hidden layer multilayer perceptron method optimized with the Levenberg-Marquardt (MLP-LM) algorithm demonstrated the highest accuracy in predicting the DES-SO₂ phase equilibrium data, achieving an R² of 0.97936.

To improve the prediction accuracy, Zhang et al. [22] proposed an accurate and interpretable DNN (AI-DNN) model for predicting the lipid solubility properties of organic compounds. Compared to previous DNNs based on QSPR methodologies, the AI-DNN approach utilizes a hybrid molecular representation that integrates descriptors from the chemistry development kit (CDK) and features learned through directed message passing neural networks (D-MPNNs). This hybrid approach captures both local and global molecules features, thereby enhancing model performance in predicting physicochemical properties.

Although ANN and DNN algorithms have been widely used in solvent selection, they still present limitations. These algorithms rely on a set of hyperparameters that must be carefully tuned [23]. Additionally, as network depth increases, model interpretability tends to decrease, and the p risk of overfitting becomes more pronounced. In this context, CNNs, as deep learning models specialized in spatial data, can effectively extract complex features through local connectivity and parameter sharing [24], often without the need for human supervision [25]. CNNs demonstrate enhanced robustness when handling high-dimensional data, effectively mitigating overfitting while maintaining interpretability.

For instance, Bahmaninia et al. [26] used a CNN model to predict CO₂ solubility in a variety of physical solvents, including methanol, ethanol, propylene glycol, n-pentanol, n-butanol, n-propanol. They also analyzed the effects of non-mass factors such as critical temperature and critical pressure on the solubility of CO₂ in physical solvents. The CNN model achieved an R² of 0.9943, outperforming other models in predictive accuracy.

Chen et al. [27] trained a transformer model on the ChEMBL database to extract molecular fingerprints, which were then used as inputs for a CNN model to predict the surface charge density profile (σ profile) and cavity volume (VCOSMO). The results demonstrate that the model could rapidly predict these properties of millions of molecules within just a few minutes. Leveraging this high computational efficiency, the researchers performed high-throughput solvent screening, offering promising implications for solvent selection and process design in chemical engineering.

2.2 Decision tree and random forest methods

In the aforementioned studies, neural network algorithms have effectively captured the nonlinear characteristics inherent in solvent selection processes. However, their practical application often encounters challenges related to feature interpretability and model stability. Consequently, researchers often incorporate decision tree models (as show in Fig.4), which offer greater interpretability, into solvent design and screening workflows. Unlike “black-box” models such as neural networks, decision tree methods can identify key factors influencing solvent performance by recursively selecting optimal split points, from which predictive models are constructed [28,29]. Random forests further enhance precision and robustness by aggregating the results of multiple decision trees, rendering it a formidable tool in addressing complex datasets.

Yin et al. [31] used three ML models, MLP, DT, and Adaboost-DT, to predict the solubility of CO₂ in potassium and sodium amino acid solutions. Their results showed that Adaboost-DT had the highest predictive accuracy with an R² of 0.998, followed by the DT model with an R² of 0.98.

Huwaimel and Alharby [32] used DT and ensemble methods, including extra tree (ET) and gradient boosting (GB), to predict the solubility of the drug lenalidomide in supercritical carbon dioxide. These models were further optimized using the sine cosine algorithm (SCA), resulting in SCA-DT, SCA-ET and SCA-GB models. The optimized models achieved R² values of 0.932, 0.951, and 0.997, respectively.

In a recent study, Wang et al. [33] integrated the RF method with the COSMO-RS method to create a comprehensive database of CO₂ solubility in various DESs, encompassing 1011 data points across a wide range of pressure and temperature conditions. A nonlinear QSPR model was developed using the RF algorithm, enabling accurate predictions of CO₂ solubility in DESs. When compared with traditional thermodynamic calculation methods, the RF-based model demonstrated significantly higher R² values, underscoring its predictive superiority.

In a related study, Wang et al. [34] also applied the RF algorithm to predict key process parameters of post-combustion CO₂ capture (PCC), including CO₂ production rate, heat duty, CO₂ absorption efficiency, and dilute solution loading. Through modeling, the RF method successfully captured complex relationships among these parameters, offering valuable insights for optimizing CO₂ separation and improving the economic performance of PCC systems. Moreover, by aggregating predictions from multiple DTs, the RF approach reduced overfitting and improved model generalizability.

2.3 Support vector machine methods

Support vector machine (SVM) is an ML model based on the concept of a hyperplane, which is used to determine the optimal boundary between different classes (as show in Fig.5). This is achieved by mapping the data to a high-dimensional space using kernel functions, which enable the identification of a suitable linear partition plane within this space. SVM is particularly effective in addressing binary classification problems, and its generalization ability often surpasses that of neural network algorithms.

In a recent study, Kataoka et al. [35] applied an SVM model, based on molecular surface information, to predict two-phase separation during CO₂ absorption. They investigated the use of a mixture of alkanolamines and organic solvents as an absorber, predicting the CO₂ absorption rate of 61 mixed solvents and their phase states before and after absorption. The SVM method demonstrated exceptional capability in addressing small datasets, maintain an accuracy rate exceeding 90% despite the limited number of mixed solvents in the dataset, which contained only 61 entries.

Abdollahzadeh et al. [37] conducted a comparative analysis of seven ML methods to predict the densities of 149 DES. Their results demonstrated that the least squares support vector regression (LSSVR) method outperformed the other methods, with an R² of 0.99798. LSSVR, a variant of the SVM method, employs a Gaussian kernel function, which is particularly effective in capturing nonlinear relationships in high-dimensional spaces.

Boobier et al. [38] employed an ML approach combined with computational chemistry to predict the solubility of organic solvents in water. Nonlinear models, such as SVM and Gaussian process (GP), demonstrated superior performance, with high consistency across five distinct datasets. However, the overall model performance was highly dependent on the quality of the chosen descriptors and the dataset used for training.

2.4 GP regression methods

Gaussian process regression (GPR), Bayesian non-parametric model, assumes that the data follow a Gaussian distribution and uses prior distributions, combined with observed data, to infer posterior distributions (as shown in Fig.6) [39]. This framework enables GPR to provide stable predictive results even when sample sizes are limited. Consequently, GPR is frequently used in process optimization, especially when handling multivariate and nonlinear objective functions. In Bayesian optimization, GPR is employed to iteratively identify potential optimal conditions, effectively reducing the number of experimental iterations while enhancing the efficiency and accuracy of the optimization process.

Toots et al. [41] employed various ML methods to predict the partition coefficient, logK, of hydrocarbon gases in ILs based on cation types. Their results demonstrated that nonlinear methods, such as SVR and GPR, exhibited higher predictive accuracy compared to linear methods like multiple linear regression (MLR). Notably, the GPR method provides uncertainty estimates for its predictions, which are critical for assessing model reliability. Moreover, since data availability for ILs is often limited, the GPR method proves particularly advantageous by delivering stable predictions even under small-sample conditions.

The GPR method is also widely applied in the field of process optimization. Zhou et al. [42] applied GPR as a surrogate model in Aspen Plus to optimize a system process involving syngas conversion, the water-gas shift reaction, and monoethanolamine-based carbon capture. Compared to detailed process simulation-based optimization, the GPR-based approach significantly reduced optimization time, from an average of 1218.72 s to just 4.12 s. Furthermore, when compared to other surrogate modeling methods, such as response surface methodology (RSM), high-dimensional model representation (HDMR), and ANN, the GPR model demonstrated superior accuracy (with R² values close to 1 and a mean absolute error of 0.027) and optimization efficiency.

In summary, different ML methods have distinct advantages and limitations regarding data requirements, computational cost, accuracy, interpretability, and robustness. To facilitate comparison of these characteristics, common ML models are summarized and compared in Tab.1, enabling readers to quickly assess their applicability and selection criteria.

3 Design of ILs

The history of IL design and screening is illustrated in Fig.7. Since 1992, when Wilkes and Zaworotko [43] synthesized a 1-ethyl-3-methylimidazolium IL with excellent thermal stability, a surge of research into ILs as environmentally friendly solvents was initiated. In the early studies, researchers explored the structures of ILs by experimentally modifying the cations (such as imidazole and pyrimidine) and anions (such as BF4− and PF6−), which initially established the foundational concept of design based on the structure‒property relationship.

Around the year 2000s, with advancements in theoretical calculations and molecular simulation techniques (such as COSMO-RS and DFT), the structure‒property relationships of ILs began to be progressively quantified. As research data accumulated, investigators started to employ statistical regression methods to build correlations between IL structures and their properties, thereby laying the foundation for future studies involving big data and ML.

In recent years, with the development of artificial intelligence, modern ML algorithms (such as gradient boosting and deep learning) have been widely adopted in the design and screening of ILs, further accelerating the shift from trial-and-error methods to more precise design strategies. Such approaches not only integrate historical experimental data but also predict the performance of novel ILs in specific applications, providing strong support for the inverse design of ILs.

Therefore, this section is divided into two parts. First, the progress made in elucidating the interrelationship between the structure and properties of ILs through conventional experimental methods and theoretical calculations is reviewed. Then, a detailed introduction to the fundamental steps of the ML-based QSPR (ML-QSPR) approach is provided.

3.1 Experimental and theoretical calculation design methods

In the early studies on the structure‒property relationships of ILs, research primarily relied on experimental data accumulation and theoretical calculations to aid modeling. Initially, researchers modified cations, anions, and alkyl side chain lengths, measuring changes in viscosity, density, and melting points through experiments. Zhang et al. [44] analyzed experimental data to summarize the influence of the macroscopic structure of ILs on their properties. Taking the melting point as an example, as the anion size increases, the interaction with the cation weakens, leading to a gradual decrease in the melting point (see Fig.8). Similarly, as the alkyl side chain lengthens, the melting point decreases. However, when the carbon chain reaches a certain length, the melting point begins to increase again due to enhanced dispersion forces (see Fig.9).

To explain the structure‒property relationships of ILs from a microscopic perspective, Shukla and Saha [45] utilized DFT and natural bond orbital (NBO) analysis to calculate the stabilization energy E(2) and correlate it with experimentally measured melting points. Fig.10 illustrates the NBO charge distribution across different atoms using various colors. The green coloration of the C2-H atom indicates that it is the most electron-deficient among all carbon atoms in the imidazolium ring and alkyl chain and is the only atom that significantly interacts with halide anions. Moreover, for chloride-based ILs, the melting point exhibits a linear correlation with the E(2) value of C2-H (see Fig.11). In contrast, for iodide-based imidazolium ILs, the relationship between melting point and E(2) follows an exponential fit (see Fig.12). The study also found that for both chloride and iodide anions, a decrease in alkyl chain length increases the E(2) value and raises the melting point.

DFT methods can capture the microscopic electronic structure of the IL surface, while MD simulations provide deeper insights into the behavior of ILs. Klahn and Seduraman [46] employed MD simulations to investigate the key factors influencing CO₂ absorption in various ILs. The results (see Fig.13) indicate that CO₂ absorption is primarily determined by the strength of electrostatic cation–anion interactions. Larger ions exhibit lower ionic densities, resulting in weaker cation–anion attraction and increased average interionic spacing, which facilitates CO₂ incorporation. Additionally, cation functionalization introduces negatively charged regions on the cation surface, which enhances ionic aggregation when in proximity to other cations, thereby reducing CO₂ absorption. Moreover, when the anion size is extremely small, unavoidable cation-cation repulsion occurs, which, to some extent, promotes CO₂ absorption.

3.2 ML methods

However, these theoretical computational methods, including the challenges of solving the Kohn-Sham (K-S) equations, are constrained by high computational costs and resource-intensive requirements [47], limiting their application to small systems and short simulation timescales. Meanwhile, experimental methods are time-consuming and labor-intensive. Furthermore, these traditional design and screening methods do not focus on the correlation between structure and properties of ILs. As a result, ML methods, based on statistical principles and data science, have emerged as an efficient screening tool [48]. They not only process large volumes of data but also accurately identify underlying patterns and enable precise classification and prediction. By integrating data from experimental and theoretical computations, ML approaches can significantly enhance the precision of ILs screening, addressing the limitations of traditional methods.

ML methods, as data-driven models that integrate experimental and theoretical approaches, have been widely applied to the screening of ILs. Their primary advantage lies in significantly reducing costs and improving efficiency. In the study of ILs, QSAR and QSPR methods have been extensively utilized.

QSPR methods were first proposed by Hansch and Fujita [16] in their study of the relationship between molecular structure characteristics and properties. These methods have now been widely applied in chemical and pharmaceutical screening and design processes. They are generally categorized into QSAR and QSPR. QSAR primarily focuses on studying the relationship between molecular structure and biological activity or chemical reactivity. It is widely used in drug design, toxicity assessment, and catalyst screening. In contrast, QSPR focuses on the relationship between molecular structure and physical properties (such as density, viscosity, and solubility) and finds extensive applications in materials science and solvent screening. Early QSPR methods, represented by the model proposed by Katritzky et al. [49] in 1995, follow a five-step process, as shown in Fig.14(a), i.e., preparation of input data, 3D geometry optimization, calculation of descriptors, statistical analysis, and QSPR report and predictions. However, with the advancement of research and computer science, ML methods have been integrated into QSPR. Fig.14(b) illustrates the basic workflow of modern QSPR methods, which are widely used in the study of ILs.

The main processes of the current ML-QSPR model are as follows:

1) Data set preparation: Currently, there are two main approaches to constructing datasets. One approach involves obtaining data directly through experiments and theoretical calculations, while the other collects data from literature sources and open-access databases.

(2) Molecular descriptor calculation and selection: In this stage, specialized software tools (such as Dragon and RDKit) are used to extract a large number of descriptors from molecular structures. These descriptors provide multidimensional information ranging from 0D to 3D (as shown in Tab.2), including molecular composition, topological structure, geometric shape, and electronic properties. Subsequently, methods such as correlation analysis and principal component analysis (PCA) are employed for feature selection and dimensionality reduction, eliminating redundant and low-relevance variables. The result is a refined set of key features that are both physically and chemically meaningful, as well as highly predictive, thus laying a solid foundation for building ML models.

(3) Building the ML-QSPR model: Traditional QSPR models rely primarily on multiple linear regression methods to establish relationships between molecular descriptors and target properties. However, with the introduction of ML techniques, researchers have increasingly adopted nonlinear approaches to construct QSPR models, such as ANN, DT, SVM, and GPR. These machine learning models are more adept at capturing complex, nonlinear relationships between molecular features and physical properties.

(4) Model Validation: In the model validation phase, statistical metrics (see Tab.3) are used to comprehensively evaluate the model’s performance across different datasets, including training, validation, and test sets. Additionally, an external independent data set is employed to assess the model’s ability to generalize to new, unseen data. Furthermore, to enhance the interpretability of the model, methods such as SHAP (Shapley additive explanations) and LIME (local interpretable model-agnostic explanations) are applied. These methods provide both local and global explanations, shedding light on the contributions of individual descriptors to the model’s predictions. Finally, the model’s applicability domain is also assessed to ensure that the ML-QSPR model delivers reliable predictive performance across diverse chemical spaces and conditions.

Among these steps in ML-based QSAR and QSPR modeling, descriptor extraction stands out as a critical process, second only to the algorithm itself. This step directly influences the accuracy of predictions related to the properties and performance of ILs. Descriptor extraction is particularly vital in QSAR and QSPR methods, as it establishes a fundamental link between molecular structure and target properties.

The core objective of molecular structure feature selection and descriptor extraction is to effectively quantify the relationship between molecular structure and the desired properties, depending heavily on the accuracy and diversity of molecular descriptors. Currently, several categories of molecular structure descriptors are widely applied in ILs research:

Group contribution (GC) descriptors: GC methods treat a molecule as a combination of various functional groups, with each group contributing uniquely to the overall molecular properties, such as melting point, boiling point, and solubility. These descriptors leverage statistical characteristics and empirical data of groups to build predictive models for molecular properties. For example, Chen et al. [50] employed GC methods to predict properties of ILs, such as density, heat capacity, viscosity, and surface tension.

Molecular structure-based descriptors: Molecular fingerprints (MF) are one of the most common types of descriptors, encoding molecular structures as binary sequences. Each bit in the sequence represents the presence or absence of a specific structural feature in the molecule. This method is useful for quickly determining whether certain functional groups, which are related to particular properties, exist in the molecule. Simplified molecular input line entry system (SMILES) descriptors, encoded as ASCII strings, provide a detailed representation of molecular composition. SMILES descriptors are often combined with deep learning models to directly extract high-dimensional features from SMILES sequences, facilitating the prediction of complex properties. For example, Ding et al. [51] used ChemDraw to generate SMILES strings for ILs, which were then converted into MF descriptors using Python’s RDKit library to predict properties such as refractive index and viscosity.

Quantum chemistry-based descriptors: Quantum chemistry-based descriptors are typically derived from molecular electronic structures and energy distributions. One of the most frequently used descriptors is the HOMO-LUMO energy gap, which reflects molecular reactivity and stability. COSMO-RS descriptors, based on quantum chemical calculations, consider molecular geometry and electronic properties and are often used to describe molecular polarity. For example, Venkatraman et al. [52] utilized the HOMO-LUMO energy gap as a descriptor in conjunction with decision with tree-based ensemble learning methods to screen the density of ILs. Their results showed consistency when compared with COSMO-RS predictions.

In summary, molecular descriptors serve as a crucial link between the molecular structure of ILs and their performance predictions in ML models. GC-based descriptors provide a straightforward and efficient way to quantify the contribution of functional groups to the overall properties of a molecule. Structure-based descriptors excel in rapid screening and the extraction of high-dimensional features, while quantum chemistry-based descriptors provide insights into the electronic properties and geometric features of molecular interactions. The accuracy and diversity of descriptor selection play a decisive role in the predictive performance of ML models, particularly when applied to QSAR and QSPR modeling.

These descriptors collectively play an indispensable role in the success of ML methods applied to ILs screening and property prediction. By ensuring that descriptors capture relevant molecular features, ML models can effectively predict and optimize IL performance, advancing the field of solvent design and application.

4 ML assisted structure‒property prediction

This section provides a detailed overview of the specific applications of ML methods in conjunction with QSAR and QSPR approaches. These methodologies have proven to be invaluable in predicting a range of properties crucial for the design and screening of ILs and other chemical compounds. The properties commonly predicted include toxicity, activity, solubility, viscosity, and density. A summary of relevant studies and results is provided in Tab.4.

4.1 Quantitative structure-activity relationship

The use of QSAR modeling in IL research has become increasingly valuable in predicting various biological and environmental properties, particularly toxicity and biodegradability. QSAR is primarily employed to analyze the relationship between molecular structure and biological activity. This method has roots in drug design and bioactivity prediction, where it was initially applied to determine the biological effects of pharmaceutical compounds based on their chemical structures. Early studies often utilized this method in drug screening and bioactivity prediction. For example, Hansch [72] applied QSAR in drug design, using X-ray crystallography and molecular graphics to predict drug molecules’ biological activities with higher precision. In the field of IL research, QSAR is mainly used to predict properties such as toxicity and biodegradability. Peric et al. [53] utilized a GC method proposed by Luis et al. [73] to develop a QSAR model that divided IL molecules into three main components: anions, cations, and cation substituents. The study compiled an experimental EC50 value dataset for 13 ILs and combined it with 42 additional EC50 values from literature, establishing an ecotoxicity testing database. The QSAR model effectively predicted the ecotoxicity of ILs, showing agreement between experimental and predicted results. GC analysis revealed that IL anions had a negative contribution to toxicity, while cations had a positive contribution. This model provides guidance for synthesizing more sustainable ILs by identifying structural groups influencing ecotoxicity.

In addition, some researchers have chosen charge density distribution area (S_{\sigma }) and surface electrostatic potential area (S_{EP}) as descriptors for constructing QSAR models. Zhu et al. [54] compared two ML algorithms, MLR and extreme learning machine (ELM), to evaluate IL the toxicity of ILs toward Acetylcholinesterase (AChE). Using a QSAR model, the study selected S_{\sigma } and S_{EP} as descriptors. The results indicated that the ELM model had higher accuracy compared to MLR, in predicting IL toxicity to AChE.

Wu et al. [55] further advanced QSAR modeling by investigating the toxicity of ILs toward AChE, identifying key molecular descriptors affecting the model (see Fig.15). They used ML algorithms such as RF and XGBoost algorithms to analyze the correlation between important molecular descriptors and IL inhibitory capacity on AChE. Through feature importance analysis, they identified 14 key molecular descriptors that influence IL toxicity. Their findings indicate that cation structure plays a crucial role in determining AChE enzyme toxicity, with hydrophobic cations exhibiting stronger inhibitory effects. Other contributing factors such as branching degree, atomic mass, and partial charges were also found to contribute to toxicity. These findings enhance the understanding of the structure-activity relationship between ILs and AChE inhibition, providing valuable guidance for the design of environmentally friendly ILs.

Hodyna et al. [56] employed QSAR modeling and experimental studies to predict the antibacterial activity of imidazolium-based ILs. They used a variety of molecular descriptors, both 2D and 3D, to construct QSAR model and adopted RF and k-nearest neighbors (k-NN) algorithms for classification and prediction of IL antibacterial activity. The results showed that RF model based on diverse descriptors achieved an 88% classification accuracy and an 80% prediction accuracy for activity, highlighting the effectiveness of using diverse molecular descriptors in predicting IL antibacterial properties.

4.2 Quantitative structure‒property relationship

QSPR methods are widely applied in the study of ILs, focusing on the quantitative relationship between their molecular structures and key physicochemical properties, such as solubility, viscosity, density, and thermal stability. These methods are more commonly used in the study of ILs, particularly in the process of IL design and screening. They enable accurate prediction of performance and stability, providing scientific support for the selection and design of ILs tailored for specific applications.

4.2.1 Viscosity

Among the various properties of ILs, viscosity has been a consistent focus of research. During acid gas desorption, high-viscosity ILs hinder gas diffusion, highlighting the critical role of ML-assisted structure‒property relationship approaches in the screening and predicting ILs viscosity. Thus, some studies have adopted the GC method to correlate functional group information with viscosity data. Chen et al. [58] collected 8672 experimental viscosity data points for IL-H₂O mixtures from literature sources. Based on this data, they proposed a nonlinear model combining the GC method and ANN. The model’s inputs included IL structural information, composition, and temperature, with output being the mixture’s viscosity. The results showed that when the hidden layer contained four or five neurons, the model provided reliable viscosity predictions (see Fig.16). For instance, with four neurons, the MAE for the training set was 0.0091, and the coefficient of determination (R²) was 0.9962; for the test set, these values were 0.0095 and 0.9952, respectively. The proposed nonlinear ANN-GC model outperformed linear mixing models in predicting IL-H₂O mixture viscosities.

Compared to GC descriptors, MOLMAP descriptors offer significant advantages. While GC descriptors may struggle to fully capture the intrinsic relationship between structure and viscosity in complex systems, especially when handling ILs viscosity data, MOLMAP descriptors integrate detailed atomic-level information, providing a more comprehensive representation of the structural characteristics of ILs and their mixtures, thereby leading to more accurate viscosity predictions. Carrera et al. [57] combined RF and Kohonen neural network techniques to predict the viscosity of ILs and their mixtures. The research team collected data for 13798 chemical systems from the NIST ILThermo database. Chemical structures were processed using the Chemaxon platform to generate MOLMAP descriptors, which were refined using RF and Kohonen neural networks. This study found that an increase in cation chain length leads to higher viscosity (see Fig.17).

In the studies mentioned above, descriptors were directly used as variables in model training, which often resulted in high feature dimensions during the training process. To enhance the performance and efficiency, recent research has increasingly focused on the preprocessing of descriptors. Huang et al. [59] combined the UNIFAC model with ML methods to predict the viscosity of IL-water mixtures. Molecular structures were represented using SMILES, which were then converted into numerical matrices for analysis. The UNIFAC model parameters (R, Q) and Stokes radius (S⁺) were used as physical information features, along with SMILES descriptors and experimental data to construct the feature set. Among five ML algorithms tested, the CatBoost model performed the best (see in Fig.18), and its integration with the UNIFAC model enhanced prediction accuracy. Additionally, SHAP analysis revealed correlations and interactions between different features and viscosity (see Fig.19). Notably, the molar fraction of ILs (x_IL) and temperature (T) have a significant impact on viscosity, with higher x_IL and lower T leading to an increase in mixture viscosity.

Acar et al. [60] collected a viscosity dataset of 922 ILs from the ILThermo database, representing each IL’s molecular structure using 5272 molecular descriptors. They then performed descriptor cleaning, which involved removing columns with low variance, missing values, or null entries. They normalized all molecular descriptors and viscosity values using the Standard Scaler function from scikit-learn and further reduced the dimensionality of the original descriptor matrix using Pearson correlation coefficients, ultimately identifying 179 key molecular descriptors. The study employed two deep learning models, DNN and CNN, to predict the viscosities of ILs. The results demonstrated that both the DNN and CNN models achieved outstanding predictive performance on the test dataset. Furthermore, the deep learning models were used to identify the most influential molecular descriptors influencing IL viscosity. The analysis indicated that viscosity reduction could be achieved by decreasing the cation size, shortening the alkyl chains, and lowering the ionization potential/energy. Additionally, for the same cation, further reductions in the size, chain length, and hydrogen bonding of the anion could contribute to additional viscosity reduction.

4.2.2 Density

The density of ILs is a critical property in applications such as energy storage systems and separation technologies, as it directly affects their performance. In the field of IL density prediction, Liu et al. [62] proposed a hybrid model combining GC methods with ML algorithms to predict the essential properties, including density and heat capacity, of IL-organic solvent binary systems. They gathered isobaric heat capacity data for 17 ILs under varying temperatures and pressures, as well as density data for 129 ILs. Three ML methods—ANN, XGBoost, and LightGBM—were used to develop the model, with SHAP analysis employed to evaluate the importance of each structural feature and parameter in the predictions. The results showed that all three algorithms delivered accurate predictions, with ANN model performing the best. For density, the ANN model achieved an MAE of 0.0049 and an R² of 0.9942 (see Tab.5). SHAP analysis revealed that the mole fraction of ILs in binary systems had the most significant positive impact on the predicted properties (see Fig.20).

Additionally, in broader application scenarios, such as screening ILs with potential for sustainable applications, a more comprehensive consideration of various factors is necessary. Venkatraman et al. [52] developed a computational screening strategy to identify ILs with sustainable solvent design potential from a library of 8 million candidate ILs. They collected data for 10 properties of ILs, including density, from over 1500 literature sources and used combinatorial methods to generate a library of approximately 8 million ILs. They employed OpenBabel to generate three-dimensional molecular geometries and optimized them using MOPAC at the semi-empirical PM6 level. Molecular descriptors, including HOMO/LUMO energy levels, were calculated and used in tree-based ensemble learning methods to develop predictive models, evaluated through 5-fold cross-validation. The research team synthesized 15 ILs to validate the model predictions, with experimentally measured density values ranging from 1060 to 1457 kg/m³. The results showed consistency between the experimental values, the ML model predictions, and COSMO-RS computational results.

4.2.3 CO₂ solubility

In response to the challenge of global warming, ILs have garnered significant attention in the field of carbon capture due to their unique properties. Many research teams currently employ traditional molecular descriptor methods for quantitative structure‒property analysis of CO₂ solubility, aiming to gain deeper insights into the relationship between IL performance and molecular structure. Song et al. [64] compiled 10116 data points of CO₂ solubility in various ILs at different temperatures and pressures. They developed two models, ANN-GC and SVM-GC, using GC methods to predict CO₂ solubility. Comparative analysis revealed that ANN-GC performed slightly better than SVM-GC (see Tab.6), demonstrating its ability to make fast and accurate predictions. However, while these ML models are effective in computer-aided molecular design (CAMD), they are data-driven and lack a theoretical foundation. This limitation affects their predictive capability, particularly beyond the temperature and pressure ranges seen in the training data. Tian et al. [66] decomposed ILs into cation fragments, anion fragments, and neutral groups (ionic fragment contribution, IFC), and built QSPR models using SVM and ANN. By training on 13055 CO₂ solubility data points, the IFC-SVM method achieved higher predictive accuracy (R² = 0.9855) than the IFC-ANN method (R² = 0.9732, see Tab.7). While the IFC method yielded satisfactory results using ionic fragments as descriptors, it did not evaluate the contribution coefficients of these fragments. Furthermore, this descriptor-based approach lacked comprehensive molecular information and thermodynamic principles, limiting its interpretability.

Traditional molecular descriptor methods provide an effective approach for predicting CO₂ solubility, based on extensive experimental data and classical molecular features. However, as research progresses, there has been a shift toward exploring ILs design from a more microscopic perspective, focusing on electronic structures and atomic-level features to develop more efficient CO₂ capture solvents. Kuroki et al. [63] calculated various physicochemical properties of ILs, including geometric and electronic features, using quantum chemistry and COSMO-RS theoretical calculations. Using the GPR method, they predicted the Henry’s constant (HCO₂) for CO₂ solubility in ILs. Based on these predictions, the IL [P₆₆₆₁₄][PFOS] was synthesized and experimentally verified to have superior CO₂ absorption capabilities compared to the previously best-performing ILs. They found that CO₂ solubility increases in the order of PF6⁻ < TFSA⁻ < PFOS⁻, due to the higher number of fluorine atoms and S = O bonds in PFOS⁻, which enhance its interactions with CO₂ and result in the most stable Gibbs free energy of absorption.

Designing ILs from the perspective of electronic structures and atomic-level features provides valuable theoretical insights for enhancing the accuracy of CO₂ solubility predictions. However, in practical applications, it is equally important to understand the mechanisms behind model predictions and evaluate their reliability. This necessity has led to the development of interpretability methods that provide deeper insights into the prediction process. Jian et al. [65] innovatively applied graph neural network (GNN) to predict the CO₂ absorption capacity of ILs without relying on traditional molecular descriptors. Using a dataset of 10117 data points, they compared the predictive performance of GNNs with that of traditional ML methods (SVM, RF, XGBoost, and MLP). The GNN model outperformed traditional models, with an R² of 0.9884 (see Tab.8). Additionally, they developed an IL interpreter based on the GNN model, representing IL molecules as graphs with atoms as nodes and chemical bonds as edges. Node feature vectors encoded atomic properties, and the GNN model aggregated information from neighboring nodes to rank the importance of individual atoms and functional groups. Their study found that in physical absorption, anions play a dominant role, while fluorination of alkyl chains on cations and the presence of longer alkyl chains enhance CO₂ solubility by creating more free volume. Yang et al. [67] employed three ML methods (GPR, LightGBM, CatBoost) and three molecular descriptor types (GC, molecular structure descriptors, and a hybrid GC-MSD) to predict CO₂ solubility in ILs. All models showed good predictive performance (see Tab.9), with the CatBoost-GC-MSD model achieving the best results (R² = 0.9925). SHAP analysis was used for interpretability (see Fig.21), identifying pressure, temperature, Chi0, Kappa2, and EState_VSA10 as the top five influential features for CO₂ solubility.

Interpretability methods provide valuable insights into the relationship between molecular structures and CO₂ absorption by analyzing the internal mechanisms of predictive models. Additionally, evaluating uncertainty in model predictions is crucial, leading to research in uncertainty analysis methods to enhance the reliability of screening and design processes. Zhong et al. [68] explored four different molecular representation methods—molecular fingerprints, molecular descriptors, molecular images, and molecular graphs—and introduced a representing uncertainty (RU) method to quantify prediction uncertainty. Compared to traditional model uncertainty (MU) approaches, RU excelled in identifying unpredicted regions. Using this model, they screened 37 ILs with low viscosity, low toxicity, and high CO₂ absorption capacity from 1420 ILs. Experimental validation confirmed the reliability of the model predictions.

4.2.4 H₂S solubility

In industrial removal processes, H₂S and CO₂ often coexist, but their differing physicochemical properties, toxicity, and corrosiveness require distinct treatment methods. While research on H₂S is relatively limited, its unique industrial demands and the challenges of experimental studies make structure‒property-based solubility prediction particularly important. This imperative has driven researchers to conduct extensive studies in this area. Mousavi et al. [69] applied several advanced models, including CNNs, recurrent neural networks (RNNs), deep belief networks (DBNs), and DNNs based on initialized decision trees (DJINN), to predict H₂S solubility in ILs. The data set comprised 1516 data points for 37 different ILs, using IL chemical structures, temperature, and pressure as inputs. Among the models, CNN demonstrated the highest accuracy and speed in predicting H₂S solubility. Sensitivity analysis revealed that solubility increases with pressure, while temperature and the presence of ‒OH groups inversely affect solubility. In a follow-up study, Mousavi et al. [70] employed white-box ML approaches (group method of data handling, GMDH; genetic programming, GP), deep learning models (DBN), and ensemble methods (XGBoost) to further analyze and predict the same dataset. In addition to temperature and pressure, thermodynamic properties such as critical temperature (T_c), critical pressure (P_c), acentric factor (ω), boiling point (T_b), and molecular weight (M_w) were introduced as descriptors to comprehensively characterize ILs properties from a macroscopic thermodynamic perspective, aiming to explore their relationship with H₂S solubility. This study also found that ILs with longer alkyl chains have higher solubility under the same anion conditions (see Fig.22.). Their larger free volume enhances van der Waals interactions, accommodating more H2S molecules. Additionally, the fluorine content in the anion significantly affects H₂S solubility (see Fig.23.). At 313.15 K, increasing the fluorine content improves the solubility of H₂S in ILs under the same cation conditions.

The above studies employed white-box ML models for prediction but did not offer a detailed quantitative evaluation of the influence exerted by individual variables on H₂S solubility. To address this limitation, some studies adopted the SHAP method for model interpretability analysis. Liu et al. [71] combined DNN and RF methods with QSPR models, using various structural representations (e.g., molecular identifiers MI, MD, and combined MD-MI). They established the relationship between IL structure, temperature, pressure, and H₂S solubility. The results showed that structural representation significantly influenced prediction accuracy, with the ranking of MI > MD-MI > MD (see Tab.10). Among the ML models, DNN outperformed RF in terms of predictive accuracy, achieving an R² of 0.9923. Moreover, the study utilized the SHAP method to identify and quantify the effect and factors such as pressure, temperature, and specific molecular structural features (e.g., double bonds and carbon chain length) on H₂S solubility. This analysis provided valuable theoretical insights for designing and screening of ILs with high H₂S solubility.

5 Challenge of ML methods applied in ILs

Currently, ML methods demonstrate significant potential in the structure‒property prediction and design of ILs. However, large-scale applications still face numerous challenges. This section focuses on how to select appropriate ML models based on factors such as data volume and task complexity. Additionally, it emphasizes the importance of integrating ML with traditional experimental methods and molecular simulation techniques to achieve complementary advantages and enhance overall prediction accuracy. In practical engineering applications, multi-property prediction and screening of ILs are essential for achieving high-performance targeted optimization. However, incorporating multi-property prediction within a single model remains a complex issue. To address this, it provides an in-depth analysis of the current challenges and proposes corresponding recommendations, offering both theoretical and practical guidance for the precise design of ILs in the future.

5.1 Model selection based on data scale and task complexity

Based on the current state of research, it is clear that no single ML model can be universally applicable to all prediction tasks. Different models have their distinct advantages and limitations depending on data scale and task complexity. The strengths, weaknesses, and applicable conditions of several widely used ML methods are summarized (see Tab.11) to guide model selection and adjustment in practical applications.

When the data set is large and the task is complex, DNN or GNN models are particularly effective in leverage big data and automatically extract features, leading to high-precision predictions. Baskin et al. [61] compared the performance of various ML methods and molecular representations in constructing QSPR models to predict the physicochemical properties of ILs. Their study identified the optimal models and descriptor methods for different property predictions (see Tab.12). The results showed that transformer-based neural networks performed best, nonlinear methods outperformed linear methods, and SMILES-based modeling yielded better results than traditional molecular descriptors. However, these deep learning models typically require significant computational resources and are often difficult to interpret.

When the dataset is limited, SVM, GPR, and ensemble methods may be more suitable. In the study conducted by Mohan et al. [74], the predictive accuracy of various ML methods was compared on a small dataset containing sonic velocity data for 218 ILs (see Tab.13). The results showed that SVM and ensemble tree-based methods, particularly XGBoost, exhibited strong predictive performance for small-sample, high-dimensional data. Additionally, in the work of Mazari et al. [75], which focused on predicting the CO₂ solubility in [Bmim][PF6] (see Fig.24), GPR-based methods not only performed well under sparse data conditions but also provided uncertainty estimates for predictions, facilitating risk assessment.

Therefore, a stepwise strategy is recommended when using ML methods for structure‒property prediction. First, a simple linear regression model can be employed to capture overall trends and preliminarily identify key factors. Then, based on the data scale and task requirements, more complex models, such as neural networks or ensemble learning, can be introduced to refine the prediction of critical parameters. This approach not only enhances the reliability of prediction results but also mitigates the risks associated with insufficient data. Additionally, it provides clear guidance for selecting appropriate models in different application scenarios for the design of ILs.

5.2 Model interpretation through the integration of physical information

In traditional ML models, statistical indicators such as R² are commonly used to assess predictive accuracy. While these metrics provide an intuitive sense of model performance, they do not reveal how input features are utilized internally, nor do they shed light on the physical mechanisms underlying the prediction outcomes. This “black box” issue limits the deeper physical interpretation of the results. To address this challenge, many researchers have adopted the SHAP analysis method, which quantifies the contribution of each feature to the prediction, clearly demonstrating which structural descriptors play a critical role in predicting the properties of ILs.

However, the interpretability of the SHAP method is highly dependent on the model structure, and different models may yield varying feature importance rankings [76]. Moreover, although SHAP based on conventional structural descriptors can reveal statistical correlations, it does not capture the physical mechanisms at the atomic or molecular level, nor can it elucidate how intermolecular interactions, hydrogen bond networks, and electron distributions in ILs influence property predictions.

Therefore, integrating MD and DFT methods, based on theoretical calculations, is recommended for more precise characterizations. Li et al. [77] calculated electronic properties—such as the nucleophilicity index, electrophilicity index, dipole moment, and HOMO and LUMO energies—using DFT as supplementary descriptors. These were combined with experimental conditions and structural descriptors as input features to predict the yield of IL-catalyzed CO₂ cyclization reactions. In the RF model that achieved the highest predictive accuracy, SHAP analysis revealed that four of the five most important descriptors for reaction yield were experimental parameters (see Fig.25). Notably, the energy gap (ΔE_H–L₁) between the IL’s HOMO and the substrate’s LUMO had the greatest impact, with a smaller gap facilitating the reaction. This study provides a valuable reference for constructing ML models in multicomponent reactions by incorporating high-accuracy DFT-calculated electronic features with physicochemical interpretability, thereby enhancing overall predictive accuracy.

A subsequent study utilized an ML model incorporating DFT calculations to screen IL catalysts for CO₂ cycloaddition reactions. Li et al. [78] not only experimentally validated the six selected ILs but also employed DFT calculations to investigate the reaction mechanism after prediction (see Fig.26). The proposed mechanism consists of three main steps: IL-induced epichlorohydrin ring-opening, CO₂ insertion, and subsequent cyclization to form cyclic carbonate. Among these, the IL-induced substrate ring-opening is identified as the rate-determining step. This study highlights the dual role of DFT calculations, both in providing molecular descriptors and elucidating reaction mechanisms, offering a solid theoretical foundation for the rational design of ILs.

In summary, an integrated approach combining MD simulations, DFT calculations, and ML models offer a comprehensive explanatory framework that spans from the molecular to the system level. This strategy not only addresses the limitations of traditional “black-box” models but also provides a thorough and precise theoretical foundation for the targeted design of novel ILs.

5.3 Multi-objective property prediction and optimization

In the application of ILs, multiple physical property criteria often need to be met simultaneously. For example, in electrochemical CO₂ reduction, in addition to high CO₂ solubility, factors such as conductivity, viscosity, thermal stability, and safety must also be considered. Traditional single-objective prediction methods often struggle to optimize multiple properties simultaneously, typically improving one characteristic at the expense of others, making it difficult to achieve an overall optimal design. To address this issue, multi-objective property prediction has become an essential aspect of IL design.

In this context, multi-task learning offers an effective strategy. By incorporating a shared feature extraction layer in the model and constructing separate output layers for each target property, multi-task neural networks can simultaneously capture both the common influences and individual differences of IL structures on various property indices [79]. An integrated multi-task neural network can predict multiple properties, such as density, viscosity, surface tension, and melting point simultaneously (see Fig.27). The shared layer extracts general features that provide complementary information for different properties, while the independent output layers ensure that the optimization of each property occurs without interference from the others, thereby achieving multi-objective collaborative optimization.

In addition, multi-objective optimization based on Pareto Front analysis holds great application potential. Zhang et al. [80] formulated the design problem of ILs as a mixed-integer nonlinear programming (MINLP) problem, aiming to maximize CO₂ solubility under specified conditions. After identifying eight IL combinations with the highest CO₂ solubility, the sigma constraint method was employed by selecting 10 equally spaced sigma values between the minimum and maximum viscosities. This approach converted the multi-objective optimization problem into 10 single-objective optimization problems. Ultimately, this approach resulted in a Pareto-optimal curve, illustrating the trade-off between IL viscosity and CO₂ solubility (see Fig.28).

This section systematically examines the challenges and strategies for improving the application of ML in the design of IL. First, the choice of an appropriate model should depend on data availability and task complexity. Deep learning and GNNs are effective for capturing nonlinear relationships in complex tasks with abundant data, while SVM, GPR, and ensemble tree methods are more suitable for small datasets or simpler tasks. Second, traditional ML models, though evaluated using statistical metrics like R² and RMSE, often act as “black boxes” with limited interpretability. To address this, integrating theoretical computation methods, such as MD and DFT with ML models are recommended. This integration facilitates the calculation of physically meaningful descriptors, which can validate and supplement model interpretations, thereby creating a comprehensive explanatory framework that spans from the atomic-scale details to macroscopic properties. Lastly, multi-objective property prediction is essential for practical applications. This section highlights the role of multi-task learning and Pareto optimization in enabling multi-property optimization of ILs in complex scenarios. In conclusion, this section provides valuable technical guidance and theoretical insights for the precise design and multi-objective optimization of ILs in the future.

6 Conclusions

In the context of the global challenge of climate change, ILs have garnered widespread attention in carbon capture research due to their customizable properties. The integration of ML methods has significantly accelerated the design and screening of IL solvents, providing a data-driven approach to overcome the inefficiencies of traditional experimental and theoretical methods. This review offers an in-depth and comprehensive analysis of the current progress in ML applications for ILs design and screening, emphasizing the crucial role of ML in addressing the limitations of conventional approaches and advancing the field.

Recent studies on ILs using ML have increasingly moved away from solely relying on physical properties (e.g., molecular weight, acentric factor, critical temperature, and pressure) as input parameters for property prediction. Instead, the focus has shifted to modern QSAR and QSPR models, which leverage ML to effectively establish relationships between the molecular structures of ILs and their performance. This review highlights the applications of various structure-based descriptors in QSAR and QSPR methods, particularly in predicting key properties such as toxicity, bioactivity, viscosity, density, and gas solubility of ILs. Currently, ML-based research in the structure‒property relationships of ILs has evolved from simple models to advanced and hybrid models, such as GNN [65] and VAE-MLP-PSO [81]. These advanced methods have demonstrated significant promise in capturing the nonlinear relationships between complex structures and properties.

This review examines the challenges faced in applying ML to ILs design and screening. It provides an overview of the advantages and limitations of different ML models when dealing with datasets of varying sizes and task complexities. For complex tasks, it recommends a stepwise strategy that not only enhances the reliability of prediction results but also mitigates the risk of overfitting. This approach offers clear guidance on model selection for IL design across different application scenarios. To further address the challenge of interpreting ML “black-box” models, it introduces a ML interpretability approach that integrates MD and DFT. By incorporating physicochemical information obtained from theoretical calculations as descriptors, this approach provides insights into the role of key molecular structures in property prediction. It also enables an atomic-level explanation of intermolecular interactions in ILs, validating the physicochemical interpretability of ML models.

Finally, this review emphasizes the importance of multi-property prediction in the engineering design of ILs. In applications like post-combustion CO₂ capture, ILs used as absorbents must exhibit low toxicity, low viscosity, and, high CO₂ solubility. To achieve the simultaneous optimization of multiple properties, it recommends using multi-task learning, Pareto optimization, and ensemble learning methods. These approaches provide comprehensive theoretical and practical framework for the targeted design of novel ILs.

In conclusion, ML methods hold great promise in the field of ILs research, with the potential to achieve significant breakthroughs in solvent screening and property prediction. By addressing critical challenges related to model selection, data availability, and interpretability, ML can substantially accelerate the development of ILs as sustainable and multifunctional solvents. This, in turn, will contribute to the transformation of global sustainable energy and chemical industries.

References

[1] Erythropel H C, Zimmerman J B, de Winter T M. . The Green ChemisTREE: 20 years after taking root with the 12 principles. Green Chemistry, 2018, 20(9): 1929–1961

[2] Shah P, Parikh S, Shah M. . A holistic review on application of green solvents and replacement study for conventional solvents. Biomass Conversion and Biorefinery, 2022, 12(5): 1985–1999

[3] GhandiK. A review of ionic liquids, their limits and applications. Green and Sustainable Chemistry. 2014, 4(1): 44–45

[4] Zeng S, Zhang X, Bai L. . Ionic-liquid-based CO₂ capture systems: Structure, interaction and process. Chemical Reviews, 2017, 117(14): 9625–9673

[5] Wang M, Zhang L, Liu H. . Studies on CO₂ absorption performance by imidazole-based ionic liquid mixtures. Journal of Fuel Chemistry & Technology, 2012, 40(10): 1264–1268

[6] Liu F, Shen Y, Shen L. . Novel amino-functionalized ionic liquid/organic solvent with low viscosity for CO₂ capture. Environmental Science & Technology, 2020, 54(6): 3520–3529

[7] Pârvulescu V I, Hardacre C. Catalysis in ionic liquids. Chemical Reviews, 2007, 107(6): 2615–2665

[8] Guerfi A, Dontigny M, Charest P. . Improved electrolytes for Li-ion batteries: Mixtures of ionic liquid and organic electrolyte with enhanced safety and electrochemical performance. Journal of Power Sources, 2010, 195(3): 845–852

[9] RogersR D, Seddon K R. Ionic liquids–solvents of the future? Science, 2003, 302(5646): 792–793

[10] Gardas R L, Coutinho J A. A group contribution method for viscosity estimation of ionic liquids. Fluid Phase Equilibria, 2008, 266(1–2): 195–201

[11] Kolbeck C, Lehmann J, Lovelock K. . Density and surface tension of ionic liquids. Journal of Physical Chemistry B, 2010, 114(51): 17025–17036

[12] Wang Y, Jiang W, Yan T. . Understanding ionic liquids through atomistic and coarse-grained molecular dynamics simulations. Accounts of Chemical Research, 2007, 40(11): 1193–1199

[13] Kroon M C, Buijs W, Peters C J. . Quantum chemical aided prediction of the thermal decomposition mechanisms and temperatures of ionic liquids. Thermochimica Acta, 2007, 465(1–2): 40–47

[14] Lei Z, Zhang J, Li Q. . UNIFAC model for ionic liquids. Industrial & Engineering Chemistry Research, 2009, 48(5): 2697–2704

[15] Seybold P G, May M, Bagal U A. Molecular structure: Property relationships. Journal of Chemical Education, 1987, 64(7): 575

[16] Hansch C, Fujita T. p-σ-π Analysis. A method for the correlation of biological activity and chemical structure. Journal of the American Chemical Society, 1964, 86(8): 1616–1626

[17] Sun Y, Chen M, Zhao Y. . Machine learning assisted QSPR model for prediction of ionic liquid’s refractive index and viscosity: The effect of representations of ionic liquid and ensemble model development. Journal of Molecular Liquids, 2021, 333: 115970

[18] ZouJ, HanY, SoS S. Overview of artificial neural networks. In: Livingstone D J, ed. Artificial Neural Networks: Methods and Applications. Totowa, NJ: Humana Press, 2009

[19] Agatonovic-Kustrin S, Beresford R. Basic concepts of artificial neural network (ANN) modeling and its application in pharmaceutical research. Journal of Pharmaceutical and Biomedical Analysis, 2000, 22(5): 717–727

[20] Ardeshiri A, Rashidi H. Performance of water-lean solvent for postcombustion carbon dioxide capture in a process-intensified absorber: Experimental, modeling, and optimization using RSM and ML. Industrial & Engineering Chemistry Research, 2023, 62(48): 20821–20832

[21] Zhu X, Khosravi M, Vaferi B. . Application of machine learning methods for estimating and comparing the sulfur dioxide absorption capacity of a variety of deep eutectic solvents. Journal of Cleaner Production, 2022, 363: 132465

[22] Zhang J, Wang Q, Su Y. . An accurate and interpretable deep learning model for environmental properties prediction using hybrid molecular representations. AIChE Journal, 2022, 68(6): e17634

[23] Bischl B, Binder M, Lang M. . Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews. Data Mining and Knowledge Discovery, 2023, 13(2): e1484

[24] GoodfellowI. Deep Learning. Cambridge: MIT Press, 2016

[25] Gu J, Wang Z, Kuen J. . Recent advances in convolutional neural networks. Pattern Recognition, 2018, 77: 354–377

[26] Bahmaninia H, Shateri M, Atashrouz S. . Predicting the equilibrium solubility of CO₂ in alcohols, ketones, and glycol ethers: Application of ensemble learning and deep learning approaches. Fluid Phase Equilibria, 2023, 567: 113712

[27] Chen G, Song Z, Qi Z. Transformer-convolutional neural network for surface charge density profile prediction: Enabling high-throughput solvent screening with COSMO-SAC. Chemical Engineering Science, 2021, 246: 117002

[28] BouzidaY,Cuppens F. Neural networks vs. decision trees for intrusion detection. In: IEEE/IST Workshop on Monitoring, Attack Detection and Mitigation (MonAM), 28:29

[29] Kabari L, Nwachukwu E. Decision support system using decision tree and neural networks. Computer Engineering and Intelligent Systems., 2013, 4(7): 8–20

[30] Nakhaei-Kohani R, Amiri-Ramsheh B, Pourmahdi M. . Extensive data analysis and modelling of carbon dioxide solubility in ionic liquids using chemical structure-based ensemble learning approaches. Fluid Phase Equilibria, 2024, 585: 114166

[31] Yin G, Jameel Ibrahim Alazzawi F, Bokov D. . Multiple machine learning models for prediction of CO₂ solubility in potassium and sodium based amino acid salt solutions. Arabian Journal of Chemistry, 2022, 15(3): 103608

[32] Huwaimel B, Alharby T N. Development of computational intelligence models for assessment of drug nanonization using green chemistry technique: Improvement of drug solubility. Case Studies in Thermal Engineering, 2023, 45: 103005

[33] Wang J, Song Z, Chen L. . Prediction of CO₂ solubility in deep eutectic solvents using random forest model based on COSMO-RS-derived descriptors. Green Chemical Engineering., 2021, 2(4): 431–440

[34] Wang X, Chan C W, Li T. High accuracy prediction of the post-combustion carbon capture process parameters using the Decision Forest approach. Chemical Engineering Science, 2024, 290: 119878

[35] Kataoka T, Hao Y, Hung Y C. . Prediction of biphasic separation in CO₂ absorption using a molecular surface information-based machine learning model. Environmental Science. Processes & Impacts, 2022, 24(12): 2409–2418

[36] Huang M W, Chen C W, Lin W C. . SVM and SVM ensembles in breast cancer prediction. PLoS One, 2017, 12(1): e0161501

[37] Abdollahzadeh M, Khosravi M, Hajipour Khire Masjidi B. . Estimating the density of deep eutectic solvents applying supervised machine learning techniques. Scientific Reports, 2022, 12(1): 4954

[38] Boobier S, Hose D R J, Blacker A J. . Machine learning with physicochemical relationships: Solubility prediction in organic solvents and water. Nature Communications, 2020, 11(1): 5753

[39] Schulz E, Speekenbrink M, Krause A. A tutorial on Gaussian process regression: Modelling, exploring, and exploiting functions. Journal of Mathematical Psychology, 2018, 85: 1–16

[40] Deringer V L, Bartók A P, Bernstein N. . Gaussian process regression for materials and molecules. Chemical Reviews, 2021, 121(16): 10073–10141

[41] Toots K M, Sild S, Leis J. . Machine learning quantitative structure‒property relationships as a function of ionic liquid cations for the gas-ionic liquid partition coefficient of hydrocarbons. International Journal of Molecular Sciences, 2022, 23(14): 7534

[42] Zhou J, Liu C, Ren J. . Targeting carbon-neutral waste reduction: Novel process design, modelling and optimization for converting medical waste into hydrogen. Energy, 2024, 310: 133272

[43] Wilkes J S, Zaworotko M J. Air and water stable 1-ethyl-3-methylimidazolium based ionic liquids. Journal of the Chemical Society. Chemical Communications, 1992, (13): 965–967

[44] Zhang S, Sun N, He X. . Physical properties of ionic liquids: Database and evaluation. Journal of Physical and Chemical Reference Data, 2006, 35(4): 1475–1517

[45] Shukla M, Saha S. Relationship between stabilization energy and thermophysical properties of different imidazolium ionic liquids: DFT studies. Computational & Theoretical Chemistry, 2013, 1015: 27–33

[46] Klahn M, Seduraman A. What determines CO₂ solubility in ionic liquids? A molecular simulation study. Journal of Physical Chemistry B, 2015, 119(31): 10066–10078

[47] del RioB G, Phan B, RamprasadR. A deep learning framework to emulate density functional theory. npj Computational Materials. 2023, 9(1): 158

[48] Zhong S, Zhang K, Bagheri M. . Machine learning: New ideas and tools in environmental science and engineering. Environmental Science & Technology, 2021, 55(19): 12741–12754

[49] Katritzky A R, Lobanov V S, Karelson M. QSPR: The correlation and quantitative prediction of chemical and physical properties from structure. Chemical Society Reviews, 1995, 24(4): 279–287

[50] Chen Y, Kontogeorgis G M, Woodley J M. Group contribution based estimation method for properties of ionic liquids. Industrial & Engineering Chemistry Research, 2019, 58(10): 4277–4292

[51] Ding Y, Chen M, Guo C. . Molecular fingerprint-based machine learning assisted QSAR model development for prediction of ionic liquid properties. Journal of Molecular Liquids, 2021, 326: 115212

[52] Venkatraman V, Evjen S, Lethesh K C. . Rapid, comprehensive screening of ionic liquids towards sustainable applications. Sustainable Energy & Fuels, 2019, 3(10): 2798–2808

[53] Peric B, Sierra J, Marti E. . Quantitative structure-activity relationship (QSAR) prediction of (eco)toxicity of short aliphatic protic ionic liquids. Ecotoxicology and Environmental Safety, 2015, 115: 257–262

[54] Zhu P, Kang X, Zhao Y. . Predicting the toxicity of ionic liquids toward acetylcholinesterase enzymes using novel QSAR models. International Journal of Molecular Sciences, 2019, 20(9): 2186

[55] Wu X, Gong J, Ren S. . A machine learning-based QSAR model reveals important molecular features for understanding the potential inhibition mechanism of ionic liquids to acetylcholinesterase. Science of the Total Environment, 2024, 915: 169974

[56] Hodyna D, Kovalishyn V, Rogalsky S. . Antibacterial activity of imidazolium-based ionic liquids investigated by QSAR modeling and experimental studies. Chemical Biology & Drug Design, 2016, 88(3): 422–433

[57] Carrera G V S M, Nunes da Ponte M. Machine-learning approaches to tune descriptors and predict the viscosities of ionic liquids and their mixtures. Chemistry Methods, 2020, 1(5): 214–223

[58] Chen Y, Peng B, Kontogeorgis G M. . Machine learning for the prediction of viscosity of ionic liquid-water mixtures. Journal of Molecular Liquids, 2022, 350: 118546

[59] Huang M, Deng J, Jia G. Predicting viscosity of ionic liquids-water mixtures by bridging UNIFAC modeling with interpretable machine learning. Journal of Molecular Liquids, 2023, 383: 122059

[60] Acar Z, Nguyen P, Cui X. . Room temperature ionic liquids viscosity prediction from deep-learning models. Energy Materials, 2023, 3: 300039

[61] Baskin I, Epshtein A, Ein-Eli Y. Benchmarking machine learning methods for modeling physical properties of ionic liquids. Journal of Molecular Liquids, 2022, 351: 118616

[62] Liu X, Gao J, Chen Y. . Machine learning-assisted modeling study on the density and heat capacity of ionic liquid-organic solvent binary systems. Journal of Molecular Liquids, 2023, 390: 122972

[63] Kuroki N, Suzuki Y, Kodama D. . Machine learning-boosted design of ionic liquids for CO₂ absorption and experimental verification. Journal of Physical Chemistry B, 2023, 127(9): 2022–2027

[64] Song Z, Shi H, Zhang X. . Prediction of CO₂ solubility in ionic liquids using machine learning methods. Chemical Engineering Science, 2020, 223: 115752

[65] Jian Y, Wang Y, Farimani A B. Predicting CO₂ absorption in ionic liquids with molecular descriptors and explainable graph neural networks. ACS Sustainable Chemistry & Engineering, 2022, 10(50): 16681–16691

[66] Tian Y, Wang X, Liu Y. . Prediction of CO₂ and N₂ solubility in ionic liquids using a combination of ionic fragments contribution and machine learning methods. Journal of Molecular Liquids, 2023, 383: 112066

[67] Yang A, Sun S, Su Y. . Insight to the prediction of CO₂ solubility in ionic liquids based on the interpretable machine learning model. Chemical Engineering Science, 2024, 297: 120266

[68] Zhong S, Chen Y, Li J. . Screening environmentally benign ionic liquids for CO₂ absorption using representation uncertainty-based machine learning. Environmental Science & Technology Letters, 2024, 11(11): 1193–1199

[69] Mousavi S P, Atashrouz S, Nakhaei-Kohani R. . Modeling of H₂S solubility in ionic liquids using deep learning: A chemical structure-based approach. Journal of Molecular Liquids, 2022, 351: 118418

[70] Mousavi S P, Nakhaei-Kohani R, Atashrouz S. . Modeling of H₂S solubility in ionic liquids: Comparison of white-box machine learning, deep learning and ensemble learning approaches. Scientific Reports, 2023, 13(1): 7946

[71] Liu T, Dong Z, Zhu W. . Prediction of the solubility of acid gas hydrogen sulfide in green solvent ionic liquids via quantitative structure–property relationship models based on the molecular structure. ACS Sustainable Chemistry & Engineering, 2023, 11(9): 3917–3931

[72] Hansch C. The physicochemical approach to drug design and discovery (QSAR). Drug Development Research, 1981, 1(4): 267–309

[73] Luis P, Ortiz I, Aldaco R. . A novel group contribution method in the development of a QSAR for predicting the toxicity (Vibrio fischeri EC50) of ionic liquids. Ecotoxicology and Environmental Safety, 2007, 67(3): 423–429

[74] Mohan M, Smith M D, Demerdash O. . Predictive understanding of the surface tension and velocity of sound in ionic liquids using machine learning. Journal of Chemical Physics, 2023, 158(21): 214502

[75] Mazari S A, Siyal A R, Solangi N H. . Prediction of thermo-physical properties of 1-butyl-3-methylimidazolium hexafluorophosphate for CO₂ capture using machine learning models. Journal of Molecular Liquids, 2021, 327: 114785

[76] Lundberg S M, Lee S I. A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 2017, 30: 4765–4774

[77] Li J, Dong S, An B. . Machine learning for the yield prediction of CO₂ cyclization reaction catalyzed by the ionic liquids. Fuel, 2023, 335: 126942

[78] Li J, Qi X, Zhang Z. . Screening of ionic liquids for efficient CO₂ cycloaddition catalysis under mild condition: A combined machine learning and DFT approach. ACS Sustainable Chemistry & Engineering, 2024, 12(48): 17512–17522

[79] Lemaoui T, Eid T, Darwish A S. . Revolutionizing inverse design of ionic liquids through the multi-property prediction of over 300,000 novel variants using ensemble deep learning. Materials Science and Engineering R Reports, 2024, 159: 100798

[80] Zhang X, Wang J, Song Z. . Data-driven ionic liquid design for CO₂ capture: Molecular structure optimization and DFT verification. Industrial & Engineering Chemistry Research, 2021, 60(27): 9992–10000

[81] Liu X, Chu J, Zhang Z. . Data-driven multi-objective molecular design of ionic liquid with high generation efficiency on small dataset. Materials & Design, 2022, 220: 110888

RIGHTS & PERMISSIONS

Higher Education Press 2025