Ionic liquid diffusion coefficient unified prediction method and system based on interpretable machine learning

Abstract

The invention belongs to the technical field of chemical informatics and machine learning crossing, and relates to an interpretable machine learning method for predicting the diffusion coefficient of an ionic liquid solution system. According to the method, three data sets of cation self-diffusion, anion self-diffusion and ionic liquid-solute mutual diffusion are respectively constructed by collecting experimental diffusion data of different ionic liquids and solute systems thereof in a wide temperature and wide pressure range. The method comprises the following steps: extracting multi-scale molecular descriptors from an SMILES structure, including a molecular structure descriptor (MSD) of global information and a fragment feature descriptor (FFD) of local information, and forming a high-dimensional feature matrix; and after feature engineering processing, modeling is carried out by adopting a multi-layer perceptron (MLP) and a random forest (RF) model, and automatic hyper-parameter adjustment is realized through Bayesian optimization. After the model is subjected to cross validation evaluation, the MSD-MLP model has the best performance in all diffusion behaviors, and the prediction decision coefficient R exceeds 0.99. The model is further subjected to interpretability analysis in combination with a Shapril additive feature interpretation (SHAP) method, temperature, electrostatic surface area, molecular structure complexity and other key features influencing the diffusion coefficient are identified, and main physical and chemical driving factors and structural dependency rules of the ionic liquid diffusion process are clarified.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of cheminformatics and machine learning, specifically relating to a method and system for predicting the diffusion coefficient of ionic liquids based on interpretable machine learning. This method can simultaneously predict the self-diffusion and inter-diffusion coefficients of ionic liquids within a unified model framework, and is applicable to electrolyte optimization, separation medium design, and mass transfer performance regulation in green chemical processes. Background Technology

[0002] Ionic liquids, as low-melting-point salts composed of organic cations and inorganic or organic anions, have been widely used in electrolyte materials, catalytic systems, and adsorption separation due to their excellent electrochemical stability, low volatility, and designability. The diffusion coefficient is a crucial kinetic parameter characterizing the migration ability of ions or molecules in a liquid phase, playing a decisive role in understanding the transport behavior, viscosity characteristics, and conductivity of ionic liquids.

[0003] Traditional experimental measurements and molecular dynamics simulations, while capable of obtaining diffusion coefficients, both have significant limitations: experimental measurements are costly, time-consuming, and highly condition-dependent; while molecular dynamics simulations can reveal microscopic mechanisms, they are computationally expensive and have limited applicability to specific systems. Existing diffusion prediction methods based on empirical correlations or single-system modeling often fail to account for the differences between various ionic liquid systems, lacking a unified prediction framework and physical interpretation capabilities.

[0004] With the rapid development of machine learning methods in physical chemistry, their application in modeling the properties of ionic liquids has shown significant advantages. However, most current models are still built only for a single diffusion type (such as self-diffusion of cations or anions), failing to achieve unified modeling of multiple systems. Furthermore, the "black box" nature of these models limits their application in structure-property analysis and mechanistic interpretation. Therefore, there is an urgent need to develop an intelligent prediction method that can accurately predict self-diffusion and inter-diffusion coefficients within a unified framework and possesses interpretability. Summary of the Invention

[0005] To overcome the problems of poor model applicability, insufficient interpretability, and low prediction efficiency in existing technologies, this invention proposes a method and system for predicting the diffusion coefficient of ionic liquids based on interpretable machine learning. This method constructs a unified modeling framework and achieves high-precision prediction of cation self-diffusion, anion self-diffusion, and ionic liquid-solute interdiffusion by combining multi-source experimental data, multi-scale feature extraction, feature engineering processing, and interpretable algorithms.

[0006] The main innovative points of this invention include: 1. For the first time, a unified machine learning framework was established to achieve collaborative prediction of three types of diffusion coefficients; 2. Constructing a multi-scale feature system includes molecular structure descriptors (MSDs) with global information and fragment feature descriptors (FFDs) with local information; 3. Random Forest (RF) and Multilayer Perceptron (MLP) models are adopted, and automatic hyperparameter adjustment is achieved through Bayesian optimization; 4. The model's prediction results, evaluated by multiple indicators, demonstrate excellent stability and generalization ability.

[0007] 5. Use the Shapley Additive Characteristic Interpretation (SHAP) method to perform character importance analysis and identify key molecular structural factors affecting diffusion behavior; This invention can significantly improve the accuracy and interpretability of diffusion coefficient prediction, providing a theoretical basis and technical support for understanding the mechanism of ionic liquid transport behavior and functional design.

[0008] [Technical problems to be solved] This invention aims to solve the following key problems: 1. Overcome the high cost and low throughput problems of traditional experimental and molecular dynamics methods in diffusion coefficient research; 2. Achieve unified modeling and high-precision prediction of self-diffusion and inter-diffusion of ionic liquids; 3. Provides model interpretability analysis to achieve mechanistic mapping from molecular structure to diffusion performance.

[0009] [Technical Solution]

[0010] The proposed method for predicting the diffusion coefficient of ionic liquids based on interpretable machine learning includes the following steps: Data Collection and Construction: Experimental data on cation self-diffusion, anion self-diffusion, and ionic liquid-solute interdiffusion were collected from public databases and literature, covering different temperature and pressure conditions. The data were organized into three datasets based on system composition, corresponding to the three diffusion types.

[0011] Feature extraction and preprocessing: Multi-scale molecular descriptors, including MSDs with global information and FFDs with local information, are extracted from the SMILES structures of ionic liquids and solutes. High-quality feature matrices are obtained through missing value removal, unique value filtering, collinearity removal, and normalization, and are then combined with environmental parameters (temperature, mole fraction) to form the model input set.

[0012] Model Building and Training: Two types of machine learning models, RF and MLP, were constructed, and Bayesian optimization was used to systematically search for their hyperparameters. The optimization objective was to maximize R² on the test set, and the process was iterated for 500 rounds to obtain the globally optimal parameter combination.

[0013] Model performance evaluation: The predictive performance of the model is systematically evaluated using statistical indicators such as the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) to comprehensively measure the model's performance in terms of accuracy, stability, and generalization ability.

[0014] Model interpretability analysis: Based on the SHAP method, the contribution of each feature to the predicted output is calculated, and key molecular features affecting diffusion behavior are identified, such as temperature, electrostatic surface area (PEOE_VSA type features), molecular complexity (Kappa index), specific groups, etc., thereby revealing the main physicochemical and structural driving factors of the diffusion process.

[0015] [Beneficial Effects]

[0016] Compared with the prior art, the present invention has the following significant advantages: 1. Unified prediction framework: For the first time, it achieves unified modeling of self-diffusion and inter-diffusion, breaking through the limitation of single-system modeling in traditional methods; 2. High prediction accuracy and generalization ability: The model's R² exceeds 0.99 on the test set, demonstrating excellent stability and cross-system applicability; 3. High interpretability: The SHAP method is introduced to reveal the key structural driving factors of diffusion, achieving molecular-level mechanistic analysis; 4. High efficiency and scalability: Bayesian optimization significantly improves model training efficiency and can be extended to other thermodynamic or kinetic property prediction tasks.

[0017] This invention not only provides a reliable computational tool for the efficient prediction of ionic liquid diffusion coefficients, but also establishes a general modeling framework applicable to the study of system structure-performance relationships in complex chemical systems. The method exhibits good scalability and can be further extended to predicting other thermodynamic or transport properties, providing solid computational support for the performance optimization and molecular design of ionic liquid systems. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall modeling process of the present invention; Detailed Implementation

[0019] This embodiment provides a machine learning-based method for predicting the diffusion coefficient of ionic liquids, applicable to the modeling and analysis of self-diffusion and interdiffusion in ionic liquid-solute systems. The method includes the following steps: (1) Dataset Construction: Experimental data for three categories—cation self-diffusion coefficient (DC), anion self-diffusion coefficient (DA), and ionic liquid-solute interdiffusion coefficient (DM)—were collected from publicly available literature and the NIST ILThermo database (v2.0). The data covered a temperature range of 253–373 K and a pressure range of 1–2500 bar. After data cleaning, each sample record was ensured to include the SMILES structural formula of the ionic liquid, temperature, pressure, mole fraction of the solute, and diffusion coefficient. Finally, three standardized datasets were formed.

[0020] (2) Feature extraction: Using the cheminformatics tool RDKit, the SMILES structures of all ionic liquids and solutes were analyzed, and two types of complementary molecular descriptors were extracted: Molecular Structure Descriptor (MTD): A 208-dimensional two-dimensional structure descriptor is calculated from SMILES, including molecular weight, topological polar surface area (TPSA), number of aromatic atoms, hydrogen bond donor / acceptor count, Wiener index, Balaban index, etc.; combined with temperature and pressure variables, a 210-dimensional feature matrix is ​​formed. For ionic liquid-solute systems, descriptors of ionic liquid and solute are extracted in parallel through a two-way structure verification mechanism, generating 208-dimensional feature vectors respectively. Combined with temperature (T), pressure (P), and mole fraction (MF_Solvent), a 419-dimensional co-located feature matrix is ​​formed.

[0021] Fragment Feature Descriptor (FFD): Based on local atomic chemical environment information (element type, chirality, charge, bond type, etc.), a fragment dictionary is generated, mapping each molecule to a fixed-length sparse vector to form an interpretable atomic-level feature matrix.

[0022] (3) Feature preprocessing: Perform the following operations on the generated feature matrix: 1. Remove features with a missing rate exceeding 60%. 2. Delete columns with unique values; 3. Calculate the Pearson correlation coefficient and use 0.98 as the threshold to remove highly collinear features; 4. Use Z-score standardization to unify the numerical scale; 5. Divide the training set and the test set into an 8:2 ratio to ensure randomness and repeatability.

[0023] (4) Model Hyperparameter Optimization: To further improve the accuracy and generalization ability of the constructed model in the task of predicting the diffusion coefficient of ionic liquids, this invention introduces a global optimization algorithm—Bayesian optimization—for systematically tuning the hyperparameters of various machine learning models. Bayesian optimization is an intelligent global search method based on a surrogate model, which is particularly suitable for problems with high computational cost of objective function, complex parameter space dimension, and inability to directly obtain gradient information. It has significant advantages in hyperparameter optimization of machine learning models. This method dynamically updates the posterior distribution of the objective function and continuously optimizes the sampling strategy based on historical search results, thereby achieving efficient approximation of the optimal solution with fewer iterations.

[0024] The goal of the optimization process is to maximize the R-value of the model on the test set. 2 Even on new, unseen data, the model maintains excellent predictive performance. Optimized hyperparameters include, but are not limited to, the following: For the RF model, the optimization parameters include the number of decision trees and the maximum tree depth; For MLP models, optimization parameters include the number of hidden layers, the number of neurons in each layer, the initial learning rate, and the optimizer type.

[0025] After optimization, the resulting model exhibits higher fitting accuracy and stability on both the training and test sets. Specifically, the MSD-MLP model achieves R² exceeding 0.99 in both self-diffusion and inter-diffusion tasks, significantly outperforming the unoptimized default model configuration. This optimization strategy effectively enhances the model's sensitivity to complex molecular structures and properties, providing a solid algorithmic foundation for high-precision prediction of ionic liquid diffusion behavior.

[0026] (5) Model Evaluation and Interpretation: After the model training is completed, R is first used to evaluate and interpret the model. 2The model performance was comprehensively evaluated using multiple evaluation metrics, including MAE and RMSE. Evaluation results show that the MSD-MLP model exhibits the best prediction accuracy on both the test and training sets, with the smallest error and the highest good fit. Furthermore, to further improve the model's interpretability, this invention employs the SHAP method to analyze the model's prediction results. The SHAP method reveals the importance of each feature in diffusion behavior by calculating the contribution of each input feature to the final prediction result. This method allows for the quantitative analysis of the positive or negative impact of each feature on the prediction results. This invention reveals the main driving factors and mechanisms influencing the diffusion behavior of ionic liquids, verifying the physical rationality and chemical interpretability of the constructed model's prediction results. Temperature was identified as the main controlling variable in the self-diffusion process of ionic liquids. With increasing temperature, the thermal motion of molecules within the system significantly increases, weakening the interionic potential barrier, thereby promoting ion migration and diffusion. The SHAP value analysis results show that there is a significant positive correlation between temperature and diffusion coefficient, indicating that increasing temperature can effectively improve ion kinetic energy and reduce the viscous resistance of the system, thereby accelerating the diffusion process, demonstrating the model's accurate response to thermodynamic characteristics.

[0027] Furthermore, the electrostatic surface area characteristics (ILs_PEOE_VSA2) and solute complexity (Solvent_Kappa3) of ionic liquids were identified as key structural features influencing the interdiffusion behavior of ionic liquid-solute systems. ILs_PEOE_VSA2 reflects the electrostatic potential distribution and polarizable surface area of ​​ionic liquid molecules, exhibiting the highest average SHAP contribution, indicating that the potential distribution of the ionic surface significantly affects the diffusion rate. Increasing the electrostatic surface area reduces the diffusion rate. Solvent_Kappa3, as an index of the topological complexity of solute molecules, characterizes the branching degree and ring structure of the molecular skeleton. Results show a negative correlation between this index and the diffusion coefficient; that is, the more complex the solute structure, the more significant its volume effect and steric hindrance, thereby increasing the viscous resistance of the system and reducing the interdiffusion rate. Simultaneously, the group characteristics of ionic liquids also show a significant structure-dependent influence on diffusion behavior. Ionic liquids containing structural units such as methylene (–CH2–) or alkynyl (–C≡C–) also have high contribution values ​​in the model.

[0028] Comprehensive analysis results show that the machine learning model established in this invention not only achieves a high level of prediction accuracy (R²>0.99), but also reveals the physicochemical essence of ionic liquid diffusion behavior at the feature level. The temperature-dominated thermal effect, the shielding effect of electrostatic surface area, the steric hindrance effect caused by molecular structural complexity, and specific functional groups jointly determine the kinetic differences between self-diffusion and inter-diffusion in ionic liquid systems. The analytical results of this invention provide important theoretical basis and technical support for the rational molecular design of ionic liquid systems and the optimization of their diffusion performance.

Claims

1. A method for predicting the diffusion coefficient of ionic liquids based on interpretable machine learning, characterized in that, Includes the following steps: Data collection and construction: Experimental data on the diffusion coefficients of ionic liquids were collected from multiple data sources and divided into three datasets: cation self-diffusion coefficient, anion self-diffusion coefficient, and ionic liquid-solute interdiffusion coefficient. Feature extraction and preprocessing: Based on the molecular structure information of ionic liquids and / or solutes, multi-scale molecular descriptors are extracted to form an initial feature matrix; The initial feature matrix is ​​preprocessed and merged with environmental parameters to obtain a high-quality feature input set; Model building and training: Construct at least one machine learning model and use Bayesian optimization to automatically optimize the hyperparameters of the machine learning model, with the optimization objective being to maximize the coefficient of determination R² on the test set, to obtain the best trained model; Model prediction step: Using the trained optimal model, predict the diffusion coefficient of the new ionic liquid system; Model interpretability analysis: The Shapley additivity feature interpretation method is used to calculate the contribution of the features to the prediction results and identify the key features that affect diffusion behavior.

2. The method according to claim 1, characterized in that, The multi-scale molecular descriptor includes: Molecular structure descriptors with global information, calculated from the overall molecular structure, and fragment feature descriptors with atomic-level local information, generated based on local atomic chemical environment information.

3. The method according to claim 1, characterized in that, In the feature extraction and preprocessing steps, for the ionic liquid-solute interdiffusion system, molecular descriptors of the ionic liquid and solute are extracted in parallel through a two-way structure verification mechanism, and combined with temperature, pressure and solute mole fraction to form a synergistic feature matrix.

4. The method according to claim 1, characterized in that, Bayesian optimization is used to optimize the hyperparameters of the model, and a high-precision diffusion coefficient prediction framework is constructed.

5. The method according to claim 1, characterized in that, In the model interpretability analysis step, the identified key factors are used to reveal the intrinsic relationship between diffusion behavior and structural response.

Images