Method for screening ionic liquid for catalyzing CO2 cycloaddition reaction based on machine learning
By constructing a three-dimensional convolutional neural network model based on quantum chemical calculations, catalytic activity features can be directly extracted from the three-dimensional electronic structure of ionic liquids. This solves the problem of relying on artificial descriptors in existing technologies, and enables efficient and accurate screening of ionic liquids, leading to the discovery of novel highly active catalysts.
Patent Information
- Application Number
- CN202511876522.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-03-17
AI Technical Summary
Existing machine learning models rely on predefined descriptors in ionic liquid-catalyzed CO2 cycloaddition reactions, which cannot automatically extract local electron distribution features directly related to catalytic activity, resulting in low prediction accuracy and weak generalization ability.
A three-dimensional spatial electronic property data field based on quantum chemical calculations was constructed, and the local electronic structure features of ionic liquids were automatically extracted through a three-dimensional convolutional neural network model to establish a mapping relationship from three-dimensional electronic structure to catalytic activity.
It improves prediction accuracy and generalization ability, enabling the discovery of novel highly active ionic liquids, achieving high-throughput and efficient virtual screening, and shortening the catalyst development cycle.
Smart Images
Figure CN121687249A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of information and communication technology, which is particularly applicable to specific application areas, and specifically relates to an ionic liquid screening method for catalytic CO2 cycloaddition reactions based on machine learning. Background Technology
[0002] With the deepening of the carbon neutrality strategy, the resource utilization of carbon dioxide (CO2) has become an important research direction in the fields of green chemistry and sustainable energy. Among them, the cycloaddition reaction of CO2 with epoxides is considered one of the key pathways to achieve efficient CO2 conversion due to its high atom economy and high added value of the products. This reaction usually relies on catalysts to activate CO2 or epoxides, and ionic liquids, due to their tunable structure, good thermal stability, and strong CO2 solubility, are widely used as the core component of catalytic systems. In recent years, to accelerate the discovery of high-performance ionic liquids, researchers have attempted to introduce machine learning methods to guide the screening process by constructing activity prediction models, significantly improving research and development efficiency.
[0003] Among these, machine learning-based ionic liquid screening methods focus on rapidly identifying highly active catalysts from a large pool of candidate structures. The basic principle is to convert the molecular structure of the ionic liquid into a numerical descriptor, using this as input features to train a regression or classification model, thereby predicting its catalytic performance in the CO2 cycloaddition reaction. However, existing methods generally rely on artificially predefined physicochemical descriptors (such as hydrophobicity logP, Hansen solubility parameter, molecular weight, etc.). While these descriptors offer some interpretability, they are insufficient to fully characterize the microscopic electronic structure features that influence catalytic activity.
[0004] Existing technologies have significant limitations in the feature construction stage: On the one hand, artificial descriptor engineering relies heavily on the experience of domain experts and is prone to overlooking implicit structural information closely related to reaction mechanisms; On the other hand, the key active sites for ionic liquid-catalyzed CO2 cycloaddition are often closely related to the local electron density, electrostatic potential distribution, or orbital interactions of specific atoms (such as oxygen atoms) in their anions. Such fine features at the quantum chemical level cannot be effectively characterized by traditional descriptors.
[0005] Due to a lack of ability to automatically extract highly sensitive implicit correlation features from first-principles calculations, existing machine learning models often suffer from large prediction biases and weak generalization ability due to the omission of key features, especially when dealing with ionic liquids with novel structures. Therefore, in the intelligent screening of ionic liquids for CO2 cycloaddition reactions, there is an urgent need for a machine learning method that can eliminate dependence on manual descriptors and adaptively extract decisive microscopic features directly from quantum chemical calculation results. Summary of the Invention
[0006] The purpose of this invention is to provide an ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning. This method aims to solve the problem that existing machine learning models rely on manually predefined descriptors and cannot automatically extract local electron distribution features directly related to catalytic activity from the three-dimensional spatial structure of ionic liquids, thus leading to the omission of key features and low prediction accuracy.
[0007] To achieve the above objectives, this invention provides a machine learning-based method for screening ionic liquids for catalytic CO2 cycloaddition reactions, comprising the following steps: First, a three-dimensional structure database of ionic liquids is constructed. The construction steps include: providing an initial set containing chemical structure identifiers of various cations and anions; traversing all combinations of cations and anions in the initial set to generate a series of ionic liquid pairs; for each ionic liquid pair, optimizing the initial geometric conformation using the molecular force field method to obtain its three-dimensional spatial coordinate structure data, and storing all obtained three-dimensional spatial coordinate structure data sets to constitute the three-dimensional structure database of ionic liquids.
[0008] Next, quantum chemical calculations are performed on each ionic liquid three-dimensional structure in the ionic liquid three-dimensional structure database to generate its corresponding three-dimensional spatial electronic property data field. The quantum chemical calculations are based on density functional theory. The three-dimensional spatial electronic property data field includes a three-dimensional electrostatic potential data field and a three-dimensional electron density data field. Both the three-dimensional electrostatic potential data field and the three-dimensional electron density data field are stored in the form of cubic grid files, which define the scalar values of the electrostatic potential and electron density at each grid point on a predefined three-dimensional grid around the molecule.
[0009] Then, the three-dimensional spatial electronic property data field corresponding to each ionic liquid is voxelized to generate a standardized multi-channel three-dimensional voxel tensor. The voxelization process specifically includes: setting a fixed-size three-dimensional voxel grid, the grid size completely covering the van der Waals volume of the target ionic liquid and including its surrounding interaction space; interpolating and mapping the grid point data of the three-dimensional static potential data field and the three-dimensional electron density data field to the center of each voxel in the three-dimensional voxel grid, forming electrostatic potential channels and electron density channels; and normalizing the values in the electrostatic potential channels and electron density channels, linearly scaling all values to a preset numerical range, thereby obtaining the standardized multi-channel three-dimensional voxel tensor.
[0010] Subsequently, a three-dimensional convolutional neural network regression model is constructed and trained. This model establishes a nonlinear mapping relationship from the standardized multi-channel three-dimensional voxel tensor to the catalytic activity of the ionic liquid in the carbon dioxide cycloaddition reaction. The three-dimensional convolutional neural network regression model includes a spatial feature autoencoding module and a catalytic activity prediction module. The spatial feature autoencoding module receives the multi-channel three-dimensional voxel tensor as input and extracts the local electronic structure features of the ionic liquid in three-dimensional space layer by layer through the stacking of multiple three-dimensional convolutional layers, activation function layers, and three-dimensional pooling layers, compressing these features into low-dimensional feature vectors. The catalytic activity prediction module receives the low-dimensional feature vectors as input and performs nonlinear combination and transformation on the features through at least one fully connected layer, finally outputting a scalar value. This scalar value predicts the catalytic activity of the corresponding ionic liquid, which is quantified by the reaction energy barrier or conversion frequency obtained in advance through high-precision quantum chemical calculations or experimental measurements. The training process of the model uses a backpropagation algorithm and a gradient descent optimizer to iteratively adjust the weight parameters in the three-dimensional convolutional layers and fully connected layers, aiming to minimize the mean squared error loss function between the model-predicted catalytic activity value and the actual catalytic activity value.
[0011] Finally, the trained 3D convolutional neural network regression model is used to perform high-throughput virtual screening of the candidate ionic liquid library. The high-throughput virtual screening steps include: for each candidate ionic liquid in the library, repeatedly performing the steps of constructing a 3D structure, performing quantum chemical calculations to generate a 3D spatial electronic property data field, and performing voxelization to generate a standardized multi-channel 3D voxel tensor; inputting each generated multi-channel 3D voxel tensor into the trained 3D convolutional neural network regression model to obtain its predicted catalytic activity value; sorting all candidate ionic liquids according to their predicted catalytic activity values, and selecting ionic liquids with a preset catalytic activity threshold or higher as high-activity candidates.
[0012] In one embodiment of the present invention, the molecular force field method is a universal force field or a MMFF 94 force field. The initial geometric conformation optimization employs the conjugate gradient method or the steepest descent method until the system energy gradient is less than a preset convergence threshold.
[0013] In one embodiment of the present invention, the density functional theory-based calculation method employs a B3LYP hybrid functional and a 6-31G(d,p) basis set. The grid spacing of the cubic grid file is set to 0.2 angstroms.
[0014] In one embodiment of the present invention, the size of the three-dimensional voxel mesh is 64×64×64 voxels. The normalization process scales the values to the range of -1 to +1.
[0015] In one embodiment of the present invention, the spatial feature autoencoding module consists of three consecutive three-dimensional convolutional-activation-pooling units. The first unit contains a three-dimensional convolutional layer with 32 filters, a kernel size of 3×3×3, and a stride of 1; the second unit contains a three-dimensional convolutional layer with 64 filters, a kernel size of 3×3×3, and a stride of 1; the third unit contains a three-dimensional convolutional layer with 128 filters, a kernel size of 3×3×3, and a stride of 1. Each of the three-dimensional convolutional layers is followed by a linear rectified unit as an activation function layer. Each activation function layer is followed by a three-dimensional max-pooling layer with a pooling window size of 2×2×2.
[0016] In one embodiment of the present invention, the catalytic activity prediction module consists of two fully connected layers and an output layer. The first fully connected layer receives the feature vector output by the spatial feature autoencoding module and flattens it, mapping it to 256 neurons; the second fully connected layer maps the output of the first fully connected layer to 128 neurons. Each fully connected layer is followed by a linear rectifier activation function and a loss layer with a loss rate of 0.5. The output layer consists of linear neurons without activation functions, directly outputting the predicted scalar value of catalytic activity.
[0017] In one embodiment of the present invention, the model is trained using an adaptive moment estimation optimizer. The initial learning rate during the training process is set to 1×10⁻⁶. -4 The batch size is set to 32.
[0018] This invention also provides an ionic liquid screening system for catalytic CO2 cycloaddition reactions based on machine learning, the system comprising: The ionic liquid structure database construction module is used to generate a series of three-dimensional geometric conformations of ionic liquid molecules according to the combination rules of cations and anions, and store them as a structure database. The quantum chemistry calculation module is connected to the ionic liquid structure database construction module and is used to perform density functional theory-based calculations on the three-dimensional geometric conformation of each ionic liquid in the database to generate three-dimensional electrostatic potential and electron density data fields characterizing its spatial electronic properties. A data voxelization preprocessing module, connected to the quantum chemical calculation module, is used to convert the three-dimensional data field into a multi-channel three-dimensional voxel tensor with a uniform size and numerical range, wherein different channels correspond to electrostatic potential and electron density, respectively. A three-dimensional convolutional neural network regression model is connected to the data voxelization preprocessing module. The model contains an automatic spatial feature encoding unit and a catalytic activity prediction unit. The automatic spatial feature encoding unit is used to automatically learn and extract spatial electronic structure features related to catalytic activity from the three-dimensional voxel tensor to generate a compact feature representation vector. The catalytic activity prediction unit is used to regress and predict the catalytic activity value of the ionic liquid based on the feature representation vector. The high-throughput screening execution module, connected to the three-dimensional convolutional neural network regression model, is used to receive a library containing a large number of candidate ionic liquids, call the aforementioned modules to process and predict the activity of each candidate ionic liquid, and sort and screen the candidate library according to the prediction results, and finally output a sorted list of highly active ionic liquids.
[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. This invention introduces a three-dimensional spatial electronic property data field generated by quantum chemical calculations and voxels it into a multi-channel three-dimensional tensor. It directly uses the basic physicochemical information of ionic liquids as input to a machine learning model, bypassing the dependence on artificially designed molecular descriptors.
[0020] 2. This invention uses a three-dimensional convolutional neural network as the core model, which can automatically and end-to-end learn and extract multi-scale, high-order spatial electronic features related to catalytic activity that are difficult for humans to define intuitively, especially the local charge enrichment and hole distribution features in the interaction region between anions and cations, thereby avoiding the omission of key features due to improper manual selection of descriptors.
[0021] 3. The model constructed in this invention directly establishes the mapping relationship from the three-dimensional electronic structure of ionic liquids to catalytic activity. This modeling method based on physical essence significantly improves the prediction accuracy and generalization ability of the model, and can discover novel highly active ionic liquid structures that transcend existing chemical knowledge and intuition.
[0022] 4. This invention systematizes and automates the entire screening process, forming a complete data and computational flow from initial structure generation to final output of highly active candidates. This enables high-throughput and high-efficiency virtual screening of large-scale ionic liquid chemical spaces, greatly shortening the development cycle of novel catalysts. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the overall technical architecture of the ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the three-dimensional convolutional neural network regression model in this invention; Figure 3 This is a flowchart illustrating the logical process of constructing a three-dimensional structure database of ionic liquids and generating quantum chemical electronic property data fields in this invention. Figure 4 This is a flowchart illustrating the logical flow of the voxelization processing of three-dimensional spatial electronic property data fields and the standardization of multi-channel three-dimensional voxel tensors in this invention. Figure 5 This is a schematic diagram of the internal structure and information flow of the spatial feature automatic encoding module and the catalytic activity prediction module in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow of candidate ionic liquid processing and sorting output during the high-throughput virtual screening process in this invention; Detailed Implementation
[0024] Please refer to Figures 1 to 6 This invention provides a machine learning-based method for screening ionic liquids in the catalytic carbon dioxide cycloaddition reaction. The core of this method lies in constructing an end-to-end data-driven workflow. Starting from the three-dimensional geometric conformation of the ionic liquid, this workflow generates a three-dimensional data field characterizing its spatial electronic properties through high-precision quantum chemical calculations. This data field is then converted into a standardized multi-channel three-dimensional voxel tensor as input. A three-dimensional convolutional neural network regression model is used to automatically extract local electronic structure features directly related to catalytic activity, ultimately achieving high-throughput virtual screening of a large-scale candidate ionic liquid library. The specific implementation of this method will be described in detail below according to the S-step sequence clearly specified in the invention description.
[0025] The method first performs step S1: constructing a three-dimensional structure database of ionic liquids. This step begins with an initial set containing various cation and anion chemical structure identifiers. The cations in the initial set include, but are not limited to, imidazolium, pyridinium, quaternary ammonium, and quaternary phosphorus structures; the anions include, but are not limited to, halide ions, bis(trifluoromethanesulfonyl)imide, tetrafluoroborate, hexafluorophosphate, bicarbonate, and carboxylate ions. A full combinatorial traversal of all cations and anions in the initial set is performed to generate a series of ionic liquid pairs.
[0026] For each ionic liquid pair, initial geometric conformation optimization was performed using a molecular force field method. The molecular force field method was either a universal force field or the MMFF94 force field. The optimization process employed the conjugate gradient method or the steepest descent method, iteratively calculating until the total energy gradient of the system was less than a preset convergence threshold, which was set to 0.001 eV per angstrom. After optimization, atomic coordinate data for each ionic liquid in three-dimensional space were obtained, including atom type, atomic number, three-dimensional Cartesian coordinates, and charge distribution information. All optimized three-dimensional structural data were stored in a standard molecular file format, forming an ionic liquid three-dimensional structure database. Each record in this database uniquely corresponds to the chemical structure of the ionic liquid and contains its complete geometric conformation information, providing accurate initial input for subsequent quantum chemical calculations.
[0027] Then, step S2 is performed: for each ionic liquid three-dimensional structure in the ionic liquid three-dimensional structure database, quantum chemical calculations are performed to generate its corresponding three-dimensional spatial electronic property data field. The quantum chemical calculations are based on density functional theory, using B3LYP mixed functionals and 6-31G(d,p) basis sets for self-consistent field iterative solutions. During the calculation, firstly, single-point energy calculations are performed on the molecular system to obtain its ground-state electronic wavefunction; then, based on this wavefunction, the electrostatic potential and electron density distribution in the space surrounding the molecule are calculated.
[0028] The three-dimensional electrostatic potential data field is defined as the scalar value of the electrostatic potential at each point on a cubic grid divided by a uniform mesh within a spatial region 5 Å epitaxially extending from the van der Waals surface of the molecule; the three-dimensional electron density data field is defined as the scalar value of the electron density at each point within the same spatial region. The grid spacing of the cubic grid file is set to 0.2 Å to ensure high-resolution capture of local electronic structure details. After calculation, both data fields are output in cubic file format, each file containing the coordinates of the grid origin, the grid dimensions (number of grid points in the X, Y, and Z directions), and the scalar values at all grid points arranged in row priority. This step ensures that each ionic liquid is endowed with spatially distributed electronic properties that strictly correspond to its three-dimensional geometric conformation and have clearly defined physical meaning.
[0029] Next, step S3 is performed: the three-dimensional spatial electronic property data field corresponding to each ionic liquid is voxelized to generate a standardized multi-channel three-dimensional voxel tensor. This voxelization process first establishes a fixed-size three-dimensional voxel grid, with dimensions of 64×64×64 voxels. This size is sufficient to completely encompass the van der Waals volume of any ionic liquid and include its surrounding five angstroms of interaction space. For each ionic liquid, with its geometric center as the origin, the cubic grid data generated by the aforementioned quantum chemical calculations is mapped to the center position of each voxel in the fixed voxel grid using a trilinear interpolation algorithm.
[0030] After mapping, two independent three-dimensional arrays are formed, corresponding to the electrostatic potential channel and the electron density channel, respectively. Then, all values in these two channels are normalized. Normalization uses linear scaling, mapping the minimum value to -1, the maximum value to +1, and other values proportionally to this range. The two normalized three-dimensional arrays are stacked along the channel dimensions to form a four-dimensional tensor of size 64×64×64×2, i.e., a standardized multi-channel three-dimensional voxel tensor. This tensor serves as the unified input format for subsequent machine learning models, ensuring comparability of different ionic liquids in terms of spatial scale and numerical range.
[0031] The following step, S4, involves constructing and training a three-dimensional convolutional neural network regression model. This model establishes a nonlinear mapping from the standardized multi-channel three-dimensional voxel tensor to the ionic liquid-catalyzed carbon dioxide cycloaddition reaction activity. The model consists of a cascaded spatial feature autoencoding module and a catalytic activity prediction module.
[0032] The spatial feature autoencoding module receives an input tensor of size 64×64×64×2 and passes it sequentially through three 3D convolution-activation-pooling units. The first unit contains a 3D convolutional layer with a 3×3×3 kernel, 32 filters, and a stride of 1. The convolutional output is then fed into a linear rectifier unit as the activation function. The activated output is then processed by a 3D max-pooling layer with a 2×2×2 pooling window and a stride of 2, reducing the output size to 32×32×32×32. The second unit's 3D convolutional layer also uses a 3×3×3 kernel, with 64 filters and a stride of 1. It is followed by a linear rectifier unit and a 3D max-pooling layer, resulting in an output size of 16×16×16×64. The third unit's 3D convolutional layer has 128 filters, with the other parameters remaining the same, resulting in an output size of 8×8×8×128 after pooling. This module ultimately compresses the high-dimensional spatial data into a low-dimensional feature representation. Subsequently, the feature tensor was flattened into a one-dimensional vector with a length of 8×8×8×128, or 65536 dimensions.
[0033] The catalytic activity prediction module receives the flattened one-dimensional feature vector and first maps it to 256 neurons through a first fully connected layer. The output of this layer is then connected to a linear rectifier activation function and a loss layer with a 0.5 loss rate to prevent overfitting. Next, the data enters a second fully connected layer, mapped to 128 neurons, and is similarly connected to a linear rectifier and a loss layer with a 0.5 loss rate. Finally, the data passes through a linear output layer without an activation function; this layer contains only neurons and directly outputs a scalar value, which is the catalytic activity predicted by the model. The true label of the catalytic activity is quantified by the reaction energy barrier or conversion frequency obtained in advance through high-precision quantum chemical calculations or experimental measurements, with units uniformly set to kilojoules per mole or conversions per hour, and standardized to match the model output range.
[0034] The model training process uses an adaptive moment estimation optimizer, with an initial learning rate set to 1×10⁻⁶. -4 The batch size is 32. The loss function uses mean squared error, defined as: ; in, This represents the batch sample size. For the first The true catalytic activity value of each sample These are the predicted values for the model. During training, the gradient of the loss function with respect to all trainable parameters is calculated using the backpropagation algorithm, and the weights are updated using the optimizer. Training continues until the loss function on the validation set no longer decreases for 10 consecutive epochs, or until the preset maximum training epochs of 500 are reached. After training is complete, the model parameters are frozen for subsequent selection tasks.
[0035] Finally, step S5 is executed: using the trained 3D convolutional neural network regression model, a high-throughput virtual screening of the candidate ionic liquid library is performed. The candidate library contains thousands to tens of thousands of novel ionic liquid structures that have not been experimentally verified. For each candidate ionic liquid in the library, the system sequentially calls the sub-processes described in steps S1 to S3: first, its 3D geometric conformation is generated; then, quantum chemical calculations are performed to obtain a 3D electronic property data field; and finally, voxelization is performed to generate a normalized multi-channel 3D voxel tensor. This tensor is input into the trained 3D convolutional neural network regression model, which outputs the predicted catalytic activity value. The predicted values of all candidate ionic liquids are collected and sorted from highest to lowest value. A preset catalytic activity threshold is set, determined based on the activity distribution of historically highly active ionic liquids, for example, the lower limit of activity of the top 10% of samples in the training set. All ionic liquids with predicted activity values higher than this threshold are selected, forming a list of highly active candidates. This list is sorted in descending order of predicted activity and serves as the final output for experimental chemists to perform subsequent synthesis and verification.
[0036] Throughout the methodology, data flow exhibits strict temporal dependencies. The output of step S1 is the sole input source for step S2; the cube file generated in step S2 forms the basis for voxelization in step S3; the normalized tensor generated in step S3 serves as the common input format for model training in step S4 and selection in step S5. Each step also contains mandatory sub-operation sequences. For example, in the quantum chemical calculations of step S2, single-point energy convergence of the geometric conformation must be achieved before grid-point calculations of electrostatic potential and electron density can begin; in voxelization of step S3, spatial alignment and interpolation must be completed before channel normalization can be performed. An exception handling mechanism is embedded in critical computational nodes: if the quantum chemical calculation is interrupted due to convergence failure, the system will automatically switch to an alternative functional or increase the integral grid density before retrying; if data overflow or invalid values are detected during voxelization, the system will mark the sample and skip it, ensuring the overall process remains uninterrupted.
[0037] This method completely avoids the subjectivity and limitations of manual descriptor design by directly encoding the physical nature of ionic liquids—their three-dimensional electronic structure—as input to the machine learning model. The hierarchical structure of the three-dimensional convolutional neural network enables it to capture multi-scale features, from local charge distribution at the atomic scale to the overall polarization effect of ions, layer by layer. The first convolutional layer primarily responds to regions of strong electrostatic potential gradients around individual atoms, such as the lone pair electron-rich region of oxygen atoms in anions; the second layer begins to integrate the electron density correlation between adjacent atoms, identifying regions of weak hydrogen bond interactions; the third layer further abstracts the charge transfer channels or hole structures formed between cations and anions. These higher-order features are often highly correlated with the nucleophilic attack sites of CO2 molecules. By learning from a large amount of known activity data, the model automatically establishes a quantitative relationship between these implicit spatial patterns and catalytic barriers.
[0038] At the system level, this invention also provides a complete screening system to support the automated execution of the above methods. This system includes an ionic liquid structure database construction module, which integrates a chemical structure analysis engine and a molecular force field optimizer, capable of batch processing SMILES strings or MOL files and outputting standardized three-dimensional coordinate files. A quantum chemistry calculation module encapsulates the core interface of density functional theory calculation software, supporting task queue management and resource scheduling to ensure the stability of large-scale computations. A data voxelization preprocessing module incorporates a spatial alignment algorithm and interpolation engine, capable of efficiently processing cubic data of different sizes and outputting tensors in a unified format. A three-dimensional convolutional neural network regression model is implemented using a deep learning framework, supporting GPU-accelerated inference. A high-throughput screening execution module coordinates the aforementioned modules, realizing an end-to-end pipeline operation from candidate library input to high-activity list output. The modules communicate through standardized data interfaces to ensure data consistency and process traceability.
[0039] In summary, the method and system described in this embodiment, through the deep integration of quantum chemical computation and three-dimensional deep learning technology, construct a framework for predicting and screening the catalytic activity of ionic liquids that does not require artificial feature engineering and directly starts from the physical essence. This framework not only significantly improves prediction accuracy but also has the ability to discover novel catalysts that transcend traditional chemical intuition, providing an efficient and reliable computational tool for catalyst design under the goals of green chemistry and carbon neutrality.
[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for screening ionic liquids for catalyzing CO2 cycloaddition reactions based on machine learning, characterized in that, The method comprises the following steps: constructing an ionic liquid three-dimensional structure database; performing density functional theory-based quantum chemistry calculation on each ionic liquid three-dimensional structure in the ionic liquid three-dimensional structure database to generate its corresponding three-dimensional spatial electronic property data field; and voxelizing each ionic liquid's corresponding three-dimensional spatial electronic property data field to generate a standardized multi-channel three-dimensional voxel tensor; constructing and training a three-dimensional convolutional neural network regression model; using the trained three-dimensional convolutional neural network regression model to perform high-throughput virtual screening on a candidate library of ionic liquids to be screened.
2. The ionic liquid screening method for catalytic CO2 cycloaddition reactions based on machine learning according to claim 1, characterized in that, The step of constructing the ionic liquid three-dimensional structure database comprises providing an initial set containing chemical structure identifiers of various cations and anions; all combinations of cations and anions in the initial set are traversed to generate a series of ionic liquid pairs, and initial geometric conformation optimization is performed on each ionic liquid pair using a molecular force field method to obtain its three-dimensional spatial coordinate structure data, and all obtained three-dimensional spatial coordinate structure data are stored to constitute the ionic liquid three-dimensional structure database.
3. The ionic liquid screening method for catalytic CO2 cycloaddition reactions based on machine learning according to claim 1, characterized in that, The three-dimensional spatial electronic property data field comprises a three-dimensional electrostatic potential data field and a three-dimensional electron density data field, and both are stored in the form of a cubic grid file, in which the electrostatic potential scalar value and the electron density scalar value of each grid point on a preset three-dimensional grid around the molecule are defined.
4. The ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning according to claim 1, characterized in that, The voxelization process comprises setting a three-dimensional voxel grid of fixed size, interpolating and mapping the grid point data of the three-dimensional electrostatic potential data field and the three-dimensional electron density data field to the center of each voxel of the three-dimensional voxel grid to form an electrostatic potential channel and an electron density channel, and normalizing the values in the electrostatic potential channel and the electron density channel to linearly scale all values within a preset value interval.
5. The ionic liquid screening method for catalytic CO2 cycloaddition reactions based on machine learning according to claim 1, characterized in that, The three-dimensional convolutional neural network regression model comprises a spatial feature automatic encoding module and a catalytic activity prediction module; The spatial feature automatic encoding module receives the multi-channel three-dimensional voxel tensor as input and extracts local electronic structure features of the ionic liquid in three-dimensional space layer by layer through the stacking of multiple three-dimensional convolutional layers, activation function layers and three-dimensional pooling layers to compress into a low-dimensional feature vector; The catalytic activity prediction module receives the low-dimensional feature vector as input and performs nonlinear combination and transformation of the features through at least one fully connected layer to output a predicted catalytic activity scalar value; The catalytic activity is quantified by the reaction energy barrier or conversion frequency determined in advance through high-precision quantum chemistry calculation or experimental measurement; The training process of the three-dimensional convolutional neural network regression model iteratively adjusts the weight parameters to minimize the mean square error loss function between the model prediction value and the true value by using the back propagation algorithm and the gradient descent optimizer.
6. The ionic liquid screening method for catalytic CO2 cycloaddition reactions based on machine learning according to claim 1, characterized in that, The high-throughput virtual screening comprises repeatedly performing the steps of constructing a three-dimensional structure, performing quantum chemistry calculation to generate a three-dimensional spatial electronic property data field, and voxelizing to generate a standardized multi-channel three-dimensional voxel tensor for each candidate ionic liquid in the candidate library. Each of the generated multi-channel three-dimensional voxel tensors is input into the trained three-dimensional convolutional neural network regression model to obtain a predicted catalytic activity value thereof, and all candidate ionic liquids are ranked according to the predicted catalytic activity values to screen out ionic liquids with a catalytic activity value above a preset threshold as high-activity candidates.
7. The ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning according to claim 1, characterized in that, The quantum chemical calculation based on the density functional theory adopts a B3LYP hybrid functional and a 6-31G(d, p) basis set.
8. The ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning according to claim 5, characterized in that, The training of the three-dimensional convolutional neural network regression model adopts an adaptive moment estimation optimizer.
9. The ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning according to claim 2, characterized in that, The cations include imidazolium, pyridinium, quaternary ammonium or quaternary phosphonium structures; and the anions include halide, bis-trifluoromethylsulfonylimide, tetrafluoroborate, hexafluorophosphate, bicarbonate or carboxylate.
10. The ionic liquid screening method for catalytic CO2 cycloaddition reaction based on machine learning according to claim 6, characterized in that, In the voxelization process, a trilinear interpolation algorithm is used to map cubic grid data to the center of a fixed voxel grid.