Material performance prediction method and system based on deep learning

Abstract

The invention discloses a material performance prediction method and system based on deep learning, and relates to the technical field of material performance prediction, and the method comprises the steps: building a cross-level message passing network based on a double-layer graph data structure, extracting atomic-level and functional group-level feature embedding, and generating a molecular global latent variable through bidirectional cross-scale attention interaction; fusing molecular global latent variables with atomic-scale and functional group-scale features, applying physical constraints in combination with a chemical bonding rule base and a group contribution theory, and constructing a physical property prediction model; obtaining candidate molecules through autoregression generation and molecular force field verification based on a molecular global latent variable and a physical property prediction model; and screening candidate molecules by using the physical property prediction model, and outputting a final molecule set through molecular dynamics simulation and synthesis feasibility evaluation verification. Through multi-stage verification of molecular dynamics simulation and synthesis feasibility evaluation, a high-tension ring or a non-synthesized structure is effectively eliminated, so that the generated molecule has high performance and manufacturability.

Description

Technical Field

[0001] The present invention relates to the technical field of material property prediction, and in particular to a material property prediction method and system based on deep learning. Background Art

[0002] In recent years, material property prediction technology has gradually shifted from traditional methods based on empirical formulas to data-driven deep learning models. Existing mainstream methods primarily build prediction frameworks based on molecular graph neural networks (GNNs) or simplified molecular descriptors (such as molecular fingerprints), achieving physical property regression through end-to-end training. However, these methods still have significant limitations in terms of comprehensive molecular representation and compatibility with physicochemical rules. Differences in SMILES atom ordering can generate different graph structures for the same molecule, leading to GNN misclassification as isomers. The lack of standardized atom reordering algorithms forces the model to learn redundant topological transformation patterns, reducing generalization. Chiral centers (R / S labels) and double bond cis / trans configurations are not explicitly encoded, relying solely on the adjacency matrix to convey topological information. For example, the model cannot distinguish between cis / trans butenedioic acid, resulting in significant deviations between predicted values ​​and actual physical properties. Existing methods directly use atomic-level graph structures without subgraph segmentation based on a chemical group dictionary. For example, decomposing benzene rings into isolated carbon atom nodes loses the overall electronic effects of the conjugated system and weakens the model's ability to characterize aromaticity. Molecular fingerprints use hash functions to map continuous chemical properties into discrete binary bits, losing geometric information such as bond lengths and bond angles. For example, it is impossible to distinguish between a conjugated system with alternating single and double bonds and an isolated double bond, resulting in a blurred electron cloud distribution. Fingerprint design does not incorporate prior knowledge from the group contribution method. For example, the hydrogen bond contribution of hydroxyl groups and the acidic strength of carboxylic acids are not differentially encoded, making it difficult for the model to learn non-additive effects between groups. Bond lengths are freely predicted beyond a reasonable range, and the output distribution is not constrained by physical penalty terms. For example, when predicting hypervalent nitrogen compounds, a pentacoordinate structure is generated, violating valence bond theory. When predicting physical properties (such as glass transition temperature), thermodynamic consistency (such as the positive correlation between Tg and melting point) is not enforced, resulting in model outputs that violate phase transition laws. Summary of the Invention

[0003] In view of the above existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides a material property prediction method based on deep learning to solve the problems of prediction deviation and irrational generation of molecules caused by incomplete molecular characterization and lack of physical and chemical constraints in existing material property prediction models.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a material property prediction method based on deep learning, which comprises: Parse the molecular SMILES string into an atomic-level molecular topology graph and generate a double-layer graph structure data; Based on the two-layer graph structure data, a cross-level message passing network is constructed to extract atomic-level features and functional group-level features, and a molecular global latent variable is generated through bidirectional cross-scale attention interaction; The global latent variables of the molecule are integrated with the atomic-level and functional group-level features, and physical constraints are imposed by combining the chemical bonding rule library and group contribution theory to construct a physical property prediction model; Based on the molecular global latent variables and physical property prediction model, candidate molecules are obtained through autoregressive generation and molecular force field verification; The physical property prediction model is used to screen candidate molecules, and the final molecular set is output after verification through molecular dynamics simulation and synthetic feasibility assessment.

[0006] As a preferred solution of the material property prediction method based on deep learning of the present invention, wherein: the generating of the double-layer graph structure data comprises the following steps: Using chemical informatics tools, the SMILES string is verified for atomic valence validity, bond connectivity rationality, and automatically corrected to generate a temporary SMILES sequence with syntax correction. Perform stereochemical analysis on the temporary SMILES sequence to generate a standardized SMILES sequence; Perform chemical equivalence normalization on standardized SMILES sequences, extract atomic-level molecular topology and annotate dimensional properties of atoms and chemical bonds; Based on the functional group dictionary, atomic-level molecular topology subgraph isomorphism matching and conflict optimization are performed to divide the conflict-free functional group structure and generate double-layer graph structure data.

[0007] As a preferred solution of the material property prediction method based on deep learning of the present invention, wherein: the construction of the cross-level message passing network includes the following steps: Perform multi-layer graph convolution on the atomic-level molecular graph to aggregate neighborhood features to generate atomic-level features; Based on the functional group partitioning scheme, the atomic-level features are spatially averaged and pooled, and the functional group type encoding is combined to generate functional group-level features; A bidirectional mapping between atoms and functional groups is established through cross-level pointers, a two-layer message passing network is constructed, and molecular global latent variables are generated through bidirectional cross-scale attention interaction.

[0008] As a preferred embodiment of the material property prediction method based on deep learning of the present invention, the construction of the physical property prediction model includes the following steps: Decode the atomic-level embedding vector into bond length and bond angle parameters, and combine it with the chemical bonding rule library to impose penalties on the atomic weights that violate physical constraints to generate a revised atomic feature matrix; The functional group-level features are spliced ​​with the theoretical weights of group contributions to calculate the asymmetric attention weights; By integrating atomic-level global pooling vectors, functional group-level global pooling vectors, and molecular global latent variables, a multi-objective composite loss function is constructed and the parameters of the physical property prediction model are dynamically optimized to obtain a trained physical property prediction model.

[0009] As a preferred solution of the material property prediction method based on deep learning described in the present invention, the generation through autoregression and molecular force field verification refers to generating functional group sequences based on global latent variable optimization, splicing atomic-level molecular sketches based on functional group templates, generating three-dimensional structures through molecular force field energy minimization, and verifying valence legitimacy, ring tension and chirality consistency to obtain candidate molecules.

[0010] As a preferred embodiment of the material property prediction method based on deep learning described in the present invention, the molecular dynamics simulation refers to performing energy minimization and equilibrium and production stage simulation on candidate molecules, calculating key physical properties and comparing them with predicted physical property values, and eliminating molecules with excessive physical property deviations.

[0011] As a preferred embodiment of the material property prediction method based on deep learning described in the present invention, the synthetic feasibility assessment verification refers to calculating the synthetic feasibility score, environmental complexity and functional group compatibility, and reversely predicting the feasible synthetic path, eliminating molecules that cannot be synthesized and that do not meet the process standards, and obtaining a set of verified molecules.

[0012] In a second aspect, the present invention provides a material property prediction system based on deep learning, comprising: The structure analysis module parses the molecular SMILES string into an atomic-level molecular topology graph and generates a double-layer graph structure data; A cross-layer encoding module constructs a cross-layer message passing network based on the two-layer graph structure data, extracts atomic-level features and functional group-level features, and generates molecular global latent variables through bidirectional cross-scale attention interaction; The physical property prediction module integrates the molecular global latent variables with atomic-level and functional group-level features, combines the chemical bonding rule library and group contribution theory to impose physical constraints, and constructs a physical property prediction model; A molecule generation module, based on the molecular global latent variables and physical property prediction model, obtains candidate molecules through autoregressive generation and molecular force field verification; The screening and evaluation module uses the physical property prediction model to screen candidate molecules, and outputs the final molecular set through molecular dynamics simulation and synthetic feasibility evaluation verification.

[0013] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the material property prediction method based on deep learning as described in the first aspect of the present invention is implemented.

[0014] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the material property prediction method based on deep learning as described in the first aspect of the present invention.

[0015] The beneficial effects of the present invention are as follows: through stereochemical standardization and chemical equivalence normalization, the semantic ambiguity of SMILES sequences is eliminated, and the problem of decreased model generalization caused by inconsistent molecular representation is solved; through VF2 subgraph isomorphism matching and weighted conflict resolution strategy, the precise division of functional group levels is achieved, and the intra-group synergistic effects and cross-group interactions are effectively encoded, so that the model can simultaneously capture atomic-level details and group-level chemical laws; the introduction of physical constraint penalty terms and cross-scale attention mechanisms forces model predictions to be consistent with chemical intuition, the atomic weight correction module suppresses abnormal activation of hypervalent bond regions, and the group contribution theory alignment module strengthens asymmetric associations between functional groups, ensuring the physical rationality of the prediction results from both data-driven and theory-driven dual paths; through multi-stage verification of molecular dynamics simulation and synthetic feasibility assessment, high-strain rings or non-synthesizable structures are effectively eliminated, so that the generated molecules have both high performance and preparability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 This is a flow chart of the material property prediction method based on deep learning in Example 1.

[0018] Figure 2 This is a flowchart of cross-layer coding in Example 1.

[0019] Figure 3 This is a flow chart for the physical property prediction in Example 1.

[0020] Figure 4 Flowchart for the generation and validation of molecules in Example 1. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0022] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0024] Example 1, with reference to Figures 1 to 4 This embodiment provides a material property prediction method based on deep learning, comprising the following steps: S1. Use the SMILES string as the initial data and perform a molecular integrity check through the chemical informatics toolkit. The molecular integrity check includes checking the legality of atomic valence states (such as tetravalent carbon atoms and divalent oxygen atoms), the rationality of bond connections (such as not allowing hypervalent bonds and ring closure errors), and automatically correcting invalid connections (for example, converting pentavalent nitrogen to a charged form and repairing unclosed aromatic rings). After the integrity check passes, a temporary SMILES sequence with grammatical correction is obtained. If the correction cannot be made, it is marked as invalid input and the process is terminated.

[0025] Perform stereochemical analysis on the verified temporary SMILES sequence. The specific analysis operations are as follows: Identify the chiral center and the cis-trans configuration of the double bond, calculate the lowest energy conformation using a three-dimensional coordinate generation algorithm (such as the ETKDG method), convert the relative stereo description into an absolute configuration (such as R / S labeling), and generate a standardized SMILES sequence with clear stereochemical information.

[0026] Chemical equivalence normalization is performed on standardized SMILES sequences that specify stereochemical information. This includes unifying bond representations (e.g., converting aromatic bonds to alternating double / single bonds), removing redundant hydrogen atoms (converting explicit hydrogens to implicit), and reordering atom numbers to ensure canonical uniqueness of SMILES strings (avoiding different string representations of the same molecule). This results in a unique standardized SMILES sequence. The standardized SMILES is then input into an RDKit (open source cheminformatics toolkit) molecule object to extract the atomic-level molecular topology, as follows: Each atom is annotated with 12 attributes as node features, including element type (one-hot encoding), formal charge (-1, 0, +1), valence state (current number of bonds), hybridization state (sp³ / sp² / sp), ring affiliation (whether it is inside the ring and the ring size), and chirality (R / S / None). Each chemical bond is annotated with 6 attributes as edge features, including bond type (mono / bi / tri / aromatic), bond length (Å), stereo configuration (cis / trans / unspecified), whether it is conjugated, whether it is inside the ring, and bond order (integer).

[0027] Load a predefined chemical group dictionary containing typical functional groups and perform VF2 subgraph isomorphism matching with the atomic-level molecular topology graph. The specific operations are as follows: A depth-first strategy is used to traverse the atomic-level molecular topology graph, giving priority to matching multi-atomic complex groups (such as benzene rings and acrylate chains). A temporary subgraph segmentation scheme is generated through each matching operation, and the complex group type, atomic members of the atomic-level molecular topology graph, and cross-group connection bonds are recorded to obtain a candidate functional group set segmentation scheme containing group type labels, member atom indices, and connection bond endpoint information.

[0028] Conflict detection and optimization selection are performed on the partitioning schemes of the candidate functional group set. When atoms are matched by multiple group templates at the same time, a weighted scoring strategy is adopted to prioritize the retention of groups with more matching atoms, and the remaining single atoms are classified as independent nodes into the residue category; cross-group connecting bonds are reconstructed into functional group edges based on the endpoint atom ownership relationship to ensure the topological integrity of the atomic-level molecule and generate a conflict-free functional group partitioning scheme.

[0029] A three-layer graph convolutional network is run on the atomic-level molecular topology graph for message passing. Each atomic node aggregates the atom / bond features within a three-hop neighborhood to generate an atomic-level embedding vector that encodes local chemical environment information (such as electron cloud distribution and steric hindrance effect).

[0030] The atomic-level embedding vectors are spatially averaged and pooled according to a conflict-free functional group partitioning scheme to obtain the intra-group atomic environment aggregation features. The intra-group atomic environment aggregation features are concatenated with the one-hot encoding of the functional group type and input into an MLP (multi-layer perceptron) for nonlinear transformation to generate a functional group-level feature vector. Based on the conflict-free functional group partitioning scheme and the functional group-level feature vector, a two-layer molecular graph structure data is constructed as follows: For atomic-level graphs, the original atomic-level molecular topology features are retained, and atomic-level embedding vectors are attached to the nodes; for functional group-level graphs, nodes contain functional group-level feature vectors, and edges inherit cross-group connection bond properties; a bidirectional mapping table of functional group nodes and member atom index lists is established through cross-level pointers to achieve cross-level association and obtain double-layer molecular graph structure data.

[0031] S2. Input the atomic-level molecular topology of the two-layer graph structure data (including node features and edge features) into a three-layer message passing network. Each layer performs the following operations: A message vector (edge ​​type embedding + source node features + linear transformation of target node features) is calculated for each edge. Each atomic node aggregates its adjacent edge messages (gated summation) and updates the node state through a GRU unit (gated recurrent unit). After the message vector and the updated node state are output at each layer, global average pooling is performed to generate a level-specific molecular graph summary vector.

[0032] According to the cross-level bidirectional mapping table, the specific molecular graph summary vectors of each level are grouped according to the functional groups to which they belong. A spatial average pooling operation is performed on each functional group node, and the dimension of all atomic embedding vectors in the group is reduced to the initial embedding vector of the functional group. The vectors are then concatenated with the functional group-level feature vector to form the input features of the functional group node. For each functional group node, the attention coefficient (softmax normalized weight based on feature similarity) of the current functional group node and the adjacent functional group nodes is calculated. The features of the adjacent nodes are weighted and aggregated according to the calculated attention coefficient, and the node state is updated through the LeakyReLU (linear unit with leakage correction) activation function. The updated node features are added to the functional group node input features, retaining the original topological information, to obtain the functional group-level context embedding vector.

[0033] Perform bidirectional cross-scale attention interaction on the atomic-level embedding vector and the functional group-level context embedding vector as follows: Attention weights are calculated based on the atomic-level embedding vector and the functional group-level context embedding vector to generate the atomic query vector and the functional group key vector. The atomic features are weightedly injected into the functional group node, and the functional group embedding is updated to the context-aware feature. Through the attention of the functional group-level context embedding vector and the atomic-level embedding vector, the functional group query vector and the atomic key vector are generated in reverse, and the functional group feature vector is broadcast to the atomic node to enhance the cross-level semantics of the atomic embedding. After multiple iterative cycles of optimization, a two-way embedding vector set containing atoms and functional groups is generated.

[0034] Based on the topological structure of the functional group-level graph, the number of neighbors of each functional group node is calculated, and the core group with the most cross-group chemical bonds (such as benzene ring) is preferentially selected as the root node. When there are multiple candidates, the atomic scales of the groups are further compared, and the context embedding vector of the largest group is selected as the initial input; the selected root node embedding vector is input into the batch normalization layer to eliminate the feature distribution differences between different molecules, and then the numerical range is constrained to the [-1,1] interval through the hyperbolic tangent activation function, and the standardized feature vector is output; the standardized feature is input into the variational inference processing unit and passes through two fully connected paths in parallel: the mean path passes through the middle layer activated by ReLU and finally outputs the linear mean vector; the variance path shares the middle layer of the same structure, but uses softplus activation at the end to generate the logarithmic variance vector.

[0035] S3. Input the atomic-level embedding vector into the gated attention layer, calculate the cross-attention score between atoms, and identify synergistic atomic pairs (such as alternating atoms in a conjugated system); filter low-correlation interactions through learnable gating parameters, retain significant chemical interaction patterns (such as hydrogen bond donor-acceptor pairs), and output an enhanced feature vector with attention weighting; input the enhanced feature vector into a two-layer perceptron: the first layer uses the GeLU activation function to extract nonlinear associations, and the second layer uses linear transformation to generate initial importance weights.

[0036] Furthermore, the generated initial importance weight reflects the predicted contribution of the atom to the target property (such as glass transition temperature), ranging from 0 to 1.

[0037] Based on the chemical bonding rule library (such as the reference range of C=O double bond length 1.20-1.25Å), the atomic-level embedding vector is decoded into predicted bond length and bond angle parameters. If the predicted bond length and bond angle parameters exceed the physical range in the reasonable chemical bonding rule library (such as a C=O bond length is predicted to be 1.30Å), an exponential decay penalty is applied to the initial importance weight to reduce the contribution in the high-confidence conflict area to obtain the corrected importance weight; the corrected importance weight is multiplied element-by-element with the original atomic embedding vector to suppress atomic features that do not conform to chemical intuition (such as abnormal activation values ​​in the hypervalent bonding area) while retaining physically reasonable local environment information, and the filtered atomic feature matrix is ​​output.

[0038] The functional group-level embedding vector (encoding group type and atomic environment characteristics) and the group contribution theory weight are aligned and spliced ​​according to the functional group nodes to form a functional group hybrid feature matrix that integrates theoretical prior knowledge; a learnable linear projection matrix is ​​used to map the functional group hybrid feature matrix into a query vector (capturing the information requirements of the current functional group) and a bond vector (representing the interaction potential of other functional groups), and a learnable cross-group attention mechanism is used to calculate the similarity scores between all functional group hybrid feature matrix nodes; the similarity scores are row-wise softmax normalized to generate an asymmetric attention weight matrix between functional groups, highlighting group combinations with strong synergistic or antagonistic effects (such as hydrogen bond donor-acceptor pairs) and weakening the interference of irrelevant groups; the hybrid features are weighted aggregated according to the attention weight matrix, and feature dimensionality reduction is performed through a linear projection network. Cross-correlation analysis is performed, the cosine similarity between functional group nodes is calculated, and potential synergistic group pairs are identified to generate functional group enhanced features that retain the driving characteristics and conform to the group contribution theory, and are associated with the screened atomic-level features according to a cross-level mapping table.

[0039] It is further explained that the deep correlation pattern refers to the group synergy effect, that is, the nonlinear superposition between non-homologous groups.

[0040] The screened atomic-level features are used as deep semantic queries (focusing on local chemical key points), and the functional group enhanced features are used as attribute keys (providing group-level context); according to the cross-level correlation between the screened atomic-level features and the functional group enhanced features, they are split into multiple independent attention heads. Each attention head learns different chemical interaction modes (such as van der Waals interactions and hydrogen bond formation tendencies) and captures the asymmetric relationship between atoms and functional groups in parallel in multidimensional space. Each attention head independently outputs weighted functional group feature fragments, which are concatenated according to the attention heads and then restored through linear transformation to generate a cross-scale enhanced atom-functional group association matrix; the cross-scale atom-functional group association features are activated by the hyperbolic tangent function, keeping the feature amplitude in the range of [-1, 1] to suppress the risk of information overload, while allowing bidirectional regulation of positive reinforcement and negative inhibition (such as enhancing the conjugation effect signal and weakening steric noise). ; The activated cross-scale atom-functional group association features are bitwise added to the atomic-level features to obtain enhanced atomic features that fuse local details with global group context; global average pooling is performed on the enhanced atomic features to aggregate the common information of the local environment of each atom in the entire molecule (such as the average distribution density of the electron cloud) to obtain the atomic-level global pooling vector; global average pooling is performed on the functional group enhanced features to condense the overall molecular properties under the group contribution theory framework (such as the total polar group density) to obtain the functional group-level global pooling vector; the atomic-level global pooling vector, the functional group-level global pooling vector and the molecular global latent variable (encoding the overall potential characteristics of the molecule) are spliced ​​into a hybrid global feature to retain multi-granular chemical information; the hybrid global feature is input into a multi-layer perceptron to extract high-order nonlinear correlations layer by layer (such as the non-additive relationship between the synergistic effect between groups and physical properties), and finally a low-dimensional dense vector reflecting the comprehensive characteristics of the molecule is obtained.

[0041] The low-dimensional dense vector reflecting the comprehensive characteristics of the molecule is input into the fully connected layer and mapped into a prediction vector that matches the dimension of the target property (such as glass transition temperature); the prediction vector is compared with the measured value, and the mean square error is calculated sample by sample; the mean vector and logarithmic variance vector of the latent variable are obtained from the variational inference unit, and the KL divergence with the standard normal distribution is calculated; the low-dimensional dense vector is decoded into thermodynamic parameters (such as Gibbs free energy, entropy change), the chain consistency is verified through thermodynamic relations, and the detected numerical deviation is converted into a penalty term; the functional group-level feature vector is extracted and similarity matched with the group contribution theoretical data (such as the polar contribution value of the benzene ring +0.3), and a quadratic penalty is imposed on the group contribution that exceeds the theoretical allowable range; the mean square error, KL divergence, thermodynamic penalty and base The cluster penalty is superimposed according to the weight line to generate a multi-objective composite loss function; the curriculum learning strategy is used to dynamically adjust the constraint strength, specifically including focusing on data fitting in the early stage of training and reducing the weight of physical constraints; gradually increasing the constraint strength linearly in the middle stage to balance the data rules and theoretical rules; fixing the maximum constraint weight in the later stage to force the predicted value of the physical property model to strictly comply with the laws of chemical physics; during the training process, the total loss gradient is propagated back to the atomic / functional group feature extraction and molecular global latent variable generation module through the back propagation algorithm, and the physical rationality of the feature encoder, the continuity of the latent variable distribution and the mapping accuracy of the regression prediction layer are simultaneously optimized; after multiple rounds of iterations, the physical property model parameters converge to a stable state, and a trained physical property prediction model with physical constraint enhancement capabilities is obtained.

[0042] S4. Randomly sample initial latent variables from the standard normal distribution as the starting point for molecule generation; input the initial latent variables into the physical property prediction model through forward propagation to obtain the predicted physical property value corresponding to the current latent variable; according to the target physical property, use the backpropagation method to calculate the gradient of the predicted physical property value with respect to the latent variable; use the momentum-accelerated gradient ascent method to iteratively update the latent variables. After each round of iteration, project the updated latent variables to a legal area (such as the unit sphere or the truncated normal distribution range) to avoid generating invalid latent variables that deviate from chemical reality. If multiple consecutive rounds of optimization do not improve the physical properties, trigger the early stopping mechanism and reinitialize the latent variables; decode the optimized latent variables, and gradually generate a functional group sequence through the autoregressive model.

[0043] Furthermore, the stepwise generation of functional group sequences through an autoregressive model involves predicting the first functional group type (e.g., a benzene ring) and its connection anchor point (e.g., an ortho-substitution site). Based on the first k-1 functional groups generated, the kth functional group type and connection site (e.g., an ester group connected to the para-position of the benzene ring) are predicted until the preset sequence length is reached or a terminator is generated. Based on the functional group sequence and its connection rules, the functional group templates (e.g., a carboxylic acid group template) are spliced ​​according to the connection sites to construct an atomic-level molecular sketch. The atom-functional group affiliation is restored through a cross-level pointer mapping table, and unmatched residue atoms (e.g., isolated methyl groups) are completed. The ETKDG method of RDKit is used for atomic-level molecular sketches, combined with the empirical values ​​of bond lengths, bond angles, and torsion angles in the chemical knowledge base to generate low-energy candidate conformations; the MMFF94 force field (Merck Molecular Force Field 94) is applied to each low-energy candidate conformation for energy minimization to obtain the energy-optimal three-dimensional molecular structure; the energy-optimal three-dimensional molecular structure is subjected to atomic valence legitimacy check, ring strain threshold detection, chiral center consistency verification, and spatial conflict detection. Molecules that pass the verification are retained as valid candidates, and molecules that fail the verification are eliminated and the reasons are recorded.

[0044] Furthermore, energy minimization includes energy calculation and gradient descent optimization, as follows: Energy calculation refers to the calculation of molecular potential energy (including bond stretching energy, bond bending energy, torsion angle energy, van der Waals interaction and electrostatic interaction) based on force field parameters.

[0045] Gradient descent optimization refers to the iterative adjustment of atomic coordinates to converge the potential energy to a local minimum and eliminate bond length / bond angle distortion (such as optimizing an overly long CC single bond from 1.6Å to 1.54Å).

[0046] Atomic valence validity check involves traversing all atoms to verify whether their actual bond counts are equal to the theoretical valence (e.g., carbon atom bond count = 4, nitrogen atom bond count = 3). If a hypervalent atom is detected (e.g., pentavalent nitrogen), the system automatically attempts to add formal charges (e.g., converting pentavalent nitrogen into a positively charged ammonium ion).

[0047] Ring tension threshold detection refers to a deviation threshold set based on the ideal bond angle of common ring structures. The deviation threshold allows bond angle fluctuations of ±5°~15°. The deviation of the bond angle of atoms in the ring from the ideal value is calculated. If the ring tension exceeds the deviation threshold, it is marked as a high-tension ring, triggering ring reconstruction or direct removal. If it does not exceed the deviation threshold, the structure is retained and the tension value is recorded as a molecular stability assessment parameter.

[0048] Chiral center consistency verification refers to checking whether the R / S label of the chiral center matches the low-energy conformation. If there is a conflict (such as SMILES is R configuration but it is actually S configuration), stereochemical correction or elimination is triggered.

[0049] Spatial conflict detection refers to setting an overlap threshold based on the degree of overlap of the van der Waals radius between atoms. If the overlapping distance exceeds the overlap threshold (such as greater than 0.3Å), it is judged as a steric conflict, triggering conformational optimization or elimination. If it does not exceed the overlap threshold, the current conformation is accepted and the smallest atom is recorded.

[0050] The validated molecules are input into the trained physical property prediction model to calculate the target physical property prediction values; the target physical property prediction values ​​are sorted from high to low, and the top 10% are retained as candidate molecules.

[0051] S5. Set the screening threshold according to the target property range, use the trained property prediction model to quickly predict the candidate molecules, retain the molecules with predicted property values ​​falling within the target range, and eliminate the molecules with large predicted deviations; based on the position comparison of Morgan fingerprints, calculate the similarity of the molecules retained in the target range, and use the maximum and minimum sampling method to select the subset of molecules with the largest structural differences to avoid excessive homogenization of the primary screening library and obtain the primary screening molecular subset; the molecular subset that passes the primary screening enters the molecular dynamics simulation verification stage to perform energy minimization, equilibrium and production stage simulation, compare the key physical properties with the predicted values, and eliminate molecules with excessive deviations; the molecules that pass the simulation verification are further evaluated for synthetic feasibility, evaluate the synthetic feasibility score, environmental complexity and functional group compatibility, and reversely predict the feasible synthetic path, eliminate molecules that cannot be synthesized or do not meet the process standards, and obtain a verified set of molecules.

[0052] Furthermore, key properties refer to target properties that are directly related to the property prediction model and auxiliary properties related to molecular stability and synthesizability.

[0053] This embodiment also provides a material property prediction system based on deep learning, including: Parse the molecular SMILES string into an atomic-level molecular topology graph and generate a double-layer graph structure data; Based on the two-layer graph structure data, a cross-level message passing network is constructed to extract atomic-level features and functional group-level features, and a molecular global latent variable is generated through bidirectional cross-scale attention interaction; The global latent variables of the molecule are integrated with the atomic-level and functional group-level features, and physical constraints are imposed by combining the chemical bonding rule library and group contribution theory to construct a physical property prediction model; Based on the molecular global latent variables and physical property prediction model, candidate molecules are obtained through autoregressive generation and molecular force field verification; The physical property prediction model is used to screen candidate molecules, and the final molecular set is output after verification through molecular dynamics simulation and synthetic feasibility assessment.

[0054] This embodiment also provides a computer device, which is suitable for the case of a material property prediction method based on deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the material property prediction method based on deep learning proposed in the above embodiment.

[0055] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0056] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the material property prediction method based on deep learning as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0057] In summary, the present invention eliminates the semantic ambiguity of SMILES sequences through stereochemical standardization and chemical equivalence normalization, thereby solving the problem of decreased model generalization caused by inconsistent molecular representation; through VF2 subgraph isomorphism matching and weighted conflict resolution strategy, accurate division of functional group levels is achieved, and intra-group synergy and cross-group interaction are effectively encoded, so that the model can simultaneously capture atomic-level details and group-level chemical laws; physical constraint penalty terms and cross-scale attention mechanisms are introduced to force model predictions to conform to chemical intuition, the atomic weight correction module suppresses abnormal activation of hypervalent bond regions, and the group contribution theory alignment module strengthens asymmetric associations between functional groups, ensuring the physical rationality of the prediction results from both data-driven and theory-driven dual paths; through multi-stage verification of molecular dynamics simulation and synthetic feasibility assessment, high-strain rings or non-synthesizable structures are effectively eliminated, so that the generated molecules have both high performance and preparability.

[0058] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A material property prediction method based on deep learning, characterized by: include, Parse the molecular SMILES string into an atomic-level molecular topology graph and generate a double-layer graph structure data; Based on the two-layer graph structure data, a cross-level message passing network is constructed to extract atomic-level features and functional group-level features, and a molecular global latent variable is generated through bidirectional cross-scale attention interaction; The global latent variables of the molecule are integrated with the atomic-level and functional group-level features, and physical constraints are imposed by combining the chemical bonding rule library and group contribution theory to construct a physical property prediction model; Based on the molecular global latent variables and physical property prediction model, candidate molecules are obtained through autoregressive generation and molecular force field verification; The physical property prediction model is used to screen candidate molecules, and the final molecular set is output after verification through molecular dynamics simulation and synthetic feasibility assessment.

2. The material property prediction method based on deep learning according to claim 1, characterized in that: The generating of the double-layer graph structure data comprises the following steps: Using chemical informatics tools, the SMILES string is verified for atomic valence validity, bond connectivity rationality, and automatically corrected to generate a temporary SMILES sequence with syntax correction. Perform stereochemical analysis on the temporary SMILES sequence to generate a standardized SMILES sequence; Perform chemical equivalence normalization on standardized SMILES sequences, extract atomic-level molecular topology and annotate dimensional properties of atoms and chemical bonds; Based on the functional group dictionary, atomic-level molecular topology subgraph isomorphism matching and conflict optimization are performed to divide the conflict-free functional group structure and generate double-layer graph structure data.

3. The material property prediction method based on deep learning according to claim 2, characterized in that: The construction of the cross-level message passing network includes the following steps: Perform multi-layer graph convolution on the atomic-level molecular graph to aggregate neighborhood features to generate atomic-level features; Based on the functional group partitioning scheme, the atomic-level embedding vectors are spatially averaged and pooled, and functional group type encoding is combined to generate functional group-level features. A bidirectional mapping between atoms and functional groups is established through cross-level pointers, a two-layer message passing network is constructed, and molecular global latent variables are generated through bidirectional cross-scale attention interaction.

4. The material property prediction method based on deep learning according to claim 3, characterized in that: The construction of the physical property prediction model comprises the following steps: Decode the atomic-level embedding vector into bond length and bond angle parameters, and combine it with the chemical bonding rule library to impose penalties on the atomic weights that violate physical constraints to generate a revised atomic feature matrix; The functional group-level features are spliced ​​with the theoretical weights of group contributions to calculate the asymmetric attention weights; By integrating atomic-level global pooling vectors, functional group-level global pooling vectors, and molecular global latent variables, a multi-objective composite loss function is constructed and the parameters of the physical property prediction model are dynamically optimized to obtain a trained physical property prediction model.

5. The material property prediction method based on deep learning according to claim 4, characterized in that: The generation through autoregression and molecular force field verification refers to generating functional group sequences based on global latent variable optimization, splicing atomic-level molecular sketches based on functional group templates, generating three-dimensional structures through molecular force field energy minimization, and verifying valence legitimacy, ring tension and chirality consistency to obtain candidate molecules.

6. The material property prediction method based on deep learning according to claim 5, characterized in that: The molecular dynamics simulation refers to performing energy minimization and equilibrium and production stage simulation on candidate molecules, calculating key physical properties and comparing them with predicted physical property values, and eliminating molecules with excessive physical property deviations.

7. The material property prediction method based on deep learning according to claim 6, characterized in that: The synthetic feasibility assessment verification refers to calculating the synthetic feasibility score, environmental complexity and functional group compatibility, and reversely predicting feasible synthetic pathways, eliminating molecules that cannot be synthesized and that do not meet process standards, and obtaining a set of verified molecules.

8. A material property prediction system based on deep learning, based on the material property prediction method based on deep learning according to any one of claims 1 to 7, characterized in that: include, The structure analysis module parses the molecular SMILES string into an atomic-level molecular topology graph and generates a double-layer graph structure data; A cross-layer encoding module constructs a cross-layer message passing network based on the two-layer graph structure data structure, extracts atomic-level features and functional group-level features, and generates molecular global latent variables through bidirectional cross-scale attention interaction; The physical property prediction module integrates the molecular global latent variables with atomic-level and functional group-level features, combines the chemical bonding rule library and group contribution theory to impose physical constraints, and constructs a physical property prediction model; A molecule generation module, based on the molecular global latent variables and physical property prediction model, obtains candidate molecules through autoregressive generation and molecular force field verification; The screening and evaluation module uses the physical property prediction model to screen candidate molecules, and outputs the final molecular set through molecular dynamics simulation and synthetic feasibility evaluation verification.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the material property prediction method based on deep learning are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the material property prediction method based on deep learning according to any one of claims 1 to 7 are implemented.

Images