A deep learning-based material performance prediction method and system

Abstract

The application discloses a material performance prediction method and system based on deep learning, and relates to the technical field of material performance prediction, comprising the following steps: constructing a cross-level message passing network based on a double-layer graph data structure, extracting atomic-level and functional group-level feature embeddings, generating a molecular global latent variable through bidirectional cross-scale attention interaction, fusing the molecular global latent variable with the atomic-level and functional group-level features, combining a chemical bonding rule library and a group contribution theory to apply physical constraints, and constructing a physical property prediction model; based on the molecular global latent variable and the physical property prediction model, a candidate molecule is obtained through self-recurrence generation and verification of a molecular force field; the candidate molecule is screened by using the physical property prediction model, and a final molecule set is output through verification of molecular dynamics simulation and synthetic feasibility evaluation. Through multi-stage verification of molecular dynamics simulation and synthetic feasibility evaluation, the application effectively eliminates high-tension rings or non-synthesizable structures, so that the generated molecules have both high performance and preparability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material performance prediction, in particular to a material performance prediction method and system based on deep learning. BACKGROUND

[0002] In recent years, material performance prediction technology has gradually shifted from traditional methods based on empirical formulas to data-driven deep learning models. The existing mainstream methods mainly build prediction frameworks based on molecular graph neural networks (GNN) or simplified molecular descriptors (such as molecular fingerprints), and achieve property regression through end-to-end training. However, such methods still have significant limitations in the comprehensiveness of molecular representation and the compatibility of physical and chemical rules. The same molecule generates different graph structures due to differences in SMILES atomic ordering, leading to misjudgment of isomers by GNN. Without introducing a standardized atomic reordering algorithm, the model is forced to learn redundant topological transformation patterns, reducing the generalization ability. Without explicitly encoding chiral centers (R / S markers) and double bond cis / trans configurations, the model only relies on adjacency matrices to transmit topological information. For example, the model cannot distinguish between cis / trans butenedioic acid, resulting in a significant deviation between predicted values and actual properties. Existing methods directly use atomic-level graph structures without subgraph segmentation based on chemical group dictionaries. For example, benzene rings are disassembled into isolated carbon atom nodes, losing the overall electronic effect characteristics of conjugated systems and weakening the model's ability to represent aromaticity. Molecular fingerprints map continuous chemical properties to discrete binary bits through a hash function, losing geometric information such as bond length and bond angle. For example, it cannot distinguish between conjugated systems with alternating single and double bonds and isolated double bonds, resulting in blurred electron cloud distribution characteristics. The fingerprint design does not incorporate prior knowledge from group contribution methods. For example, the hydrogen bond contribution of the hydroxyl group and the acidity strength of the carboxylic acid are not encoded differently, making it difficult for the model to learn non-additive effects between groups. Free prediction of bond lengths exceeds reasonable ranges without constraints on output distribution through physical penalty terms. For example, when predicting hypervalent nitrogen compounds, five-coordinate structures are generated, violating valence bond theory. When predicting properties such as glass transition temperature, the model does not enforce thermodynamic consistency (such as the positive correlation between Tg and melting point), resulting in model outputs that violate phase transition rules. SUMMARY

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

[0004] Therefore, the present application provides a material performance prediction method based on deep learning to solve the problem of prediction deviation and unreasonable generated molecules caused by incomplete molecular representation and missing physical and chemical constraints in existing material performance prediction models.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides a material performance prediction method based on deep learning, comprising,

[0007] parsing the molecular SMILES string into an atomic-level molecular topological graph to generate double-layer graph structure data;

[0008] constructing a cross-level message passing network based on the double-layer graph structure data, extracting atomic-level features and functional group-level features, and generating a molecular global latent variable through bidirectional cross-scale attention interaction;

[0009] fusing the molecular global latent variable with the atomic-level and functional group-level features, combining a chemical bonding rule library and a group contribution theory to impose physical constraints, and constructing a property prediction model;

[0010] based on the molecular global latent variable and the property prediction model, generating a candidate molecule through autoregression and force field verification;

[0011] filtering the candidate molecule using the property prediction model, verifying it through molecular dynamics simulation and synthesis feasibility evaluation, and outputting a final molecule set.

[0012] As a preferred scheme of the material performance prediction method based on deep learning, the method comprises the following steps:

[0013] Performing atomic valence state legality verification, bond connection rationality checking and automatic correction on the SMILES string through chemical informatics tools to generate a temporarily corrected SMILES sequence;

[0014] Performing stereochemical analysis on the temporary SMILES sequence to generate a standardized SMILES sequence;

[0015] Performing chemical equivalence normalization on the standardized SMILES sequence, extracting an atomic-level molecular topological graph, and labeling the dimensional properties of atoms and chemical bonds;

[0016] Performing atomic-level molecular topological subgraph isomorphism matching and conflict optimization based on a functional group dictionary, dividing non-conflict functional group structures, and generating double-layer graph structure data.

[0017] As a preferred scheme of the material performance prediction method based on deep learning, the method comprises the following steps:

[0018] Performing multi-layer graph convolution on the atomic-level molecular graph to aggregate neighborhood features and generate atomic-level features;

[0019] Performing spatial average pooling on the atomic-level features based on a functional group division scheme, combining functional group type encoding to generate functional group-level features;

[0020] A two-way mapping between atoms and functional groups is established through cross-level pointers to construct a two-layer message passing network, and a molecular global latent variable is generated through two-way cross-scale attention interaction.

[0021] As a preferred scheme of the material performance prediction method based on deep learning provided by the application, the property prediction model comprises the following steps,

[0022] The atomic-level embedding vector is decoded into bond length and bond angle parameters, and a revised atomic feature matrix is generated by combining a chemical bonding rule library to impose a penalty on atomic weights that violate physical constraints.

[0023] The functional group-level features are spliced with the group contribution theory weights to calculate asymmetric attention weights.

[0024] The atomic-level global pooling vector, the functional group-level global pooling vector, and the molecular global latent variable are fused to construct a multi-objective composite loss function and dynamically optimize the property prediction model parameters to obtain a trained property prediction model.

[0025] As a preferred scheme of the material performance prediction method based on deep learning provided by the application, the generation of the functional group sequence based on the global latent variable optimization refers to splicing the atomic-level molecular sketch based on the functional group template, generating a three-dimensional structure through molecular force field energy minimization, and performing valence state legality, ring strain, and chirality consistency verification to obtain a candidate molecule.

[0026] As a preferred scheme of the material performance prediction method based on deep learning provided by the application, the molecular dynamics simulation refers to performing energy minimization, equilibrium, and production phase simulation on the candidate molecule, calculating key properties and comparing them with predicted property values, and eliminating molecules with excessive property deviations.

[0027] As a preferred scheme of the material performance prediction method based on deep learning provided by the application, the synthesis feasibility evaluation verification refers to calculating synthesis feasibility scores, environmental complexity, and functional group compatibility, and inversely predicting feasible synthesis paths to eliminate molecules that cannot be synthesized or have substandard processes to obtain a verified molecule set.

[0028] In a second aspect, the application provides a material performance prediction system based on deep learning, comprising,

[0029] The structure analysis module parses the molecular SMILES string into an atomic-level molecular topology graph to generate a double-layer graph structure data.

[0030] The cross-layer encoding module constructs a cross-level message passing network based on the double-layer graph structure data, extracts atomic-level features and functional group-level features, and generates a molecular global latent variable through two-way cross-scale attention interaction.

[0031] The property prediction module integrates the global latent variables of the molecule with atomic-level and functional group-level features, and applies physical constraints by combining the chemical bonding rule library and group contribution theory to construct a property prediction model.

[0032] The molecule generation module, based on the global latent variables and property prediction model of the molecules, obtains candidate molecules through autoregressive generation and molecular force field verification;

[0033] The screening and evaluation module uses the property prediction model to screen candidate molecules, verifies them through molecular dynamics simulation and synthesis feasibility assessment, and outputs the final molecule set.

[0034] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the deep learning-based material property prediction method as described in the first aspect of the present invention.

[0035] Fourthly, 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 deep learning-based material property prediction method as described in the first aspect of the present invention.

[0036] The beneficial effects of this invention are as follows: By stereochemical standardization and chemical equivalence normalization, semantic ambiguity of the SMILES sequence is eliminated, solving the problem of decreased model generalization caused by inconsistent molecular characterization; by using VF2 subgraph isomorphic matching and weighted conflict resolution strategies, precise functional group hierarchical division is achieved, effectively encoding intra-group synergistic effects and cross-group interactions, enabling the model to simultaneously capture atomic-level details and group-level chemical regularities; by introducing physical constraint penalty terms and cross-scale attention mechanisms, the model predictions are forced 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 the asymmetric correlation between functional groups, ensuring the physical rationality of the prediction results through both data-driven and theory-driven approaches; through multi-stage verification of molecular dynamics simulation and synthetic feasibility assessment, high-stretch rings or unsynthesizable structures are effectively eliminated, making the generated molecules both high-performance and preparable. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Fig. 1A flowchart for the deep learning-based material property prediction method in Example 1.

[0039] Fig. 2 A flowchart for the cross-layer encoding in Example 1.

[0040] Fig. 3 A flowchart for the property prediction in Example 1.

[0041] Fig. 4 A flowchart for the molecule generation and verification in Example 1. DETAILED DESCRIPTION

[0042] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0043] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the spirit of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0044] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0045] Example 1, with reference to Figs. 1-4 The embodiment provides a deep learning-based material property prediction method, comprising the following steps:

[0046] S1. Take the SMILES string as the initial data, perform molecular integrity check through the chemical information toolkit, the molecular integrity check includes checking the legality of atomic valence (such as carbon atom tetravalence, oxygen atom divalent), the rationality of bond connection (such as not allowing hypervalence bond, ring closure error) and automatically correcting invalid connection (for example, converting pentavalent nitrogen into charged form, repairing unclosed aromatic ring), after passing the integrity check, a temporarily corrected SMILES sequence is obtained, and if it cannot be corrected, it is marked as invalid input and the process is terminated.

[0047] The temporarily corrected SMILES sequence passing the check is subjected to stereochemical analysis, and the specific analysis operation is as follows:

[0048] Identify the chiral center and the double bond cis-trans configuration, use the three-dimensional coordinate generation algorithm (such as the ETKDG method) to calculate the lowest energy conformation, and convert the relative stereo description into absolute configuration (such as R / S label) to generate a standardized SMILES sequence with clear stereochemical information.

[0049] Perform chemical equivalence normalization on the standardized SMILES sequence with clear stereochemical information, including uniform representation of bond types (such as converting aromatic bonds to localized double bond / single bond alternating structures), removing redundant hydrogen atoms (explicit hydrogen to implicit), resetting the order of atomic numbers to achieve the canonical uniqueness of the SMILES string (avoiding different string representations of the same molecule), and obtaining a unique standardized SMILES sequence; input the standardized SMILES into the RDKit (open source chemical information toolkit) molecular object, extract the atomic-level molecular topology graph, as follows:

[0050] Each atom is labeled with 12-dimensional attributes as node features, including element type (One-hot encoding), formal charge (-1, 0, +1), valence (current bond number), hybridization state (sp³ / sp² / sp), ring membership (whether in a ring and ring size), chirality (R / S / None); Each chemical bond is labeled with 6-dimensional attributes as edge features, including bond type (single / double / triple / aromatic), bond length (Å), stereochemistry (cis / trans / none), conjugation, ring membership, and bond order (integer).

[0051] Load a predefined dictionary of chemical groups containing typical functional groups and perform VF2 subgraph isomorphism matching with the atomic-level molecular topology graph, as follows:

[0052] Use a depth-first strategy to traverse the atomic-level molecular topology graph, preferentially match multi-atom complex groups (such as benzene rings, acrylate chains), generate temporary subgraph segmentation schemes through each matching operation, and record complex group types, atomic-level molecular topology graph atom members, and cross-group connection bonds, to obtain a candidate functional group set segmentation scheme containing group type labels, member atom indices, and connection bond endpoint information.

[0053] Perform conflict detection and optimization selection on the segmentation scheme of the candidate functional group set. When an atom is matched by multiple group templates simultaneously, use a weighted scoring strategy to preferentially retain the group with more matching atoms, and the remaining single atoms are classified as independent nodes into the residue category; The cross-group connection bond is reconstructed as a functional group edge according to the endpoint atom attribution relationship to ensure the topological integrity of the atomic-level molecule, generating a conflict-free functional group division scheme.

[0054] The 3-layer graph convolution network is run on the atomic-level molecular topology graph, each atomic node aggregates the atomic / bond features within 3-hop neighborhood, and generates an atomic-level embedding vector, which encodes the local chemical environment information (such as electron cloud distribution, steric hindrance effect).

[0055] The atomic-level embedding vector is spatially averaged and pooled according to the non-conflicting functional group division scheme to obtain the aggregated atomic environment features within the group; the aggregated atomic environment features within the group are concatenated with the functional group type One-hot encoding and input into the MLP (Multi-Layer Perceptron) for nonlinear transformation to generate the functional group-level feature vector; based on the non-conflicting functional group division scheme and the functional group-level feature vector, a double-layer molecular graph structure data is constructed, as follows:

[0056] For the atomic-level graph, the original atomic-level molecular topology graph features are retained, and the atomic-level embedding vector is attached to the node; for the functional group-level graph, the node contains the functional group-level feature vector, and the edge inherits the cross-group bond attribute; through the cross-level pointer, a bidirectional mapping table of the functional group node and the member atomic index list is established to realize cross-level association, and the double-layer molecular graph structure data is obtained.

[0057] S2. The atomic-level molecular topology graph (including node features and edge features) of the double-layer graph structure data is input into a 3-layer message passing network, and each layer performs the following operations:

[0058] The message vector (linear transformation of edge type embedding + source node feature + target node feature) is calculated for each edge; each atomic node aggregates its adjacent edge messages (gated summation) and updates the node state through the 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.

[0059] According to the bidirectional mapping table across levels, the level-specific molecular graph summary vectors are grouped according to the belonging functional group, and the spatial average pooling operation is performed for each functional group node to reduce the dimension of all atomic embedding vectors within the group to the initial embedding vector of the functional group, which is 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 adjacent node features are weighted and aggregated according to the calculated attention coefficient, and the node state is updated through the LeakyReLU (leaky rectified linear unit) activation function; the updated node features and the functional group node input features are added to retain the original topology information, and the functional group-level context embedding vector is obtained.

[0060] Bidirectional cross-scale attention interaction is performed on the atomic-level embedding vector and the functional group-level context embedding vector, as follows:

[0061] The attention weight is calculated based on the atomic-level embedding vector and the functional group-level context embedding vector, an atomic query vector and a functional group key vector are generated, and the atomic features are weighted and injected into the functional group node to update the functional group embedding as context-aware features; the functional group query vector and the atomic key vector are generated in reverse through the attention of the functional group-level context embedding vector and the atomic-level embedding vector, the functional group feature vector is broadcast to the atomic node to enhance the cross-level semantics of the atomic embedding; after multiple iteration cycles of optimization, a dual-path embedding vector set containing atoms and functional groups is generated.

[0062] Based on the topology structure of the functional group-level graph, the adjacency number of each functional group node is calculated, the core group (such as a benzene ring) with the most connections of cross-group chemical bonds is preferentially selected as the root node, when there are multiple candidates, the context embedding vector of the largest group is selected as the initial input by further comparing the group atomic size; the selected root node embedding vector is input into the batch normalization layer to eliminate the feature distribution difference among different molecules, and then the value range is restricted to the interval [-1, 1] through the hyperbolic tangent activation function, and the normalized feature vector is output; the normalized feature is input into the variational inference processing unit, and is processed in parallel through two fully connected paths: the mean path passes through the ReLU activated intermediate layer, and finally outputs a linear mean vector; the variance path shares the same structure of the intermediate layer, but the end adopts the softplus activation to generate a logarithmic variance vector.

[0063] S3. The atomic-level embedding vector is input into the gated attention layer, the cross-attention score between atoms is calculated, and the atom pairs with synergistic effect (such as alternating atoms in the conjugated system) are identified; through the learnable gating parameter, low correlation interaction is filtered, and significant chemical action mode (such as hydrogen bond donor-acceptor pair) is reserved, and an enhanced feature vector with attention weighting is output; the enhanced feature vector is input into a double-layer perception machine: the first layer uses GeLU activation function to extract nonlinear correlation, and the second layer uses linear transformation to generate initial importance weight.

[0064] Further, the generated initial importance weight reflects the prediction contribution of the atom pair to the target physical property (such as glass transition temperature), and the value range is 0 to 1.

[0065] Based on the chemical bonding rule library (such as the C=O double bond bond length reference range 1.20-1.25 Å), the atomic-level embedding vector is decoded into the 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 the C=O bond length predicted as 1.30 Å), the exponential decay penalty is applied to the initial importance weight, the contribution in the high confidence conflict area is reduced, and the modified importance weight is obtained; The modified importance weight and the original atomic embedding vector are multiplied element by element to suppress the atomic features that do not conform to the chemical intuition (such as the abnormal activation value of the hyper-valence bonding area), while retaining the physically reasonable local environment information, and output the screened atomic feature matrix.

[0066] The functional group-level embedding vector (encoding group type and atomic environment features) is aligned and spliced with the group contribution theory weight according to the functional group node to form a functional group mixed feature matrix that integrates theoretical prior knowledge; using a learnable linear projection matrix, the functional group mixed feature matrix is mapped into a query vector (capturing the information needs of the current functional group) and a key vector (representing the interaction potential of other functional groups), and a learnable cross-group attention mechanism is used to calculate the similarity score between all functional group mixed feature matrix nodes; the similarity score is normalized in the row direction by softmax, generating 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; according to the attention weight matrix, the mixed features are weighted and aggregated, and at the same time the features are reduced through a linear projection network, and cross-correlation analysis is performed, the cosine similarity between functional group nodes is calculated, potential synergistic group pairs are identified, functional group enhanced features that retain driving characteristics and conform to group contribution theory are generated, and are associated with the screened atomic-level features according to the cross-level mapping table.

[0067] Further explanation, deep correlation mode refers to group synergistic effect, that is, nonlinear superposition between non-homologous groups.

[0068] 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 of the screened atomic-level features and the functional group enhanced features, a plurality of independent attention heads are split, each attention head learns different chemical action modes (such as van der Waals interaction, hydrogen bond formation tendency), and captures the asymmetric relationship between atoms and functional groups in a multi-dimensional space in parallel, each attention head independently outputs a weighted functional group feature segment, and after splicing according to the attention head, the feature dimension is restored through linear transformation, generating a cross-scale reinforced atomic-functional group association matrix; the cross-scale atomic-functional group association features are activated by a hyperbolic tangent function, the feature amplitude is kept in the interval [-1, 1], the risk of information overload is inhibited, and at the same time, bidirectional regulation of positive reinforcement and negative inhibition is allowed (such as enhancing conjugation effect signals and weakening steric hindrance noise); the activated cross-scale atomic-functional group association features and the atomic-level features are added according to the position, and enhanced atomic features that fuse local details and global group context are obtained; global average pooling is performed on the enhanced atomic features, and the common information of the local environment of each atom in the whole molecule (such as the average distribution density of the electron cloud) is aggregated, and an atomic-level global pooling vector is obtained; global average pooling is performed on the functional group enhanced features, and the overall attributes of the molecule under the group contribution theory framework (such as the total polar group density) are condensed, and a functional group-level global pooling vector is obtained; 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 hybrid global features, and multi-granularity chemical information is retained; the hybrid global features are input into a multi-layer perceptron, high-order nonlinear correlations (such as non-additive relationships between group synergies and physical properties) are extracted layer by layer, and finally a low-dimensional dense vector reflecting the comprehensive characteristics of the molecule is obtained.

[0069] The low-dimensional dense vector reflecting the comprehensive characteristics of the molecule is input into the full connection layer, and is mapped into a prediction vector matching 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 the logarithmic variance vector of the latent variable are obtained from the variational inference unit, and the KL divergence thereof and the standard normal distribution is calculated; the low-dimensional dense vector is decoded into thermodynamic parameters (such as Gibbs free energy and entropy change), the chain consistency is verified through the thermodynamic relationship, and the detected numerical deviation is converted into a penalty term; the functional group level feature vector is extracted, similarity matching is performed with the group contribution theory data (such as the polarity contribution value of a benzene ring +0.3), and a quadratic penalty is applied to the group contribution degree exceeding the theoretical allowable range; the mean square error, the KL divergence, the thermodynamic penalty and the group penalty are superimposed according to the weight line to generate a multi-objective composite loss function; a course learning strategy is adopted to dynamically adjust the constraint strength, specifically including that in the early training stage, data fitting is mainly performed, and the physical constraint weight is reduced; in the middle stage, the constraint strength is gradually linearly increased, and the data rule and the theoretical rule are balanced; in the later stage, the maximum constraint weight is fixed, and the property model prediction value is forced to strictly comply with the chemical and physical law; during the training process, the total loss gradient is returned to the atom / functional group feature extraction and molecular global latent variable generation module through the back propagation algorithm, and the physical reasonableness 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 iterations, the property model parameters converge to a stable state, and a trained property prediction model with enhanced physical constraint ability is obtained.

[0070] S4. Randomly sample an initial latent variable from a standard normal distribution as a starting point for molecule generation; input the initial latent variable into the property prediction model to obtain the predicted property value corresponding to the current latent variable through forward propagation; according to the target property, the gradient of the predicted property value with respect to the latent variable is calculated using the back propagation method; the momentum acceleration gradient ascent method is used to iteratively update the latent variable, and after each iteration, the updated latent variable is projected into a legal region (such as a unit sphere or a truncated normal distribution range) to avoid generating invalid latent variables that deviate from chemical reality, and if the property is not improved after continuous multiple optimization, an early stopping mechanism is triggered and the latent variable is reinitialized; the optimized latent variable is decoded, and the functional group sequence is gradually generated through an autoregressive model.

[0071] Further, the functional group sequence is gradually generated through an autoregressive model, which means predicting the first functional group type (such as a benzene ring) and its connection anchor point (such as an ortho substitution site), predicting the kth functional group type and connection site (such as an ester group connected to the para position of a benzene ring) based on the first k−1 generated functional groups, and so on until reaching a preset sequence length or generating a termination symbol; according to the functional group sequence and its connection rules, the functional group templates (such as a carboxylic acid group template) are spliced according to the connection sites to construct an atomic-level molecular sketch, and the atom-functional group attribution relationship is recovered through a cross-level pointer mapping table to complete the unmatched residual atoms (such as isolated methyl groups).

[0072] The ETKDG method of RDKit is used for atomic-level molecular sketch, combined with the bond length, bond angle and torsion angle empirical values 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 an energy-optimal three-dimensional molecular structure; the energy-optimal three-dimensional molecular structure is subjected to atomic valence legality check, ring strain threshold detection, chiral center consistency verification and spatial conflict detection, and the verified molecules that pass the verification are retained as effective candidates, and the molecules that fail to pass the verification are eliminated and the reasons are recorded.

[0073] Further, the energy minimization includes energy calculation and gradient descent optimization, specifically as follows:

[0074] Energy calculation refers to calculating the potential energy of the molecule based on force field parameters (including bond stretching energy, bond angle bending energy, torsion angle energy, van der Waals interaction and electrostatic interaction).

[0075] Gradient descent optimization refers to iteratively adjusting atomic coordinates to make the potential energy converge to a local minimum, eliminating bond length / bond angle distortion (such as optimizing a too long C-C single bond from 1.6 Å to 1.54 Å).

[0076] Atomic valence legality check refers to traversing all atoms to verify whether the actual bond number is equal to the theoretical valence (such as carbon atom bond number = 4, nitrogen atom bond number = 3), and if an overvalent atom (such as a pentavalent nitrogen) is detected, an attempt is automatically made to add a formal charge (such as converting a pentavalent nitrogen to a positively charged ammonium ion form).

[0077] Ring strain threshold detection refers to setting a deviation threshold based on the ideal bond angle of common ring structures, with an allowable bond angle fluctuation of ±5° to 15°, calculating the deviation of the bond angle of the atoms in the ring from the ideal value, and if the ring strain energy exceeds the deviation threshold, marking it as a high-strain ring, triggering ring reconstruction or direct elimination, and if it does not exceed the deviation threshold, the structure is retained and the strain value is recorded as a molecular stability evaluation parameter.

[0078] Chiral center consistency verification refers to checking whether the R / S label of the chiral center matches the low-energy conformation, and if there is a conflict (such as SMILES being R configuration but actually being S configuration), triggering stereochemistry correction or elimination.

[0079] Spatial conflict detection refers to setting an overlap threshold based on the degree of overlap of atomic van der Waals radii, and if the overlap distance exceeds the overlap threshold (such as greater than 0.3 Å), it is determined to be a spatial steric conflict, triggering conformation optimization or elimination, and if it does not exceed the overlap threshold, the current conformation is accepted and the minimum atom is recorded.

[0080] The molecules passing the verification are input into the trained property prediction model to calculate target property prediction values; the target property prediction values are sorted from high to low, and the top 10% are retained as candidate molecules.

[0081] S5. According to the target property interval, a screening threshold is set, the trained property prediction model is used to quickly predict the candidate molecules, the molecules with predicted property values falling within the target interval are retained, and the molecules with large deviation are removed; based on the position comparison of the Morgan fingerprint, the similarity of the molecules retained in the target interval is calculated, and the maximum minimum sampling method is used to select a subset of molecules with the largest structural difference, to avoid excessive homogeneity of the primary screening library, to obtain a subset of molecules screened; the subset of molecules screened enters the molecular dynamics simulation verification stage to perform energy minimization, equilibrium and production stage simulation, compares the key properties with the prediction values, and removes the molecules with excessive deviation; the molecules passing the simulation verification are further subjected to synthesis feasibility evaluation, evaluate the synthesis feasibility score, environmental complexity and functional group compatibility, and inversely predict the feasible synthesis path, eliminate the molecules that cannot be synthesized or have unqualified process, to obtain a set of molecules passing the verification.

[0082] Further, the key properties refer to the target properties directly related to the property prediction model and the auxiliary properties related to the stability and synthesizability of the molecules.

[0083] The embodiment also provides a material performance prediction system based on deep learning, comprising:

[0084] The molecular SMILES string is parsed into an atomic-level molecular topological graph to generate a double-layer graph structure data;

[0085] A cross-level message passing network is constructed based on the double-layer graph structure data, atomic-level features and functional group-level features are extracted, a molecular global latent variable is generated through bidirectional cross-scale attention interaction;

[0086] The molecular global latent variable is fused with the atomic-level and functional group-level features, a property prediction model is constructed by combining a chemical bonding rule library and a group contribution theory to impose physical constraints;

[0087] Based on the molecular global latent variable and the property prediction model, a self-recurrent generation and molecular force field verification are performed to obtain candidate molecules;

[0088] The candidate molecules are screened by using the property prediction model, verified by molecular dynamics simulation and synthesis feasibility evaluation, and a final molecule set is output.

[0089] The embodiment also provides a computer device suitable for the material performance 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 the computer executable instructions to realize the material performance prediction method based on deep learning proposed in the above embodiment.

[0090] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0091] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to realize the material performance prediction method based on deep learning proposed in the above embodiment. The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0092] In summary, the present application eliminates the semantic ambiguity of SMILES sequence by stereochemical standardization and chemical equivalence normalization, solves the problem of model generalization decline caused by inconsistent molecular representation; through VF2 subgraph isomorphism matching and weighted conflict resolution strategy, the accurate division of functional group level is realized, the synergistic effect within the group and the interaction between groups are effectively encoded, so that the model can capture both atomic-level details and group-level chemical laws; the introduction of physical constraint penalty term and cross-scale attention mechanism forces the model to predict in line with chemical intuition, the atomic weight correction module suppresses the abnormal activation of hypercoordinate bond region, and the group contribution theory alignment module strengthens the asymmetric association between functional groups, which guarantees the physical reasonableness of the prediction results from data-driven and theory-driven paths; through multi-stage verification of molecular dynamics simulation and synthetic feasibility evaluation, high-tension rings or non-synthesizable structures are effectively removed, so that the generated molecules have both high performance and preparability.

[0093] It should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application 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 application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A deep learning based material performance prediction method, characterized in that: The application relates to a method for generating a molecular graph structure and a molecular property prediction model. The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:

2. The deep learning-based material property prediction method of claim 1, wherein: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:

3. The deep learning-based material property prediction method of claim 2, wherein: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps 4. The deep learning-based material property prediction method of claim 3, wherein: The molecular dynamics simulation refers to performing energy minimization and equilibrium and production stage simulation on the candidate molecules, calculating key physical properties and comparing with predicted physical property values, and eliminating molecules with excessive physical property deviation.

5. The deep learning-based material property prediction method of claim 4, wherein: The synthesis feasibility evaluation verification refers to calculating synthesis feasibility scores, environmental complexity and functional group compatibility, and inversely predicting feasible synthesis paths, eliminating molecules that cannot be synthesized and have unqualified processes, and obtaining a verified molecule set.

6. A deep learning-based material performance prediction system based on the deep learning-based material performance prediction method of any one of claims 1 to 5, characterized by: The structure analysis module parses the molecule SMILES string into an atomic-level molecular topology graph to generate a double-layer graph structure data. The cross-layer encoding module constructs a cross-layer message passing network based on the double-layer graph structure data structure, extracts atomic-level features and functional group-level features, and generates a molecular global latent variable through bidirectional cross-scale attention interaction. The physical property prediction module fuses the molecular global latent variable with atomic-level and functional group-level features, combines a chemical bonding rule library and a group contribution theory to impose physical constraints, and constructs a physical property prediction model. The molecule generation module generates candidate molecules through self-recurrence based on the molecular global latent variable and the physical property prediction model, and verifies the molecules through molecular force field verification. The screening evaluation module screens candidate molecules using the physical property prediction model, and outputs a final molecule set through molecular dynamics simulation and synthesis feasibility evaluation verification. The processor executes the computer program to implement the steps of the deep learning-based material performance prediction method of any one of claims 1-5. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The computer program is executed by the processor to implement the steps of the deep learning-based material performance prediction method of any one of claims 1-5.

8. A computer readable storage medium having stored thereon a computer program, characterized in that: ​

Images