Machine learning-based high-throughput calculation method and system for catalytic materials

Abstract

The application relates to the technical field of simulation calculation, in particular to a high-throughput calculation method and system for catalytic materials based on machine learning. The method comprises the following steps: establishing a two-dimensional sparse matrix by taking a catalyst identifier and a condition parameter as indexes and taking a known performance index value as a filling value; generating a mixed feature vector by mapping through a multilayer perception network based on non-empty coordinates and the catalyst identifier; obtaining a sample score by performing feature fusion on the mixed feature vector through a neural collaborative filtering network and a full-connection neural network, intercepting samples based on the condition parameter, and outputting a to-be-verified list; calculating performance data by using a simulation calculation model, backfilling the performance data to the two-dimensional sparse matrix, and updating the non-empty coordinates; judging based on a termination condition, and when the termination condition is not met, iteratively predicting the non-empty coordinates; and when the termination condition is met, outputting a catalytic material list. The application deeply couples structured space indexing, double-channel deep learning and adaptive closed-loop iteration, and improves data utilization.

Description

Technical Field

[0001] This invention relates to the field of simulation computing technology, specifically to a high-throughput calculation method and system for catalytic materials based on machine learning. Background Technology

[0002] Digital design and high-throughput screening of catalytic materials. In existing computer-aided materials design systems, high-throughput computing is a mainstream data processing paradigm. This process typically includes three core data processing modules: First, the system uses computational geometry algorithms to generate a three-dimensional geometric model of nanoparticles based on pre-set surface energy scalar data; second, a physical simulation engine calculates the electronic structure eigenvectors of this geometric model; finally, these eigenvectors are input into a numerical simulation solver, where microscopic dynamics are analyzed by solving a system of ordinary differential equations to predict the macroscopic performance values ​​of the material under specific boundary conditions. This linear data flow of "modeling-simulation-prediction" enables the computer system to automatically traverse a large number of candidate structures and build a material performance database.

[0003] However, existing simulation systems based on linear workflows suffer from significant data processing deficiencies when handling high-dimensional design spaces. Existing Woolf construction algorithms typically generate fixed geometric topologies based on static input parameters (vacuum surface energy). However, in actual simulations, changes in boundary conditions (such as gas partial pressure and coverage) lead to reconstructed geometric shapes of physical entities. Current technologies lack mechanisms for dynamically updating geometric models, resulting in a mismatch between the simulation model and the real-time state of the physical entity. In the microdynamics analysis module, the rigid differential equations involved often face convergence difficulties or excessive computation time when dealing with large-scale reaction networks, severely limiting the overall system throughput. Furthermore, existing high-throughput simulation systems, when handling large-scale, multivariable material design problems, fail to intelligently adjust sampling strategies to prioritize high-value samples due to their open-loop exhaustive sampling approach. This results in a significant waste of computational resources on low-value or redundant samples, leading to extremely low data utilization.

[0004] To address this, a high-throughput computational method and system for catalytic materials based on machine learning is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a high-throughput calculation method and system for catalytic materials based on machine learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a high-throughput calculation method for catalytic materials based on machine learning, comprising:

[0007] Historical simulation datasets are retrieved, and parameter values ​​are obtained using catalyst identifiers and condition parameters as indices. Known performance index values ​​are used as filler values ​​to construct a two-dimensional sparse matrix. Known data regions are marked as non-empty coordinates, and unknown regions are marked as empty value domains.

[0008] Based on non-empty coordinates and catalyst identifiers, sample feature attributes, sample identifiers and condition parameter codes are obtained by querying the embedding table, and then mapped through a multilayer perceptron network to generate a hybrid representation vector.

[0009] Linear and nonlinear features are fused into the hybrid representation vector using a neural collaborative filtering network and a fully connected neural network to generate performance prediction values ​​and prediction dispersion. Based on the performance prediction values, prediction dispersion, and hybrid representation vector, sample scores are obtained. Based on conditional parameters, samples with scores higher than a dynamic threshold are selected, and a list to be verified is output.

[0010] The simulation calculation model is used to calculate the list of tests to obtain performance data, which is then backfilled into a two-dimensional sparse matrix, and the non-empty coordinates are updated and the dynamic threshold is adjusted.

[0011] The determination is based on the termination condition. If the termination condition is not met, iterative prediction is performed on the non-empty coordinates; if the termination condition is met, a list of catalytic materials is output.

[0012] Preferably, the specific generation process of the two-dimensional sparse matrix includes: retrieving a historical simulation dataset, extracting the catalyst identifier sequence and the condition parameter sequence, converting the catalyst identifier sequence into a row logical address, and simultaneously converting the condition parameter sequence into a column logical address, and combining them to generate a physicochemical spatial positioning index pair; initializing the two-dimensional sparse matrix, and writing the known performance index values ​​from the historical simulation dataset into the corresponding storage units according to the pointers of the physicochemical spatial positioning index pairs; traversing the address space, identifying storage units containing valid values, extracting row and column position information to generate a non-empty coordinate index table; identifying blank storage units that do not contain valid values, and recording their distribution positions to generate a null value domain mask map.

[0013] Preferably, the specific generation process of the hybrid characterization vector includes: receiving the non-empty coordinate index table, extracting the catalyst discrete index and conditional discrete index by parsing the row and column positioning addresses respectively; mapping the catalyst discrete index to a preset catalyst structure fingerprint matrix, and extracting the electronic structure feature vector as the sample identifier; mapping the conditional discrete index to an environment embedding matrix, and extracting the state feature vector as the conditional parameter encoding; retrieving the corresponding attribute value sequence from the associated attribute database as the sample feature attribute, and concatenating it with the sample identifier and the conditional parameter encoding execution channel to generate an aggregated input tensor; transmitting the aggregated input tensor to a multilayer perceptron network, and performing weighted summation and nonlinear activation function transformation through multi-level neuron nodes to obtain a feature value sequence as a hybrid characterization vector containing material-environment coupling features. The sample feature attributes include electronegativity, atomic radius, and valence electron number.

[0014] Preferably, the specific generation process of the performance prediction value and the prediction dispersion includes: transmitting the hybrid representation vector in parallel to the processing branches of the neural collaborative filtering network and the fully connected neural network; in the processing branch of the neural collaborative filtering network, performing element-wise multiplication of the feature dimensions on the hybrid representation vector, analyzing the second-order cross features of the catalyst components and the reaction environment within the feature vector, and outputting a cross-interaction feature tensor; in the processing branch of the fully connected neural network, driving the multi-layer neuron weight matrix to perform continuous matrix multiplication and nonlinear activation function transformation on the hybrid representation vector, reconstructing the nonlinear spatial structure, and outputting a nonlinear activation barrier tensor; concatenating the cross-interaction feature tensor and the nonlinear activation barrier tensor to construct a fused feature vector, calculating the weighted sum value as the performance prediction value; and using the variance evaluation node to analyze the dispersion of the feature distribution as the prediction dispersion.

[0015] Preferably, the specific generation process of the list to be verified includes: receiving the performance prediction value and the prediction dispersion, and retrieving the hybrid representation vector; mapping the hybrid representation vector to the feature latent space, calculating the manifold distance between the hybrid representation vector and the sample points in the non-empty coordinate set, and calling the kernel density estimation algorithm to generate the null domain distribution density; constructing an exploration and utilization balance evaluation model, receiving the performance prediction value, the prediction dispersion, and the null domain distribution density as input variables, performing nonlinear weighted fusion operation, and generating sample scores; retrieving the dynamic threshold value of the current iteration cycle, performing a numerical comparison operation between the sample scores and the dynamic threshold, and filtering out a subset of candidate samples whose scores exceed the dynamic threshold; parsing the condition parameters associated with the candidate sample subset, performing physicochemical parameter logical verification according to the preset experimental boundary constraints, and eliminating invalid samples that do not meet the experimental boundary constraints; and assembling the passed sample indexes and corresponding parameter codes into a list to be verified.

[0016] Preferably, the specific process of updating non-empty coordinates and adjusting dynamic thresholds includes: performing format conversion on the condition parameter encoding in the list to be verified, generating an input control file stream, distributing it to the computing node, performing first-principles simulation operations through the simulation calculation model, monitoring the convergence status of the operation and capturing the output log data stream; parsing the output log data stream and extracting scalar values ​​characterizing material activity as performance data; locking the storage address of the corresponding empty value domain in the two-dimensional sparse matrix through the data storage manager and writing the performance data into the corresponding storage unit; scanning the storage unit position of the newly written data, extracting its row and column positioning information, appending the row and column positioning information to the non-empty coordinate index table, and updating the non-empty coordinate set; retrieving the performance data set and performance prediction value set corresponding to the updated non-empty coordinates, calculating the distribution gradient of the prediction error, correcting the boundary parameters of the sampling strategy based on the distribution gradient, and outputting the updated dynamic threshold value.

[0017] Preferably, the specific architecture process of the simulation calculation model includes: receiving a list of samples to be verified, calling a preset crystal structure template based on the sample identifier, and adjusting the lattice parameters according to the condition parameters to generate spatial geometric topology data characterizing the microstructure of the material; constructing a discretized numerical calculation grid based on the spatial geometric topology data, and setting boundary constraints in combination with preset environmental parameters to generate a simulation input file; loading the simulation input file into the iterative solver, initializing the system state vector, and driving the self-consistent field cyclic calculation engine to perform nonlinear variational solution of the control equations; calculating the convergence residual of the energy functional in real time during the solution process, and determining that the calculation has converged and locking the system steady state when the convergence residual falls within the preset tolerance range; triggering the post-processing extraction module to analyze the electronic structure data under steady state and calculate the scalar values ​​characterizing the catalytic activity of the material as performance data.

[0018] Preferably, the specific process of determining based on termination conditions includes: receiving the updated two-dimensional sparse matrix and retrieving the performance data distribution statistics of the most recent consecutive iterations; performing convergence evaluation operations to calculate the decay rate of performance prediction error and the gradient of high-scoring sample discovery rate as system state indicators; mapping the system state indicators to a preset termination determination logic controller; if the determination result indicates that the model state is in an oscillation period, generating an iterative control signal, defining the updated non-empty coordinate set as a new seed dataset, and transmitting it to the embedded table query interface through a feedback loop to drive the multilayer perceptron network to perform a new round of value evaluation on the remaining empty value domain and correct the sampling strategy; if the determination result indicates that the model state is in a steady-state convergence period, activating the result output generator; performing a full scan on all non-empty coordinates in the two-dimensional sparse matrix according to the final dynamic threshold standard to extract target units with qualified performance values; parsing the sample identifiers associated with the target units and constructing a list of catalytic materials containing material composition and structural characteristics.

[0019] A high-throughput computing system for catalytic materials based on machine learning, comprising:

[0020] Matrix construction module: retrieves historical simulation datasets, uses catalyst identifiers and condition parameters as indices to obtain parameter values, uses known performance index values ​​as fill values ​​to build a two-dimensional sparse matrix; marks known data regions as non-empty coordinates and unknown regions as empty value domains;

[0021] Feature mapping module: Based on non-empty coordinates and catalyst identifiers, it obtains sample feature attributes, sample identifiers and condition parameter codes by querying the embedding table, and then maps them through a multilayer perceptron network to generate a hybrid representation vector;

[0022] Sample extraction module: Linear and nonlinear features are fused into the hybrid representation vector through a neural collaborative filtering network and a fully connected neural network to generate performance prediction values ​​and prediction dispersion; Based on the performance prediction values, prediction dispersion, and hybrid representation vector, the sample score is obtained; Based on the conditional parameters, samples with scores higher than the dynamic threshold are extracted, and a list to be verified is output.

[0023] Performance backfilling module: Calculates the list to be verified using a simulation calculation model, obtains performance data, backfills it into a two-dimensional sparse matrix, updates non-empty coordinates, and adjusts dynamic thresholds;

[0024] Termination Judgment Module: Based on the termination condition, a judgment is made. If the termination condition is not met, iterative prediction is performed on the non-empty coordinates; if the termination condition is met, a list of catalytic materials is output.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. By constructing a dual-channel network architecture, this study addresses the problem of numerous low-value samples being miscalculated due to inaccurate feature extraction during catalytic material screening. Through linear and nonlinear feature fusion of mixed representation vectors using a neural collaborative filtering network and a fully connected neural network, complex structure-activity relationships can be captured when dealing with large-scale catalyst components and reaction conditions. This mechanism generates performance prediction values ​​and prediction dispersion, and accordingly truncates samples with scores higher than a dynamic threshold. Compared to traditional models, this architecture more accurately identifies materials with truly high catalytic activity, avoiding the waste of computational resources on false positives and low-value samples due to prediction bias. Feature fusion reduces the proportion of invalid simulation calculations, thereby improving the screening hit rate and data utilization efficiency for high-value catalytic materials with limited computing power.

[0027] 2. By introducing a sparse matrix and distribution density, the resource waste caused by repeated sampling in the catalytic material design space is solved. A two-dimensional sparse matrix is ​​established using catalyst identifiers and conditional parameters as indices, and unknown regions are marked as empty domains, intuitively mapping the vast catalytic chemistry space. When generating the list of samples to be validated, sample scores are obtained based on performance predictions, prediction dispersion, and the distribution density of the empty domains. By incorporating the distribution density of the empty domains into the evaluation system, information-scarce regions in the catalytic space can be intelligently identified, avoiding redundant sampling in areas with dense known data. This strategy ensures that computational resources are preferentially allocated to unknown catalytic systems that can maximize the model's generalization ability, thereby covering the largest material search space with minimal computation and improving data utilization.

[0028] 3. By constructing a closed-loop iteration and dynamic adjustment mechanism, the problem of open-loop systems failing to eliminate low-value samples and thus consuming large amounts of resources is solved. Using a simulation model to calculate the list of samples to be validated, performance data can be obtained and backfilled into a two-dimensional sparse matrix. This not only expands the catalytic database but, more importantly, updates non-empty coordinates and adjusts dynamic thresholds based on feedback. As iteration progresses, a termination condition is used for judgment; if it is not met, iterative prediction is performed on the non-empty coordinates. This process allows the screening criteria to adaptively improve as the understanding of catalytic performance deepens, thereby automatically filtering out a large number of low-to-medium performance catalytic materials in the later stages of iteration. This ensures that high-throughput simulation computing power is always concentrated on high-value samples, changing the situation of wasted computing resources caused by open-loop exhaustive search.

[0029] 4. By deeply coupling structured spatial indexing, dual-channel deep learning, and adaptive closed-loop iteration, a fully intelligent catalytic material screening system was constructed. First, a two-dimensional sparse matrix was established, and unknown regions were marked as null domains, providing structured navigation for global optimization. Based on this, the high-precision feature extraction capabilities of neural collaborative filtering networks and fully connected neural networks were utilized to obtain sample scores based on performance prediction values, prediction dispersion, and null domain distribution density, achieving multi-dimensional and accurate evaluation of sample value. By truncating samples with scores higher than a dynamic threshold for calculation and backfilling the results into the two-dimensional sparse matrix, updating non-null coordinates, and adjusting the dynamic threshold, a virtuous cycle of model-guided sampling and sampling-feedback model was achieved during iterative prediction of non-null coordinates. This overall collaborative mechanism ensures that each simulation calculation reduces model uncertainty or discovers high-performance catalysts, thereby systematically avoiding resource waste on low-value or redundant samples and improving data utilization. Attached Figure Description

[0030] Figure 1 A flowchart illustrating a high-throughput calculation method for catalytic materials based on machine learning, as proposed in an embodiment of this invention application;

[0031] Figure 2 This is a flowchart of the dual-channel linear and nonlinear feature fusion prediction process proposed in an embodiment of this invention.

[0032] Figure 3 This is a flowchart illustrating the sample selection process based on dynamic thresholds and boundary constraints, as proposed in an embodiment of this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figure 1 - Figure 3 The present invention provides a high-throughput calculation method for catalytic materials based on machine learning, the specific steps of which are as follows:

[0035] Historical simulation datasets are retrieved, and parameter values ​​are obtained using catalyst identifiers and condition parameters as indices. Known performance index values ​​are used as filler values ​​to construct a two-dimensional sparse matrix. Known data regions are marked as non-empty coordinates, and unknown regions are marked as empty value domains.

[0036] Based on non-empty coordinates and catalyst identifiers, sample feature attributes, sample identifiers and condition parameter codes are obtained by querying the embedding table, and then mapped through a multilayer perceptron network to generate a hybrid representation vector.

[0037] Linear and nonlinear features are fused into the hybrid representation vector using a neural collaborative filtering network and a fully connected neural network to generate performance prediction values ​​and prediction dispersion. Based on the performance prediction values, prediction dispersion, and hybrid representation vector, sample scores are obtained. Based on conditional parameters, samples with scores higher than a dynamic threshold are selected, and a list to be verified is output.

[0038] The simulation calculation model is used to calculate the list of tests to obtain performance data, which is then backfilled into a two-dimensional sparse matrix, and the non-empty coordinates are updated and the dynamic threshold is adjusted.

[0039] The determination is based on the termination condition. If the termination condition is not met, iterative prediction is performed on the non-empty coordinates; if the termination condition is met, a list of catalytic materials is output.

[0040] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0041] Example 1

[0042] This application discloses a high-throughput calculation method for catalytic materials based on machine learning. (See attached document.) Figure 1 The specific steps proposed in this invention include: S1. Retrieving historical simulation datasets, using catalyst identifiers and condition parameters as indices to obtain parameter values, and using known performance index values ​​as fill values ​​to establish a two-dimensional sparse matrix; marking known data regions as non-empty coordinates and unknown regions as empty value domains; S2. Based on non-empty coordinates and catalyst identifiers, obtaining sample feature attributes, sample identifiers, and condition parameter codes by querying an embedding table, and mapping them through a multilayer perceptron network to generate a hybrid representation vector; S3. Performing linear and nonlinear feature fusion on the hybrid representation vector through a neural collaborative filtering network and a fully connected neural network to generate performance prediction values ​​and prediction dispersion; S4. Obtaining sample scores based on performance prediction values, prediction dispersion, and hybrid representation vectors, and based on condition parameters, selecting samples with scores higher than a dynamic threshold, and outputting a list to be verified; S5. Calculating the list to be verified using a simulation calculation model to obtain performance data, backfilling it into the two-dimensional sparse matrix, updating the non-empty coordinates, and adjusting the dynamic threshold; S6. Judging based on termination conditions; if the termination conditions are not met, iteratively predicting the non-empty coordinates; if the termination conditions are met, outputting a list of catalytic materials.

[0043] Furthermore, historical simulation datasets are retrieved, and parameter values ​​are obtained using catalyst identifiers and condition parameters as indices. Known performance index values ​​are used as filler values ​​to construct a two-dimensional sparse matrix. Known data regions are marked as non-empty coordinates, and unknown regions are marked as empty value domains; this corresponds to step S1 above. The specific implementation process includes:

[0044] Historical simulation datasets are retrieved, and catalyst identifier sequences and condition parameter sequences are extracted. The catalyst identifier sequences are converted into row logical addresses, and the condition parameter sequences are converted into column logical addresses. These are then combined to generate physicochemical spatial positioning index pairs. A two-dimensional sparse matrix is ​​initialized, and based on the pointers of the physicochemical spatial positioning index pairs, known performance index values ​​from the historical simulation dataset are written into the corresponding storage units. The address space is traversed, and storage units containing valid values ​​are identified. Row and column position information is extracted to generate a non-empty coordinate index table. Blank storage units without valid values ​​are identified, and their distribution positions are recorded to generate a null value domain mask map.

[0045] Specifically, the historical simulation dataset is retrieved, and the catalyst identifier sequence and condition parameter sequence are extracted. The catalyst identifier sequence is converted into a row logical address, and the condition parameter sequence is converted into a column logical address, which are then combined to generate a physicochemical spatial location index pair. In this embodiment, the historical simulation dataset is accessed first. This dataset stores records of the catalytic performance of perovskite oxides accumulated through high-throughput experimental tests. Each record contains a complete material chemical formula (e.g., LaCoO3, BaO.5Sr0.5Co0.8FeO.2O3, etc.) as a unique catalyst identifier, and a corresponding description of the reaction environment (e.g., pH value, applied potential, crystal orientation) as condition parameters. First, a deduplication and cleaning operation is performed to extract M unique catalyst components to form the catalyst identifier sequence, and N different combinations of environmental conditions to form the condition parameter sequence. Subsequently, a hash mapping table is established, using lexicographical or atomic number priority rules to map each text-formatted catalyst identifier to a unique integer index i, with a value ranging from 0 to M-1; this is the row logical address. Similarly, each combination of conditional parameters is mapped to a unique integer index j, with a value ranging from 0 to N-1; this is the column logical address. For example, for the material LaNiO3 at pH 14 and a potential of 1.6V (relative to the standard hydrogen electrode RHE), if it is mapped to row 105 and column 3, the generated physicochemical spatial location index pair is (105, 3).

[0046] Specifically, for initializing the two-dimensional sparse matrix, based on the pointers of the physicochemical spatial positioning index pairs, the known performance index values ​​from the historical simulation dataset are written into the corresponding storage units. In this embodiment, storage space is allocated in memory to construct a two-dimensional sparse matrix with dimensions M multiplied by N. Since known experimental or computational data is extremely scarce relative to the vast chemical space in actual high-throughput screening scenarios (typically with sparsity exceeding 99%), the values ​​of all elements in the matrix are set to specific null values ​​(e.g., NaN, Not a Number, or set to negative one) during initialization to clearly distinguish the physical meaning of "unknown" and "zero value". Next, the processor traverses the historical dataset, reading the known performance index values ​​from each record. In this embodiment, this performance index is uniformly set as the overpotential of the oxygen evolution reaction (OER), in volts (V). Addressing is performed based on the aforementioned generated physicochemical spatial positioning index pairs (i, j), directly locating the storage unit in the i-th row and j-th column of the two-dimensional sparse matrix, and writing the corresponding overpotential value. For example, if historical data shows that the overpotential of LaNiO3 in an alkaline environment is 0.33V, then the value at matrix coordinates (105, 3) is overwritten as 0.33. This process transforms unstructured database records into a structured matrix form, providing a standard input format for subsequent matrix factorization-based machine learning algorithms. Simultaneously, a conflict detection mechanism is implemented; if multiple historical data points exist at the same coordinate (e.g., calculation results from different batches), their arithmetic mean is calculated as the final padding value to eliminate data noise.

[0047] Specifically, the address space is traversed to identify storage units containing valid values, and row and column position information is extracted to generate a non-empty coordinate index table. In this embodiment, after the matrix is ​​filled, in order to efficiently train the machine learning model, it is necessary to clearly distinguish which data is the "fact" used for training and which is the "unknown" to be predicted. The processor starts a full matrix scan program to traverse M multiplied by N storage units. At each coordinate point (u, v), it is determined whether the value stored in the unit is equal to the initialized null value marker. If the stored value is not NaN (for example, the stored value is 0.33), it is determined that the unit contains a valid value and belongs to the "known data area". Then, the row index u and column index v of the unit are extracted, formed into a tuple (u, v) and appended to a dynamic list structure, finally forming a non-empty coordinate index table. This index table essentially defines the training set sample space in the supervised learning process. Taking a matrix containing 5000 materials and 10 conditions as an example, if only 800 cells are filled with valid overpotential data, then the length of the non-empty coordinate index table is 800. The data output from this step, the non-empty coordinate index table, will be directly transmitted to the subsequent neural network module as the direct basis for extracting positive sample features and calculating the gradient of the loss function, ensuring that the model only learns real physical laws and avoids interference from blank noise.

[0048] In this embodiment, spectral clustering is performed on a two-dimensional sparse matrix to rearrange rows and columns, constructing local high-density data blocks and establishing a hierarchical index mapping to optimize the matrix topology. Specifically, after constructing the two-dimensional sparse matrix, a Laplacian matrix is ​​constructed using a spectral clustering algorithm. Eigenvectors of the matrix's row and column vectors are calculated, and the row and column orders of the two-dimensional sparse matrix are rearranged based on the similarity of the eigenvectors. The originally discrete numerical points are aggregated into several locally continuous high-density data blocks, forming a block-like diagonalized or strip-like distribution structure. Based on the rearranged topology of the two-dimensional sparse matrix, a hierarchical index table containing global block indexes and local unit indexes is constructed, transforming random access to the sparse matrix into continuous access to high-density blocks, which serve as the input structure for subsequent convolution operations. By aggregating highly correlated data through matrix rearrangement, the locality and signal-to-noise ratio of feature extraction are improved, invalid computations on low-value sparse regions are reduced, and data utilization is increased.

[0049] Specifically, for blank storage cells that do not contain valid values, their distribution locations are recorded to generate a null value domain mask. In this embodiment, the explicit definition of unknown regions is performed simultaneously with the extraction of non-null coordinates. When a traversal check finds that the value in a storage cell (x, y) is still an initialized null value, it is identified as a "blank storage cell," indicating that the material has not yet been experimentally verified or simulated under the corresponding conditions. In order to process these positions in parallel and automatically ignore them in subsequent matrix operations, a binary tensor with the same dimensions as the original two-dimensional sparse matrix (i.e., M multiplied by N) is generated, called the null value domain mask. In this mask, the position corresponding to the valid value coordinate (u, v) is assigned a value of 1 (or True), while the position corresponding to the blank storage cell (x, y) is assigned a value of 0 (or False). For example, if the coordinate (105, 4) is empty in the original matrix, then the value of (105, 4) in the mask is 0. The mask image serves as a crucial control signal input to subsequent loss function calculation nodes: when calculating the mean squared error, the predicted value matrix is ​​first multiplied element-wise with the mask image, thereby forcing the prediction errors at all unknown locations to be set to zero, preventing the model from fitting non-existent "null values". Furthermore, the set of coordinates with values ​​of 0 in the mask image directly constitutes the "exploration space," providing a candidate pool for generating the "validation list" in subsequent active learning strategies.

[0050] By constructing a two-dimensional sparse matrix and converting catalysts and conditional parameters into unique logical addresses, a structured index of the physicochemical space is achieved. This not only clearly delineates the known and unknown domains, avoiding redundant calculations on existing data, but also locks in unexplored blank areas through a null value mask map. This ensures that each historical simulation record can support subsequent predictions, maximizing the value of historical data and preventing the waste of storage and computing resources in meaningless or known areas.

[0051] Furthermore, based on non-empty coordinates and catalyst identifiers, the sample feature attributes, sample identifiers, and conditional parameter codes are obtained by querying the embedding table, and then mapped through a multilayer perceptron network to generate a hybrid representation vector; corresponding to step S2 above; the specific implementation process includes:

[0052] The non-empty coordinate index table is received, and the catalyst discrete index and conditional discrete index are extracted by parsing the row and column positioning addresses, respectively. The catalyst discrete index is mapped to a preset catalyst structure fingerprint matrix, and the electronic structure feature vector is extracted as the sample identifier. The conditional discrete index is mapped to an environment embedding matrix, and the state feature vector is extracted as the conditional parameter encoding. The corresponding attribute value sequence is retrieved from the associated attribute database as the sample feature attribute, and concatenated with the sample identifier and the conditional parameter encoding execution channel to generate an aggregated input tensor. The aggregated input tensor is transmitted to a multilayer perceptron network, and weighted summation and nonlinear activation function transformation are performed through multi-level neuron nodes to obtain a feature value sequence, which serves as a hybrid representation vector containing material-environment coupling features. The sample feature attributes include electronegativity, atomic radius, and valence electron number.

[0053] Specifically, upon receiving the non-empty coordinate index table, the catalyst discrete index and conditional discrete index are extracted by parsing the row and column positioning addresses, respectively. In this embodiment, upon receiving the non-empty coordinate index table, the processor starts a parallel parsing thread to decouple the coordinates (u, v) of each storage unit. The row positioning address u is extracted and defined as the catalyst discrete index, which uniquely points to a specific material composition; for example, index value 1024 corresponds to the perovskite material LaNiO3. The column positioning address v is extracted and defined as the conditional discrete index, which uniquely points to a specific set of reaction environment parameters; for example, index value 7 corresponds to the combination of an alkaline electrolyte environment with a pH of 14 and the (110) crystal plane orientation. To improve throughput, index parsing adopts a GPU parallel computing mode based on the CUDA kernel, which can complete the address separation and index extraction of hundreds of thousands of coordinate points within milliseconds, laying the foundation for feature mapping of massive data.

[0054] Specifically, the catalyst discrete index is mapped to a pre-set catalyst structure fingerprint matrix, and the electronic structure feature vector is extracted as the sample identifier. In this embodiment, a catalyst structure fingerprint matrix with dimension M×D1 is pre-constructed, where M represents the total number of candidate catalysts (e.g., 5000), and D1 represents the dimension of the feature vector (e.g., 32-dimensional). When the catalyst discrete index u is input, a lookup operation is performed to directly locate the u-th row of the matrix and read the corresponding row vector. This row vector is not a random number, but an electronic structure feature vector generated using the Coulomb matrix based on material crystallographic data, which compresses and characterizes the basic topological structure and composition information of the material. Taking LaNiO3 as an example, this vector implicitly contains structural fingerprint information such as the space occupancy ratio of lanthanum elements at A-site and nickel elements at B-site, the degree of lattice distortion, and the connection mode of oxygen octahedra. This vector is defined as the sample identifier, serving as a "digital ID card" for the neural network to identify different individual materials.

[0055] Specifically, the conditional discrete index is mapped to the environment embedding matrix, and the state feature vector is extracted as the conditional parameter encoding. In this embodiment, an environment embedding matrix with dimension N×D2 is constructed to address the diversity of reaction environments, where N is the total number of condition combinations (e.g., 9) and D2 is the embedding dimension (e.g., 16 dimensions). Unlike fixed atomic properties, the parameters of this matrix are dynamically updated entirely through a backpropagation algorithm. When the conditional discrete index v is input, the vector in the v-th row is extracted as the conditional parameter encoding. For example, for index 7 (pH=14, (110) crystal plane), the extracted 16-dimensional vector can implicitly characterize the chemisorption potential field of high-concentration hydroxide ions on the surface and the atomic arrangement density of a specific crystal plane in a high-dimensional space. Through this embedding technique, the originally discrete and discontinuous physical conditions (such as acidity, neutrality, and alkalinity) are transformed into a continuous and differentiable dense vector space, enabling the neural network to capture the potential correlations and transition patterns between different environmental conditions.

[0056] Specifically, the corresponding attribute value sequence is retrieved from the associated attribute database as the sample feature attribute, and concatenated with the sample identifier and the condition parameter encoding through the execution channel to generate an aggregated input tensor. In this embodiment, to enhance the interpretability and physical constraints of the model, explicit physicochemical parameters matching the current catalyst index are precisely retrieved from the associated attribute database. Specifically, for LaNiO3, the Pauling electronegativity (1.91), ionic radius (0.69 Å), and valence electron number (7) of its B-site nickel are extracted. These scalar data are standardized (normalized to the 0-1 range) to form an attribute value sequence of length D_3 (e.g., 3). Subsequently, the processor performs a channel concatenation operation, stacking the aforementioned 32-dimensional sample identifier, 16-dimensional condition parameter encoding, and 3-dimensional sample feature attribute along the feature channel dimension to generate an aggregated input tensor with a total dimension of 32+16+3=51. This tensor fully integrates the three-in-one information of "material identity, environmental state, and physical essence," providing holographic data input for subsequent nonlinear mapping.

[0057] Specifically, the aggregated input tensor is transmitted to a multilayer perceptron network. Weighted summation and nonlinear activation function transformations are performed through multi-level neuron nodes to obtain a feature value sequence, which serves as a hybrid representation vector containing material-environment coupling features. In this embodiment, the aggregated input tensor is fed into a multilayer perceptron network containing an input layer, two hidden layers, and an output layer. The first hidden layer has 128 neurons. Matrix multiplication is performed between the input tensor and a fully connected weight matrix, and a bias vector is superimposed. This is then processed by a ReLU (Rectified Linear Unit) activation function to remove negative noise and introduce nonlinear features. The second hidden layer has 64 neurons, which further compress the feature space and extract higher-order combined features to generate the final output sequence. In each layer's calculation, the weight parameters are continuously optimized using gradient descent to minimize prediction error. The final output 64-dimensional feature value sequence is the hybrid representation vector, where the first 32 bits are defined as the component feature mapping region, and the last 32 bits are defined as the environment feature mapping region, together forming the hybrid representation vector. This vector is a mathematical representation generated after being "digested" by a deep neural network. It profoundly encodes complex material-environment coupling characteristics such as the "synergistic effect of high-valence nickel ions with oxygen vacancies in an alkaline environment," providing a high-density information carrier for accurate prediction of catalytic performance.

[0058] In this embodiment, a channel attention-based weight redistribution mechanism is performed on the hybrid representation vector, and a soft orthogonalization constraint is applied to output a deredundant feature vector. Specifically, the generated hybrid representation vector is recalibrated, a compressed excitation network is used to calculate the global descriptor of the feature channels, and channel weight coefficients are generated through fully connected neuron layers. The original hybrid representation vector is then weighted channel-by-channel to suppress low-information feature responses. Simultaneously, a soft orthogonalization regularization term is introduced into the loss function, and singular value decomposition is performed on the feature vector matrix to constrain the inner product of feature vectors of different dimensions to approach zero, forcing feature decoupling. This process eliminates redundant coupling information at the feature input stage, outputting a highly significant orthogonalized feature vector. This enhances the model's ability to identify key features of high-value samples, reduces misjudgments of low-value samples due to feature redundancy, and avoids wasting subsequent simulation resources.

[0059] A hybrid representation generation technique based on embedding tables and multilayer perceptrons was adopted to map discrete atom and environment indices into high-dimensional dense vectors containing physicochemical connotations, thereby enhancing the feature representation capability of the input data. This means that the model can learn more accurate structure-property relationships with fewer samples, improving the utilization efficiency of simulation data in a single run. By capturing deep non-second-order cross features, the model can more accurately predict the value of unknown samples, thus reducing misjudgments of low-value samples and ineffective trial and error caused by unclear feature representation at the source.

[0060] Furthermore, linear and nonlinear features are fused onto the hybrid representation vector using a neural collaborative filtering network and a fully connected neural network to generate performance prediction values ​​and prediction dispersion; this corresponds to step S3 above; see [link to relevant documentation]. Figure 2 The specific implementation process includes:

[0061] The hybrid representation vector is transmitted in parallel to the processing branches of the neural collaborative filtering network and the fully connected neural network. In the neural collaborative filtering network processing branch, element-wise multiplication of the feature dimensions is performed on the hybrid representation vector to analyze the second-order cross-features of the catalyst components and the reaction environment within the feature vector, and outputs a cross-interaction feature tensor. In the fully connected neural network processing branch, the weight matrix of the multilayer neurons is driven to perform continuous matrix multiplication and nonlinear activation function transformation on the hybrid representation vector to reconstruct the nonlinear spatial structure and output a nonlinear activation barrier tensor. The cross-interaction feature tensor and the nonlinear activation barrier tensor are concatenated to construct a fused feature vector, and the weighted sum is calculated as the performance prediction value. The dispersion of the feature distribution analyzed by the variance evaluation node is used as the prediction dispersion.

[0062] Specifically, in the neural collaborative filtering network processing branch, element-wise multiplication of the feature dimensions is performed on the hybrid representation vector to analyze the second-order cross-features of the catalyst components and the reaction environment within the feature vector, and output the cross-interaction feature tensor. In this embodiment, the hybrid representation vector (e.g., a floating-point vector with dimension 64) is first imported into the dedicated processing channel of the neural collaborative filtering network processing branch. In this branch, the processor performs vector slicing, extracting the first 32 dimensions as component feature sub-vectors and the last 32 dimensions as environment feature sub-vectors based on a preset logical index. Due to the use of block output constraints, these two sub-vectors maintain relatively independent physical properties, and both have a dimension of 32. Subsequently, the tensor processing unit is called to perform the Hadamard product operation, i.e., element-wise multiplication. Specifically, the first element of the component feature sub-vector is multiplied by the first element of the environment feature sub-vector, and so on, until the multiplication operation of all corresponding dimensions is completed. This mathematical process physically simulates the direct proportional interaction between the catalyst active site and the electrolyte environment, such as the linear proportional relationship between the d-band center of metal ions and the adsorption energy of hydroxide ions. Through this explicit cross-operation, the model can efficiently capture low-order feature combination effects, and the generated output is called the cross-interaction feature tensor. This tensor retains the dimensional information of the original feature space, and its numerical value directly reflects the degree of matching of a specific material under a specific environment (i.e., the affinity strength of active sites), providing a physically intuition-based linear benchmark for subsequent performance prediction.

[0063] Specifically, in the fully connected neural network processing branch, the driving multilayer neuron weight matrix performs continuous matrix multiplication and nonlinear activation function transformation on the hybrid representation vector to reconstruct the nonlinear spatial structure and output a nonlinear activation barrier tensor. In this embodiment, the fully connected neural network branch, which works in parallel with the neural collaborative filtering branch, aims to mine the high-order complex patterns hidden in the data. The complete hybrid representation vector is input into a deep multilayer perceptron structure. This network is configured with a "tower" structure, containing three hidden layers with the number of neuron nodes set to 128, 64, and 32 respectively. In each layer, the input vector is first multiplied by a randomly initialized (and subsequently optimized by backpropagation) weight matrix, and a bias term is added. Subsequently, the calculation result is transformed by a nonlinear activation function. In this embodiment, a linear rectifier unit with leakage is selected to avoid the "death" of neurons in the negative interval while preserving the gradient in the positive interval. This series of nonlinear transformations aims to reconstruct the complex potential energy surface in the catalytic reaction, that is, to simulate the nonlinear activity changes (e.g., the vertex region of the volcano-shaped curve relationship) caused by electron spin state flipping, lattice distortion, or solvation effects during the reaction of a real catalyst. After multiple layers of abstraction and feature recombination, the 32-dimensional vector output by the last layer is the nonlinear activation barrier tensor, which encodes the intrinsic ability of a material to overcome the reaction barrier on a high-dimensional manifold.

[0064] Specifically, the cross-interaction feature tensor and the nonlinear activation barrier tensor are concatenated to construct a fused feature vector, and the weighted sum is calculated as the performance prediction value. The dispersion of the feature distribution is analyzed by the variance evaluation node as the prediction dispersion. In this embodiment, a feature fusion operation is performed to comprehensively utilize the linear matching law and the nonlinear structure-property relationship. The processor concatenates the cross-interaction feature tensor output from the NCF branch and the nonlinear activation barrier tensor output from the FCNN branch in the channel dimension to generate a fused feature vector with a dimension of 64. This vector is then transmitted to the output layer, which contains two independent prediction heads. The first prediction head is a regression node that uses a fully connected layer to calculate the weighted sum of the fused vector and directly outputs a scalar value, which is the predicted oxygen evolution reaction overpotential (in V). For example, the predicted overpotential of a certain double perovskite material at a current density of 10 mA / cm² is 0.35 V. The second prediction head is a variance evaluation node, which is designed based on the variance estimation mechanism of Bayesian neural network principles. This node does not directly predict performance, but rather analyzes the sparsity and dispersion of the fused feature vectors in the latent space, outputting a non-negative scalar as the prediction dispersion (i.e., uncertainty σ). 2If the input sample is located in a dense region of the training data, the feature distribution is compact, and the output dispersion value is small (e.g., 0.001); conversely, if the sample is located in an unknown region of chemical space, the feature distribution is divergent, and the dispersion value increases significantly (e.g., 0.05). These two output values, the performance prediction value and the prediction dispersion, together constitute a complete evaluation of the candidate material. This variance evaluation node is trained to output the log-variance of the prediction value. The training uses a Gaussian negative log-likelihood loss function, which is equal to the sum of the weighted squared error term (whose weight is determined by the log-variance) and the prediction log-variance (as a regularization term). By minimizing this loss function, the network can simultaneously learn the uncertainty (arbitrary uncertainty) of the prediction value and the data itself.

[0065] By employing a dual-parallel network architecture, a fully connected network reconstructs the nonlinear energy barrier, while neural collaborative filtering captures linear interactions. This combination not only improves prediction accuracy but, more importantly, outputs prediction dispersion through variance evaluation nodes. This metric directly quantifies the uncertainty of the samples, acting as a detector to help identify samples with high prediction values ​​but potentially redundant or repetitive patterns, as well as samples that are full of unknowns but highly valuable for exploration. This guides computing power to avoid redundant regions with low information gain and focus on information-rich areas.

[0066] Further, based on the performance prediction value, prediction dispersion, and mixed representation vector, sample scores are obtained. Based on conditional parameters, samples with scores higher than a dynamic threshold are selected, and a list of samples to be verified is output; this corresponds to step S4 above; see [link / reference]. Figure 3 The specific implementation process includes:

[0067] The algorithm receives the predicted performance value and predicted dispersion, and retrieves the hybrid representation vector. It maps the hybrid representation vector to the latent feature space, calculates the manifold distance between the hybrid representation vector and the sample points in the non-empty coordinate set, and calls the kernel density estimation algorithm to generate the null domain distribution density. It constructs an exploration-utilization balance evaluation model, receiving the predicted performance value, predicted dispersion, and null domain distribution density as input variables, and performs a nonlinear weighted fusion operation to generate sample scores. It retrieves the dynamic threshold value for the current iteration period, performs a numerical comparison operation between the sample scores and the dynamic threshold, and filters out a subset of candidate samples whose scores exceed the dynamic threshold. It parses the conditional parameters associated with the candidate sample subset, performs physicochemical parameter logical verification based on preset experimental boundary constraints, and eliminates invalid samples that do not meet the experimental boundary constraints. Finally, it assembles the passed sample indices and corresponding parameter codes into a list to be verified.

[0068] Specifically, for the predicted value and predicted dispersion of the receiving performance, a hybrid representation vector is retrieved; the hybrid representation vector is mapped to the feature latent space, the manifold distance between the hybrid representation vector and the sample points in the non-empty coordinate set is calculated, and the kernel density estimation algorithm is called to generate the null domain distribution density. In this embodiment, during the iterative process of active learning, three key data streams for the currently traversed sample are first received from the upstream neural network module: the predicted oxygen evolution reaction overpotential value (e.g., 0.35V), the model uncertainty variance (e.g., 0.02), and a 64-dimensional hybrid representation vector. To quantify the novelty of the current sample in the chemical space, the processor projects this 64-dimensional vector into a feature latent space optimized by the t-SNE algorithm. This structure can maintain local neighborhood relationships, and the dimension of the latent space is strictly limited to 2 to 4 dimensions to ensure the efficiency of distance calculation. Subsequently, a full-to-full distance calculation is performed, using Euclidean metric to calculate the manifold distance between the test sample vector and all known training sample vectors in the non-empty coordinate set. To reduce computational complexity, a KD-Tree data structure is used to accelerate nearest neighbor search, selecting only the K nearest neighbors (e.g., setting K to 10). Based on these manifold distances, a kernel density estimation algorithm is invoked. In this embodiment, a Gaussian kernel function is used, with its core hyperparameter bandwidth set to 1.0. The kernel function values ​​of all neighbors are weighted and accumulated, outputting a scalar between 0 and 1, which is the null domain distribution density. If this density value is lower than a preset sparsity threshold (e.g., 0.001), it indicates that the sample is located in an unknown region of the existing knowledge base, belonging to a potential new material region with high exploration value; conversely, if the density value is close to 1, it indicates that the region has been fully explored and has high information redundancy. To ensure iterative efficiency, this projection mapping process is implemented through a pre-trained parametric encoder network, which fits the nonlinear mapping relationship of t-SNE, thereby enabling rapid mapping of new sample vectors to the existing low-dimensional space. Subsequently, the manifold distance between the test sample vector and all known training sample vectors in the non-empty coordinate set is calculated using Euclidean metric. The parametric encoder network is a four-layer fully connected neural network (with a structure of 64-32-16-2). Its training process is as follows: select the mixed representation vector of all samples in the historical dataset, first execute the standard t-SNE algorithm to obtain the target low-dimensional coordinates; then use the mixed representation vector as input, the low-dimensional coordinates generated by t-SNE as labels, and use the mean squared error as the loss function to train the parametric encoder network, thereby giving it the ability to directly perform low-dimensional mapping on new input samples.

[0069] Specifically, to construct an exploration-utilization balance evaluation model, the performance prediction value, the prediction dispersion, and the null value domain distribution density are received as input variables. A nonlinear weighted fusion operation is performed to generate sample scores. In this embodiment, a nonlinear weighted fusion equation, namely the exploration-utilization balance evaluation model, is constructed. This equation contains three core indicators: the first is the utilization term, which is the negative of the normalized performance prediction value, aiming to screen out materials with the best predicted performance (i.e., the lowest overpotential); the second is the exploration term, which is the square root of the prediction dispersion, aiming to give more attention to samples with high model uncertainty; the third is the novelty term, which is the reciprocal of the null value domain distribution density, aiming to reward samples located in the sparse chemical space. These three indicators are fused through adjustable weight coefficients. For example, in the early stages of iteration, the exploration weight and novelty weight are set to 2.0, and the utilization weight is set to 1.0 to encourage the model to cast a wide net; as the number of iterations increases, the utilization weight gradually increases to 5.0, and the exploration weight decreases to 0.5, thereby causing the search focus to converge to the high-performance region. In the specific calculation, the input variables are non-linearly scaled using the Sigmoid function to prevent any single indicator value from dominating the overall score. The final output sample score is a dimensionless real number, such as 85.5 points, which comprehensively reflects the overall value of a material in the three dimensions of "good performance", "high uncertainty" and "sufficient novelty".

[0070] Specifically, the dynamic threshold value of the current iteration cycle is retrieved, and the sample score is compared with the dynamic threshold to filter out a subset of candidate samples whose scores exceed the dynamic threshold. The conditional parameters associated with the candidate sample subset are parsed, and physicochemical parameter logic verification is performed based on preset experimental boundary constraints to eliminate invalid samples that do not meet the experimental boundary constraints. The passed sample indices and corresponding parameter codes are assembled into a list to be verified. In this embodiment, the dynamic threshold in the global state register is read. This threshold is dynamically generated based on the 95th percentile of the sample score distribution in the previous iteration. Assuming the current threshold is 80.0, all calculated sample scores are compared with this value, and only samples with scores greater than 80.0 are retained in the candidate pool to obtain a subset of candidate samples. Subsequently, the physical validity verification stage begins. The chemical formulas and reaction conditions of the candidate samples are reverse-analyzed, and a strict logical check, i.e., physicochemical parameter logic verification, is performed. First, the crystal structure stability is verified by calculating the Goldschmidt tolerance factor t, which must fall within the range of 0.80 to 1.00, i.e., within the set experimental boundary constraints, to ensure the geometric stability of the perovskite structure. Second, the charge conservation is verified by using a pre-set database of common oxidation states (e.g., Fe at +3 or +4 valence) to verify whether the total valence state of the A-site and B-site cations can be balanced with the oxygen anion, eliminating combinations that are not charge neutral. Finally, the Pourbaix stability is verified by querying the decomposition energy of the material at pH 14 and a potential of 1.6V, eliminating unstable phases with a decomposition energy greater than 0.2eV. After this screening, approximately 50 to 100 samples that pass the verification are packaged, and their corresponding material IDs (e.g., ID2048) and condition codes are formatted into a JSON file as the final list to be verified.

[0071] A screening mechanism based on dynamic threshold and kernel density estimation was established. By calculating manifold distance and null domain density, the relative value of a sample in the current iteration stage can be dynamically evaluated, and redundant samples in high-density known regions are eliminated. Combined with experimental boundary constraint verification, invalid samples that are not experimentally feasible are forcibly filtered out before simulation. This dual filtering mechanism acts like an intelligent funnel, strictly intercepting low-value and invalid candidates, ensuring that every sample entering the simulation stage is a rigorously selected target, and improving the conversion rate of computational resources.

[0072] Furthermore, the performance data is obtained by calculating the list to be verified using a simulation model, backfilled into a two-dimensional sparse matrix, and the non-empty coordinates are updated and the dynamic threshold is adjusted; this corresponds to step S5 above; the specific implementation process includes:

[0073] The system performs format conversion on the encoded condition parameters in the verification list, generates an input control file stream, distributes it to the computing nodes, performs first-principles simulation operations through the simulation model, monitors the convergence status of the operation, and captures the output log data stream. It then parses the output log data stream to extract scalar values ​​characterizing material activity as performance data. Using the data storage manager, it locks the storage address of the corresponding empty value domain in the two-dimensional sparse matrix and writes the performance data to the corresponding storage cell. It scans the storage cell location of the newly written data, extracts its row and column positioning information, appends this information to the non-empty coordinate index table, and updates the non-empty coordinate set. Finally, it retrieves the performance data set and performance prediction value set corresponding to the updated non-empty coordinates, calculates the distribution gradient of the prediction error, corrects the boundary parameters of the sampling strategy based on the distribution gradient, and outputs the updated dynamic threshold value.

[0074] The system receives a list of samples to be verified, calls a preset crystal structure template based on the sample identifier, and adjusts the lattice parameters according to the condition parameters to generate spatial geometric topology data characterizing the microstructure of the material. Based on the spatial geometric topology data, a discretized numerical computation grid is constructed, and boundary constraints are set in conjunction with preset environmental parameters to generate a simulation input file. The simulation input file is loaded into the iterative solver, the system state vector is initialized, and the self-consistent field cyclic computation engine is driven to perform nonlinear variational solutions to the control equations. During the solution process, the convergence residual of the energy functional is calculated in real time. When the convergence residual falls within a preset tolerance range, the calculation is considered converged, and the system steady state is locked. The post-processing extraction module is triggered to analyze the electronic structure data under steady state and calculate scalar values ​​characterizing the catalytic activity of the material as performance data.

[0075] Specifically, for the received list of samples to be verified, a preset crystal structure template is invoked based on the sample identifier, and the lattice parameters are adjusted according to the condition parameters to generate spatial geometric topology data characterizing the microstructure of the material; a discretized numerical calculation grid is constructed based on the spatial geometric topology data, and boundary constraints are set in combination with preset environmental parameters to generate a simulation input file; in this embodiment, the list of samples to be verified is parsed, and the highest priority sample is extracted, such as the perovskite catalyst numbered ID-2048, whose chemical formula is LaFeO.5NiO.5O3, and the corresponding environmental condition code is alkaline electrolyte (pH equals 14). The crystal structure generator is invoked, and a standard orthorhombic Pnma space group perovskite unit cell template is loaded. Based on the compositional characteristics of LaFeO.5NiO.5O3, the processor performs a supercell expansion operation to construct a 2x2x1 supercell containing 40 atoms. A special quasi-random structure algorithm is used to randomly assign Fe and Ni atoms to B-site metal nodes to simulate the disordered distribution of a real solid solution. The lattice constant is initially adjusted based on Shannon ion radius data and Vigarde's law, generating initial spatial geometric topology data (i.e., the atomic coordinate file POSCAR). Subsequently, based on the crystal orientation index (e.g., the 110 crystal plane) in the condition parameters, the crystal is sliced ​​along this direction to construct a flat plate model with a thickness of four atomic layers. A 15 Å thick vacuum layer is added in the Z-axis direction to eliminate periodic mirror interactions. Based on this, a discretized numerical computation grid is constructed, and the plane wave cutoff energy is set to 500 eV. Based on preset environmental parameters, a series of simulation input files are automatically generated: In the input parameter file, the generalized gradient approximation functional is set as the exchange correlation functional, spin polarization is enabled (ISPIN equals 2), and Hubbard U correction parameters are set for strongly correlated electron systems, with the effective U value of iron set to 4.0 eV and the effective U value of nickel set to 6.4 eV to describe the electron localization effect of 3d orbitals; at the same time, an implicit solvation model is configured, and the dielectric constant is set to 78.4 to simulate the boundary conditions of the aqueous environment.

[0076] Specifically, the simulation input file is loaded into the iterative solver, the system state vector is initialized, and the self-consistent field cyclic computation engine is driven to perform nonlinear variational solution of the governing equations. During the solution process, the convergence residual of the energy functional is calculated in real time. When the convergence residual falls within the preset tolerance range, the calculation is considered converged and the system steady state is locked. In this embodiment, the generated input parameter file, atomic coordinate file, pseudopotential file, and K-point grid setting file are distributed to the computation nodes. The iterative solver first initializes the system wavefunction state vector, usually using a random wavefunction or superimposed atomic charge density as the initial guess. Subsequently, the self-consistent field cyclic computation engine of the VASP kernel is started, and the Cohen-Shen Lüjiu equation is solved nonlinearly using the RMM-DIIS algorithm. In each electron step iteration, the engine calculates the total energy and Hamiltonian eigenvalues ​​of the system, and monitors in real time whether the convergence residual of the energy functional falls within the tolerance range. The set electron step convergence tolerance is 1 x 10^-5 eV. When the energy difference between two consecutive iterations is less than the threshold, the electronic structure is considered to have reached the ground state. Then, ion-step relaxation is initiated, calculating the Hermann-Feynman force acting on each atom. The atom positions are adjusted using the conjugate gradient method until the force on all atoms is less than 0.02 eV per angstrom. When both the electronic and ion-step convergence conditions are met simultaneously, the calculation is considered to have fully converged, the current steady-state geometry and electron cloud density distribution are locked, and the calculation process is terminated.

[0077] Specifically, the post-processing extraction module analyzes the steady-state electronic structure data and calculates scalar values ​​characterizing the catalytic activity of the material as performance data. The storage address corresponding to the empty domain in the two-dimensional sparse matrix is ​​locked by the data storage manager, and the performance data is written to the corresponding storage unit. In this embodiment, after the calculation task is completed, the processing script automatically mounts the output directory and parses the log file data stream. To obtain scalar values ​​characterizing the oxygen evolution reaction (OER) activity, the Gibbs free energy of four steady-state configurations—pure surface, adsorbed OH groups, adsorbed O groups, and adsorbed OOH groups—is extracted. The free energy of each intermediate state is calculated using the formula G=E+ZPE-TS, where E is the total energy calculated by DFT, ZPE is the zero-point energy correction, and TS is the entropy change contribution. Based on the four-electron reaction mechanism, the free energy change of the rate-determining step in the reaction pathway is calculated. G max Then, using the formula η= G maxThe performance data of the material, i.e., the theoretical overpotential value, is calculated using / e-1.23V, for example, 0.325V. At this point, the data storage manager regenerates the physicochemical spatial positioning index pair based on the original material identifier (ID-2048) and condition parameter (pH=14), precisely locks the corresponding empty domain coordinates (i, j) in the two-dimensional sparse matrix, and writes the calculated scalar value 0.325 into the storage unit, completing the mapping from the physical simulation space to the digital feature space.

[0078] In this embodiment, the energy convergence curve and atomic force residual of the electronic iteration step associated with the performance data can be analyzed to perform physical convergence logic verification and eliminate pseudo-converged data. Specifically: after acquiring the performance data, the self-consistent field loop record in the simulation calculation model log is retrieved, the second derivative of the energy change in the electronic iteration step is calculated, and the presence of high-frequency oscillation characteristics is identified; the maximum atomic force component of the final ion relaxation step is extracted, and it is determined whether it is strictly lower than the preset convergence tolerance. If a pseudo-convergence state is detected where the energy oscillation has not decayed or the force residual exceeds the tolerance, the abnormal circuit breaker logic is triggered, the performance data is marked as invalid noise, and removed from the list to be verified. Performance data is only allowed to enter the subsequent backfilling and model update process when the physical convergence constraint is met. This prevents low-value noise data with pseudo-convergence from entering the iteration closed loop, prevents model parameters from being updated in the wrong direction, and ensures that computing resources are always used for the effective iteration of high-value samples.

[0079] Specifically, for the storage unit location of newly written data, its row and column positioning information is extracted and appended to the non-empty coordinate index table to update the non-empty coordinate set. The updated performance data set and performance prediction value set corresponding to the non-empty coordinates are retrieved, the distribution gradient of the prediction error is calculated, the boundary parameters of the sampling strategy are corrected based on the distribution gradient, and the updated dynamic threshold value is output. In this embodiment, in the synchronous callback function triggered by data backfilling, the process of identifying the matrix coordinates (i, j) of the newly written value and updating the non-empty coordinate set involves: flipping the state flag of the matrix coordinates to "non-empty" and appending it to the global non-empty coordinate index table. Next, the true performance data set (DFT calculated value) of all new validation samples in this iteration and the previous machine learning performance prediction value set are retrieved. The mean absolute error and root mean square error of both are calculated, and combined with the error statistics of the previous iteration, the distribution gradient of the prediction error is calculated. For example, if it is found that the prediction error of high-scoring samples in this round has decreased by 20% compared to the previous round, it indicates that the confidence of the model in the high-performance region has improved (the gradient is negative). Based on this distribution gradient, the boundary parameters of the sampling strategy are adjusted: when the error gradient indicates that the model is becoming more accurate, the dynamic threshold is increased, for example, from the top 5% of predicted scores to the top 2%, thus adopting a more aggressive "exploitation" strategy in the next round of screening, calculating only samples with a very high probability of becoming the optimal solution; conversely, if the error gradient fluctuates significantly, the threshold is lowered and the exploration weight is increased to introduce more diverse samples to correct model bias. Finally, the updated dynamic threshold value is output to guide the truncation of the next round of the list to be validated.

[0080] By employing a real-time feedback update strategy based on physical simulation, the model automatically schedules first-principles calculations to obtain the true performance values ​​and backfills them into the matrix to correct the non-empty coordinate set. Through real-time calculation of the prediction error gradient, the model can quickly detect its own cognitive biases and adjust its sampling strategy. Once it detects that the current direction is low-value, it can immediately stop the loss and adjust the dynamic threshold to shift to a new region. This rapid closed-loop correction mechanism ensures that the model always evolves towards high-value regions, avoiding a large number of ineffective and redundant iterations within incorrect physical and chemical spaces.

[0081] Furthermore, a judgment is made based on the termination condition. If the termination condition is not met, iterative prediction is performed on the non-empty coordinates; if the termination condition is met, a list of catalytic materials is output; this corresponds to step S6 above; the specific implementation process includes:

[0082] The system receives the updated two-dimensional sparse matrix and retrieves the performance data distribution statistics from the most recent consecutive iterations. It performs convergence evaluation, calculating the decay rate of performance prediction error and the gradient of high-scoring sample discovery rate as system state indicators. These indicators are then mapped to a pre-defined termination judgment logic controller. If the judgment result indicates the model is in an oscillation period, an iterative control signal is generated, defining the updated non-empty coordinate set as a new seed dataset. This dataset is then transmitted to the embedded table query interface via a feedback loop, driving the multilayer perceptron network to perform a new round of value evaluation on the remaining empty value domain and correct the sampling strategy. If the judgment result indicates the model is in a steady-state convergence period, the result output generator is activated. Based on the final dynamic threshold standard, a full scan is performed on all non-empty coordinates in the two-dimensional sparse matrix to extract target units with satisfactory performance values. The sample identifiers associated with the target units are parsed, and a list of catalytic materials containing material composition and structural characteristics is constructed.

[0083] Specifically, for the received updated two-dimensional sparse matrix, the performance data distribution statistics of the most recent consecutive iterations are retrieved; convergence evaluation is performed to calculate the decay rate of performance prediction error and the gradient of high-scoring sample discovery rate as system state indicators; in this embodiment, at the end of each round of active learning, the current global two-dimensional sparse matrix is ​​locked, which has been updated with the latest simulation data. The most recent three iterations (denoted as T, T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15, T16, T17, T18, T19, T10 ... -1 T -2 The system records historical data. For each cycle, two core statistics are extracted: first, the "validation set error," which is the average absolute error between the true DFT values ​​of all newly calculated samples in that cycle and the model's predicted values; second, the "high-scoring sample count," which is the number of newly discovered high-quality materials with overpotentials below a specific threshold (e.g., 0.35V) in that cycle. Based on this fundamental data, the processor performs convergence evaluation calculations: first, it calculates the decay rate of the performance prediction error, expressed as: the error of the previous cycle minus the error of the current cycle, then divided by the error of the previous cycle. If this rate is positive and large, it indicates that the model is learning rapidly; if it is close to zero, it indicates that the model's ability is approaching saturation. Simultaneously, it calculates the gradient of the high-scoring sample discovery rate, which is the difference between the number of high-quality samples discovered in the current cycle and the previous cycle. These two calculated scalar values—error decay rate and discovery rate gradient—are combined and encapsulated into a system state index vector, serving as a quantitative basis for determining whether the entire high-throughput screening process should continue.

[0084] Specifically, the system state indicators are mapped to a preset termination decision logic controller. If the decision result indicates that the model state is in an oscillation period, an iterative control signal is generated, the updated non-empty coordinate set is defined as a new seed dataset, and transmitted to the embedded table query interface through a feedback loop. This drives the multilayer perceptron network to perform a new round of value evaluation on the remaining empty value domain and correct the sampling strategy. In this embodiment, the termination decision logic controller has a built-in rule-based decision tree. When the input system state indicators show that the error decay rate is negative (i.e., error bounce) or fluctuates significantly near zero, and the high-scoring sample discovery rate is unstable, the controller determines that it has entered an "oscillation period." This usually means that the current sampling strategy has fallen into a local minimum, or the model has not yet mastered the physical laws of a certain type of complex material. At this time, the controller generates a high-level signal for "continue iteration" and triggers a strategy correction subroutine. All non-empty coordinate sets (which may contain thousands of data points) that have been filled with real data in the current matrix are packaged and defined as a new full seed dataset. This dataset is transmitted back to the front end through the feedback bus, triggering the model's "full retraining" or "fine-tuning" mode, driving the multilayer perceptron (MLP) network to relearn the feature mapping weights. Meanwhile, the sampling strategy parameters were modified: the exploration weight coefficient λ in the sampling function was automatically increased, for example, from 0.5 to 2.0, forcing the model to give higher scores to those null samples with larger prediction variance (uncertainty) in the next round of evaluation, so as to explore the unknown areas in the chemical space that have not yet been explored.

[0085] Specifically, if the determination result indicates that the model is in a steady-state convergence period, the result output generator is activated. Based on the final dynamic threshold standard, a full scan is performed on all non-empty coordinates in the two-dimensional sparse matrix to extract target cells with satisfactory performance values. In this embodiment, when the logic controller detects that the decay rate of the prediction error is lower than the preset convergence tolerance (e.g., 5%) for three consecutive cycles, and the gradient of the high-scoring sample discovery rate approaches zero (meaning it is difficult to find better materials), the model is determined to have entered a "steady-state convergence period." At this time, a termination command is issued, the iteration loop is cut off, and the result output generator is activated. This generator no longer makes predictions but performs deterministic data mining on the final state of the two-dimensional sparse matrix. A final dynamic threshold standard is set, for example, filtering records with an overpotential less than 0.32V and a model prediction confidence higher than 90%. The processor performs a full scan on all non-empty coordinates in the matrix (including both historical data and data calculated and verified during iteration), logically filters out all storage cells that meet the conditions, and marks them as "target cells." This step ensures that the final output is not only a theoretical prediction, but also includes data rigorously validated by the DFT, or a prediction at a high confidence level.

[0086] Specifically, for each target unit associated with a sample identifier, a list of catalytic materials containing material composition and structural features is constructed. In this embodiment, a reverse analysis operation is performed for each selected target unit. Based on its row index i in the sparse matrix, a reverse lookup of the hash mapping table is performed to recover the corresponding catalyst chemical formula (e.g., LaFeO.5NiO.5O3); based on the column index j, the corresponding reaction conditions (e.g., pH=14, 110 crystal plane) are recovered. The associated attribute database is called to supplement each candidate material with detailed physicochemical descriptors, including structural feature data such as lattice constant, d-band center position of B-site metal, and charge transfer impedance. This information is structured and assembled to generate a detailed list of catalytic materials. This list is usually exported as a CSV file, with each row representing a recommended material and columns containing key fields such as "material ID," "chemical composition," "simulation environment," "predicted overpotential," and "prediction uncertainty."

[0087] By monitoring the performance error decay rate and the high-scoring sample discovery rate, it can be determined whether the mining has reached its limit. When the model enters the steady-state convergence period, it means that continuing the calculation will only yield extremely low marginal benefits (i.e., entering a low-value redundancy state), and the iteration will automatically terminate and the results will be output. This mechanism effectively prevents the disproportionate consumption of computing power for a negligible improvement in accuracy, ensuring that the entire computing task stops in a timely manner when the data utilization and resource consumption ratio reach their optimal levels.

[0088] This invention provides a high-throughput computational method for catalytic materials based on machine learning. It utilizes a two-dimensional sparse matrix to integrate historical data, firstly addressing the problem of redundant computation caused by data fragmentation. Through a multilayer perceptron and a neural collaborative filtering network, a comprehensive value assessment of the physicochemical space is performed before computation to locate high-potential regions. The introduced predictive dispersion mechanism identifies information-scarce areas, guiding computational power to avoid low-value redundant sampling traps and achieving a dynamic balance between exploration and utilization. With the backfilling of simulation data and dynamic adjustment of thresholds, each first-principles calculation is ensured to target the most promising samples. This mechanism improves data utilization, concentrating limited computational power on key samples.

[0089] Example 2

[0090] This embodiment details the application of a machine learning-based high-throughput computational system for catalytic materials in the development of high-performance cobalt-iron-based perovskite oxygen evolution catalysts. Addressing the technical challenges of complex active site mechanisms and vast compositional control in the alkaline water electrolysis oxygen evolution reaction (OER) of non-noble metal catalysts, this embodiment selects La1-xSrxCo1-yFeyO3 (LSCF) and its derived A-site defect-type double perovskite system as the research object; the specific implementation method is as follows:

[0091] In the step of constructing the two-dimensional sparse matrix, the system first retrieves a dedicated dataset containing 3,500 historical records. This dataset covers crystal structure data of cobalt-based and iron-based perovskites from the Materials Project database, as well as electrochemical test data accumulated in the laboratory. The processor executes a data cleaning procedure to remove data items containing radioactive elements or non-oxides, ultimately extracting 2,400 unique catalyst component identifiers, covering various stoichiometric combinations from the parent LaCoO3 to highly doped La0.4Sr0.6Co0.2FeO.8O3. Simultaneously, the system sets six key electrochemical reaction condition parameters, including electrolyte environments with pH values ​​of 13 and 14, and three main crystal plane orientations: (100), (110), and (111). Based on the above identifiers and condition parameters, the system generates row logical addresses (0 to 2399) and column logical addresses (0 to 5), constructing a two-dimensional matrix with a dimension of 2400 multiplied by 6. During initialization, the matrix fill rate is NaN (not a number). Then, based on the physicochemical spatial location index pairs, the 680 known effective overpotential values ​​(in V) are written into the corresponding storage cells. For example, for La0.6Sr0.4Co0.2FeO.8O3 under the condition of pH 14 and exposure of the (110) crystal plane, the historical experimental value of 0.36V is written. The address space is traversed to generate a non-empty coordinate index table containing 680 coordinate items, and the corresponding empty value domain mask map is generated simultaneously, in which the unknown region accounts for approximately 95.3%.

[0092] The system receives a non-empty coordinate index table, parses the discrete index of the catalyst being processed (e.g., index value 1560 pointing to SrCoO.8FeO.2O3), maps this index to the catalyst structure fingerprint matrix, and extracts the electronic structure feature vectors of Sr, Co, Fe, and O elements. Key parameters such as the ionic radius of Sr atom (1.44 Å), the number of 3d orbital electrons of Co atom (7), and the Pauling electronegativity of Fe atom (1.83) are retrieved. To characterize the complex A-site and B-site doping effects, the average radius difference (variance) of the A-site cation and the average d-band center position of the B-site metal are calculated and encoded as a 32-dimensional sample identifier. For the alkaline environment of pH=14 corresponding to the column index, a 16-dimensional state feature vector is extracted through the environment embedding matrix. This vector implicitly includes the influence weight of the hydroxide ion concentration field on the surface adsorption energy. Explicit attribute numerical sequences such as electronegativity, atomic radius, and number of valence electrons are retrieved from the associated attribute database and concatenated with the aforementioned vector to generate a aggregated input tensor with a total dimension of 52. The tensor is fed into a multilayer perceptron network, and after layer-by-layer dimensionality reduction mapping with 128 and 64 nodes and ReLU nonlinear activation, a 64-dimensional hybrid characterization vector is finally output. This vector not only represents the material identity of SrCoO.8FeO.2O3 in the high-dimensional feature space, but also deeply couples its potential tendency for surface lattice oxygen to participate in the reaction under a strongly alkaline environment.

[0093] In the steps of generating performance prediction values ​​and prediction dispersion, the hybrid characterization vector is transmitted in parallel to a two-stream network. In the neural collaborative filtering network branch, the 32-dimensional vector is divided into a 16-dimensional sub-vector characterizing the intrinsic properties of the material and a 16-dimensional sub-vector characterizing the external environment. The two sub-vectors undergo element-wise multiplication of their feature dimensions, outputting a cross-interaction feature tensor. This operation physically simulates the adsorption-desorption equilibrium constant of Co-Fe active sites on the catalyst surface and OH- groups in the electrolyte, capturing the linear structure-activity relationship that conforms to the Sabatier principle. In parallel, a fully connected neural network branch drives a four-layer neuron weight matrix to perform a deep nonlinear transformation on the complete vector, focusing on reconstructing the nonlinear effect of electronic spin state transitions (such as the change of Co ions from low-spin to medium-spin states) on the OER overpotential, outputting a nonlinear activation barrier tensor. The fused feature vector, obtained by concatenating the two, is used by a regression prediction head to calculate a weighted sum and output a performance prediction value. For example, for the test sample SrCoO.8FeO.2O3, the model predicts an overpotential of 0.285V. The average variance of the prediction results of each of the five independent sub-models in the deep ensemble module is calculated, and the output prediction dispersion is 0.0045. This low dispersion value indicates that the model has a high confidence level in predicting this region, which is related to the presence of a large amount of similar Sr-based perovskite data in the training set.

[0094] In the step of generating the list to be verified, the system constructs an acquisition function based on the calculated performance prediction value (0.285V) and prediction dispersion (0.0045), combined with the null domain distribution density. Considering that it is currently in the middle of the iteration, the system sets the exploration weight to 1.5 and the utilization weight to 3.0. After nonlinear weighted fusion, the sample obtains a comprehensive score of 92.4. The system retrieves the dynamic threshold of 88.0 for the current period and determines that the sample score exceeds the threshold. It is included in the candidate subset for rigorous physical validity verification. The system calculates that the Goldschmidt tolerance factor t under this ratio is 0.98, which is within the stable range of cubic perovskite (0.9 to 1.0). Moreover, the charge balance verification shows that the average valence state of the B-site metal is +3.6, which meets the electronic structure characteristics required for high OER activity. After verification, this sample and 49 other high-scoring samples are assembled into a list to be verified and output to the simulation calculation module.

[0095] In the simulation calculation and data backfilling steps, the system automatically schedules high-performance computing cluster resources to perform first-principles calculations on SrCoO.8FeO.2O3, which has the highest priority in the list. Based on the sample identifier, the system calls the cubic Pm-3m space group template to construct a 2x2x2 supercell model containing 40 atoms, and according to the random distribution algorithm, Sr atoms occupy the A site, while Co and Fe atoms randomly occupy the B sites in an 8:2 ratio. After generating the POSCAR file, the system configures the INCAR parameters, specifically including: selecting the generalized gradient approximation PBEsol functional, setting the plane wave cutoff energy to 550 eV, applying a Hubbard U correction of 3.32 eV for the 3d orbitals of Co, and a U correction of 4.0 eV for Fe, correcting the self-interaction error of the strongly correlated electron system, enabling spin polarization calculation (ISPIN=2), and setting the initial magnetic moment of ferromagnetism. The simulation job is submitted to a 64-core computing node, and after 48 hours of electronic and ionic step iterations (convergence criterion -0.02 eV / A), the system locks the steady-state structure. The post-processing module parses the OSZICAR and OUTCAR files and extracts the adsorption free energy of the four-step reaction intermediates (OH, O, OOH). Calculations show that the rate-determining step of this material is the conversion from O to OOH, with a theoretical overpotential of 0.292 V. The data storage manager then writes this true value of 0.292 to the corresponding coordinates of the two-dimensional sparse matrix, overwriting the original null value.

[0096] In the closed-loop iteration and termination decision steps, as the actual performance data (0.292V) is backfilled, the system updates the non-empty coordinate set and calculates the absolute error between the predicted value (0.285V) and the actual value to be 0.007V. Combined with data from other validation samples in this round, the system calculates that the mean absolute error (MAE) has decreased to 0.035V, and the error decay rate has been below 3% for two consecutive rounds, indicating that the model has entered a steady-state convergence period. Simultaneously, the gradient of the high-scoring sample discovery rate approaches zero, indicating that all superior materials in the current chemical space have been explored. Based on this, the termination decision logic controller activates the result output generator, performs a full scan of the updated matrix, and finally outputs a list containing 15 preferred catalysts. Among them, SrCoO.8FeO.2O3 ranks first with a low overpotential of 0.292V.

[0097] The matrix construction module maximizes the reuse of historical data, the feature mapping and extraction module accurately identifies high-value samples, and the performance backfilling and termination module eliminates ineffective consumption through closed-loop feedback. This system-level design frees computing power from tedious data cleaning and ineffective trial and error, allowing the entire process to focus on the exploration of high-potential materials, thus maximizing the efficiency of computing resources.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning based high-throughput computational method for catalytic materials, characterized in that, The method comprises the following steps: Call the historical simulation data set, use the catalyst identifier and the condition parameter as the index, get the parameter value, use the known performance index value as the filling value, and establish a two-dimensional sparse matrix; mark the known data area as non-empty coordinates, and mark the unknown area as null value domain; Based on the non-empty coordinates and the catalyst identifier, the sample characteristic attribute, the sample identifier and the condition parameter code are obtained by querying the embedded table, and the mixed feature vector is generated by mapping through the multi-layer perception network; The mixed feature vector is transmitted in parallel to the neural collaborative filtering network processing branch and the fully connected neural network processing branch; the mixed feature vector is subjected to linear and nonlinear feature fusion by the neural collaborative filtering network and the fully connected neural network, the mixed feature vector is subjected to element-by-element multiplication operation in the neural collaborative filtering network processing branch to analyze the second-order cross features of the catalyst components and the reaction environment inside the feature vector, and a cross interaction feature tensor is output; in the fully connected neural network processing branch, a multi-layer neuron weight matrix is driven to perform continuous matrix multiplication operation and nonlinear activation function transformation on the mixed feature vector to reconstruct the nonlinear space structure, and output a nonlinear activation energy barrier tensor; the cross interaction feature tensor and the nonlinear activation energy barrier tensor are spliced to construct a fusion feature vector, and the weighted sum value is calculated as a performance prediction value; The performance prediction value and the prediction dispersion are generated by evaluating the dispersion degree of the feature distribution as the prediction dispersion through the variance evaluation node; according to the performance prediction value, the prediction dispersion and the mixed feature vector, a sample score is obtained, based on the condition parameter, the samples with sample scores higher than the dynamic threshold are intercepted, and a to-be-verified list is output; The performance data is obtained by calculating the to-be-verified list through the simulation calculation model, and is backfilled to the two-dimensional sparse matrix to update the non-empty coordinates and adjust the dynamic threshold; Based on the termination condition, it is judged whether the termination condition is met or not, and the non-empty coordinates are iteratively predicted when the termination condition is not met; When the termination condition is met, a list of catalytic materials is output.

2. The machine learning based high-throughput computational method for catalytic materials of claim 1, wherein, The specific generation process of the two-dimensional sparse matrix comprises the following steps: calling the historical simulation data set, extracting the catalyst identifier sequence and the condition parameter sequence, converting the catalyst identifier sequence into row logical addresses, and simultaneously converting the condition parameter sequence into column logical addresses to generate a physical-chemical space positioning index pair; initializing the two-dimensional sparse matrix, writing the known performance index values in the historical simulation data set into the corresponding storage units according to the pointing of the physical-chemical space positioning index pair; traversing the addressing space, identifying the storage units containing effective values, extracting the row and column position information to generate a non-empty coordinate index table; identifying the blank storage units not containing effective values, and recording the distribution positions to generate a null value domain mask graph.

3. The machine learning based high-throughput computational method for catalytic materials of claim 2, wherein, The specific generation process of the mixed representation vector includes: receiving the non-empty coordinate index table, extracting the catalyst discrete index and the condition discrete index by parsing the row positioning address and the column positioning address; mapping the catalyst discrete index to the pre-set catalyst structure fingerprint matrix to extract the electronic structure feature vector as the sample identifier; mapping the condition discrete index to the environment embedding matrix to extract the state feature vector as the condition parameter code; calling the corresponding attribute value sequence from the associated attribute database as the sample feature attribute, and performing channel splicing with the sample identifier and the condition parameter code to generate an aggregated input tensor; transmitting the aggregated input tensor to a multi-layer perception network to perform weighted summation operation and nonlinear activation function transformation through multiple neuron nodes to obtain a feature value sequence as a mixed representation vector containing material-environment coupling features; and the sample feature attribute includes electronegativity, atomic radius, and valence electron number.

4. The machine learning based high-throughput computational method for catalytic materials of claim 1, wherein, The specific generation process of the to-be-verified list includes: receiving the performance prediction value and the prediction dispersion, and calling the mixed representation vector; mapping the mixed representation vector to a feature latent space, calculating the manifold distance between the mixed representation vector and the sample points in the non-empty coordinate set, and calling a kernel density estimation algorithm to generate a null value domain distribution density; constructing an exploration and utilization balance evaluation model, receiving the performance prediction value, the prediction dispersion, and the null value domain distribution density as input variables, performing nonlinear weighted fusion operation to generate a sample score; calling the dynamic threshold value of the current iteration period, performing a numerical comparison operation between the sample score and the dynamic threshold value to select a candidate sample subset whose score exceeds the dynamic threshold value; analyzing the condition parameters associated with the candidate sample subset, performing logical verification of the physicochemical parameters according to the pre-set experimental boundary constraint condition, and removing invalid samples that do not satisfy the experimental boundary constraint condition; and assembling the sample index and the corresponding parameter code that pass the verification into the to-be-verified list.

5. The machine learning based high-throughput computational method for catalytic materials of claim 1, wherein, The specific process of updating the non-empty coordinates and adjusting the dynamic threshold value includes: performing format conversion on the condition parameter code in the to-be-verified list to generate an input control file stream, distributing it to the computing nodes, performing first-principle simulation operation through the simulation calculation model, monitoring the operation convergence state and capturing the output log data stream; analyzing the output log data stream to extract the scalar value representing material activity as performance data; locking the storage address of the corresponding null value domain in the two-dimensional sparse matrix through the data storage manager, and writing the performance data into the corresponding storage unit; scanning the storage unit position of the newly written data to extract its row and column positioning information, and appending the row and column positioning information to the non-empty coordinate index table to update the non-empty coordinate set; calling the performance data set corresponding to the updated non-empty coordinates and the performance prediction value set, calculating the distribution gradient of the prediction error, correcting the boundary parameters of the sampling strategy according to the distribution gradient, and outputting the updated dynamic threshold value.

6. The machine learning based high-throughput computational method for catalytic materials of claim 1, wherein, The specific architecture process of the simulation calculation model includes: receiving a to-be-verified list, calling a preset crystal structure template based on a sample identifier, and adjusting lattice parameters according to condition parameters to generate spatial geometric topology data representing the micro-configuration of the material; constructing a discretized numerical calculation grid according to the spatial geometric topology data, and setting boundary constraint conditions in combination with preset environment parameters to generate a simulation input file; loading the simulation input file into an iterative solver, initializing a system state vector, and driving a self-consistent field loop calculation engine to perform nonlinear variational solving on the control equation set; calculating the convergence residual of the energy functional in real time during the solving process, and when the convergence residual falls within a preset tolerance range, determining that the calculation converges and locking the system steady state; triggering a post-processing extraction module to analyze the electronic structure data at the steady state, and calculating a scalar value representing the catalytic activity of the material as performance data.

7. The machine learning based high-throughput computational method of catalytic materials according to claim 1, wherein, The specific process of judging based on the termination condition includes: receiving the updated two-dimensional sparse matrix, and calling the performance data distribution statistics of the last continuous iteration period; performing a convergence evaluation operation, calculating the decay rate of the performance prediction error and the change gradient of the high sample discovery rate as system state indicators; mapping the system state indicators to a preset termination judgment logic controller, if the determination result indicates that the model state is in a shock period, generating an iteration control signal, defining the updated non-empty coordinate set as a new seed data set, and transmitting the new seed data set to the embedded table query interface through a feedback loop to drive the multilayer perception network to perform a new round of value evaluation on the remaining empty value domain and correct the sampling strategy; if the determination result indicates that the model state is in a steady state convergence period, activating a result output generator; performing full-scan on all non-empty coordinates in the two-dimensional sparse matrix according to the final dynamic threshold standard, and extracting target units whose performance values meet the standard; analyzing the sample identifiers associated with the target units to construct a list of catalytic materials containing material components and structural features.

8. A machine learning based high-throughput computational system for catalytic materials, characterized in that, It includes: a matrix construction module: calling a historical simulation data set, using a catalyst identifier and a condition parameter as an index to obtain parameter values, and using a known performance indicator value as a filling value to establish a two-dimensional sparse matrix; marking the known data region as a non-empty coordinate and the unknown region as an empty value domain; a feature mapping module: based on the non-empty coordinates and the catalyst identifier, obtaining sample feature attributes, sample identifiers and condition parameter encodings by querying an embedded table, and generating a hybrid feature vector by mapping through a multilayer perception network; The intercepting sample module: the mixed representation vector is transmitted to the neural collaborative filtering network processing branch and the full connection neural network processing branch in parallel; linear and nonlinear feature fusion is performed on the mixed representation vector by the neural collaborative filtering network and the full connection neural network; in the neural collaborative filtering network processing branch, element-by-element multiplication operation of feature dimension is performed on the mixed representation vector to analyze the second-order cross features of catalyst components and reaction environment inside the feature vector and output cross interaction feature tensor; in the full connection neural network processing branch, a plurality of layers of neuron weight matrix are driven to perform continuous matrix multiplication operation and nonlinear activation function transformation on the mixed representation vector to reconstruct nonlinear space structure and output nonlinear activation energy barrier tensor; the cross interaction feature tensor and the nonlinear activation energy barrier tensor are spliced to construct fusion feature vector, and weighted sum value is calculated as performance prediction value; The performance backfilling module: the to-be-verified list is calculated by the simulation calculation model to obtain performance data, which is backfilled to the two-dimensional sparse matrix to update the non-empty coordinates and adjust the dynamic threshold value; The termination judgment module: based on the termination condition, if the termination condition is not met, the non-empty coordinates are iteratively predicted; If the termination condition is met, the list of catalytic materials is output. ​

Images