High-chromium cast iron material digitalization collaborative research method and system

Abstract

The application provides a high chromium cast iron material digitalization collaborative research method and system, and belongs to the cross field of artificial intelligence and material science. The application constructs a high chromium cast iron structured feature data set by high-throughput literature collection and structured extraction driven by a large language model, realizes positive prediction of alloy performance by fusing a graph attention network and a multilayer perceptron, constructs a continuous and interpolable low-dimensional hidden space by a variational autoencoder, realizes inverse formula optimization with double target constraints by a conditional generative adversarial network, integrates retrieval enhanced generated knowledge question and answer services, and forms an end-to-end digitalization material collaborative design closed loop.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and materials science, and in particular to a digital and intelligent collaborative research and development method and system for high-chromium cast iron materials. Background Technology

[0002] High-chromium cast iron is a crucial type of wear-resistant metallic engineering material in industrial applications. Due to the formation of a chromium-rich carbide network and a martensitic / austenitic composite matrix during solidification, it exhibits excellent wear resistance under harsh conditions such as mining, cement smelting, and power grinding, making it the preferred material for grinding media in industries like mining, power, and metallurgy. However, high-chromium cast iron alloys are complex systems with highly coupled multi-component chemical composition and heat treatment processes. There is an inherent physical constraint between its high hardness and high impact toughness. Traditional trial-and-error development methods heavily rely on researchers' experience, making it difficult to achieve global collaborative optimization in high-dimensional parameter spaces, resulting in long development cycles and high costs.

[0003] The existing technology has the following shortcomings: 1. A large amount of valuable composition-process-performance mapping knowledge in the field is scattered in academic literature in an unstructured form. Traditional keyword retrieval and manual extraction methods are inefficient and cannot meet the needs of machine learning models for large-scale structured feature libraries; 2. When processing alloy composition, traditional machine learning methods simply concatenate features of heterogeneous dimensions into a one-dimensional dense vector, ignoring the inherent complex relationships between alloy elements in the physicochemical reaction process, which limits the prediction accuracy of the model; 3. Material reverse design is a typical one-to-many problem in mathematics. The formulation space that satisfies specific performance constraints is not unique and is highly non-convex, lacking efficient reverse optimization tools; 4. The massive amount of unstructured domain literature knowledge has not yet been organically integrated with material performance prediction and reverse design, and there is a lack of intelligent knowledge question answering services for the material field. Summary of the Invention

[0004] To address the problems in the existing technologies, this invention provides a digital and intelligent collaborative R&D method and system for high-chromium cast iron materials. This invention utilizes a large language model-driven high-throughput document acquisition and structured extraction to construct a structured feature dataset for high-chromium cast iron. It integrates graph attention networks and multilayer perceptrons to achieve positive prediction of alloy properties, uses variational autoencoders to construct a continuously interpolable low-dimensional latent space, and employs conditional generative adversarial networks to achieve inverse formula optimization under dual-objective constraints. Simultaneously, it integrates retrieval-enhanced knowledge-based question-answering services, forming an end-to-end digital and intelligent collaborative material design closed loop. To achieve the above objectives, the technical solution is as follows: On the one hand, this invention provides a digital and intelligent collaborative research and development method for high-chromium cast iron materials, the method comprising: S1. Based on academic databases, high-throughput literature collection and dual-judgment protocol screening using large language models are used to obtain the screened literature. S2. Based on the selected literature, a high-chromium cast iron structured feature dataset is obtained through multidimensional structured feature extraction. S3. Based on the high-chromium cast iron structured feature dataset, obtain the fully connected weighted component topology map and the structured process feature vector; S4. Based on the fully connected weighted component topology graph and the structured process feature vector, obtain the low-dimensional component process features through a variational autoencoder. S5. Based on the characteristics of this low-dimensional component process, positive performance prediction data is obtained by fusing a graph attention network and a multilayer perceptron dual-branch composite prediction network. S6. Based on the positive performance prediction data and the target performance data, the reverse optimization iteration is performed with high hardness and high impact toughness as dual objective constraints, and the high chromium cast iron material formula that meets the target data requirements is obtained through variational self-decoder.

[0005] Optionally, in S1, based on academic databases, high-throughput literature collection is performed, and a dual-judgment protocol using a large language model is used to filter the literature, resulting in the following filtered literature: S11. Based on academic databases, relevant literature on high-chromium cast iron materials was obtained through a multi-threaded concurrent download architecture and a high-fault-tolerant queue mechanism based on the exponential backoff algorithm. S12. Based on relevant literature on this high-chromium cast iron material, the literature was screened through a dual-judgment protocol at the abstract level.

[0006] Optionally, the dual-determination protocol includes: Agreement 1: The main body of the research on this high-chromium cast iron material is based on high-chromium cast iron as the core material; Agreement 2: The relevant literature on this high-chromium cast iron material contains quantifiable data on composition, process, or performance.

[0007] Optionally, in S2, based on the selected literature, a high-chromium cast iron structured feature dataset is obtained through multidimensional structured feature extraction, including: S21. Based on the filtered literature, the first-stage dataset is obtained by discarding data that lacks basic skeleton elements; S22. Based on the first-stage dataset, filter out abnormal data through metallurgical prior interval constraints to obtain the second-stage dataset. S23. Based on the second-stage dataset, the hardness data is uniformly converted and duplicate entries are eliminated through the joint primary key deduplication mechanism to obtain the high-chromium cast iron structured feature dataset.

[0008] Optionally, in S4, based on the fully connected weighted component topology graph and the structured process feature vector, a variational autoencoder is used to obtain low-dimensional component process features, including: S41. Based on the fully connected weighted component topology and the structured process feature vector, the first stage features are obtained through three fully connected dense layers. S42. Based on the characteristics of the first stage, the characteristics of the second stage are obtained through reparameterization techniques. S43. Based on the characteristics of this second stage, the low-dimensional composition process characteristics are obtained through a progressive KL annealing strategy.

[0009] Optionally, in S5, based on the low-dimensional component process characteristics, a dual-branch composite prediction network fusing a graph attention network and a multilayer perceptron is used to obtain positive performance prediction data, including: S51. Based on the characteristics of the low-dimensional component process, the graph attention network branch and the multilayer perceptron branch are obtained; S52. Based on the attention network branches in the graph, feature extraction is performed through a multi-head self-attention node aggregation mechanism to obtain the output data of the first branch; S53. Based on the multilayer perceptron branch, perform deep encoding to obtain the output data of the second branch; S54. Based on the output data of the first branch and the output data of the second branch, the fused features are obtained by late-stage fusion splicing and nonlinear fusion through a fully connected layer. S55. Based on this fusion feature, positive performance prediction data is obtained by predicting using a mask-aware multi-task weighted loss function.

[0010] Optionally, the positive performance prediction data includes: Rockwell hardness and impact toughness.

[0011] Optionally, in step S6, based on the positive performance prediction data and the target performance data, a reverse optimization iteration is performed with high hardness and high impact toughness as dual objective constraints, and a variational self-decoder is used to obtain a high-chromium cast iron material formulation that meets the target data requirements, including: S61. Based on the positive performance prediction data and the target performance data, the discriminant loss function is obtained by using the Wasserstein distance and gradient penalty mechanism. S62. Based on the discriminant loss function, perform inverse optimization iteration with high hardness and high impact toughness as dual objective constraints to obtain low-dimensional optimal data. S63. Based on the low-dimensional optimal data, a high-chromium cast iron material formula that meets the target data requirements is obtained through a variational self-decoder.

[0012] Optionally, it also includes: Based on the target performance data and the high-chromium cast iron material formula that meets the target data requirements, a knowledge question answering system is obtained by accelerating semantic retrieval through HNSW graph indexing and using a two-layer prompt word constraint and source citation mechanism.

[0013] On the other hand, the present invention provides a digital and intelligent collaborative R&D system for high-chromium cast iron materials. This system is applied to a digital and intelligent collaborative R&D method for high-chromium cast iron materials. The system includes: The literature screening module is used to screen literature based on academic databases through high-throughput literature collection and a dual-judgment protocol using a large language model. The data acquisition module is used to obtain a high-chromium cast iron structured feature dataset by extracting multidimensional structured features from the filtered literature. The first feature acquisition module is used to obtain a fully connected weighted component topology map and a structured process feature vector based on the high-chromium cast iron structured feature dataset. The second feature acquisition module is used to obtain low-dimensional component process features by using a variational autoencoder based on the fully connected weighted component topology graph and the structured process feature vector. The positive prediction module is used to obtain positive performance prediction data based on the low-dimensional component process characteristics by fusing a graph attention network and a multilayer perceptron dual-branch composite prediction network. The formula acquisition module is used to perform reverse optimization iteration based on the positive performance prediction data and target performance data, with high hardness and high impact toughness as dual objective constraints, and obtain a high-chromium cast iron material formula that meets the target data requirements through a variational self-decoder.

[0014] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects: The aforementioned scheme, firstly, utilizes a high-throughput literature acquisition and structured feature extraction pipeline driven by a large language model, effectively eliminating noisy literature through a dual-judgment protocol, laying a solid data foundation for subsequent data-driven research; secondly, by fusing a dual-branch composite prediction network of graph attention networks and multilayer perceptrons, it models alloy composition as a fully connected weighted topological graph, introduces intrinsic physicochemical properties of elements as node features, and uses the product of composition concentrations as edge weights, overcoming the limitation of traditional machine learning methods that treat chemical composition as isolated numerical vectors, thus improving the accuracy of alloy performance prediction; thirdly, The variational autoencoder is trained using a progressive KL annealing strategy to compress the high-dimensional composition-process characteristics of high-chromium cast iron, providing a physically legitimate exploration space for inverse optimization. Fourthly, the Wasserstein distance and gradient penalty mechanism is used to ensure training stability, enabling inverse formulation optimization with high Rockwell hardness and high impact toughness as dual objectives. Fifthly, a retrieval-enhanced knowledge question-answering system using a vector database as the retrieval engine and a large language model as the reasoning core fundamentally suppresses the illusionary output of the large language model through a source citation mechanism, providing researchers with reliable domain knowledge query services. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of an embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention; Figure 2 This is a flowchart of the selected literature obtained in the embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention; Figure 3 This is a flowchart illustrating the process of obtaining the structured feature dataset of high-chromium cast iron in an embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention. Figure 4 This is a flowchart illustrating the low-dimensional composition process characteristics obtained in an embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention; Figure 5 This is a flowchart illustrating the positive performance prediction data obtained in an embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention; Figure 6 This is a flowchart illustrating the high-chromium cast iron material formulation that meets the target data requirements in an embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention. Figure 7 This is a system block diagram of an embodiment of the intelligent collaborative R&D system for high-chromium cast iron materials of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] like Figure 1The flowchart shown is an embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials according to the present invention. The present invention provides an intelligent collaborative R&D method for high-chromium cast iron materials, which is implemented by an intelligent collaborative R&D system for high-chromium cast iron materials. The method includes: S1. Based on academic databases, high-throughput literature collection and dual-judgment protocol screening using large language models are used to obtain the screened literature. Specifically, such as Figure 2 The flowchart shown in this embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention illustrates the process of obtaining screened literature. In step S1, based on academic databases, high-throughput literature collection is performed, and a dual-judgment protocol using a large language model is used to screen the literature, resulting in the following: S11. Based on academic databases, relevant literature on high-chromium cast iron materials was obtained through a multi-threaded concurrent download architecture and a high-fault-tolerant queue mechanism based on the exponential backoff algorithm. S12. Based on relevant literature on this high-chromium cast iron material, the literature was screened through a dual-judgment protocol at the abstract level.

[0021] Furthermore, the dual-determination protocol includes: Agreement 1: The main body of the research on this high-chromium cast iron material is based on high-chromium cast iron as the core material; Agreement 2: The relevant literature on this high-chromium cast iron material contains quantifiable data on composition, process, or performance.

[0022] Furthermore, using the Web of Science Core Collection as the primary data source, a keyword combination strategy encompassing multiple synonyms such as high chromiumcast iron, HCCI, high-Cr white iron, and Fe-Cr-C alloy was employed to construct the search query, maximizing recall within the domain. A multi-threaded concurrent download architecture was used to maintain multiple independent browser instances, automatically collecting approximately 4,300 full-text PDFs of high-chromium cast iron within several days, achieving a coverage rate of 69.4%. A high-fault-tolerant queue mechanism based on the exponential backoff algorithm effectively addressed network jitter and API rate limitations, with the waiting time increasing after each failed request according to the following formula: (1) in, W(n) For the first n Waiting time after a failed request n For the current retry round, W 0 Based on waiting time, r The introduced microsecond-level random noise is used to prevent concurrent peak conflicts from occurring when multiple parallel threads retry.

[0023] Then, a large language model was used to conduct in-depth semantic cognitive screening of each literature abstract. The first criterion required that the research subject of the literature focus on high-chromium cast iron; the second criterion required that the literature contain quantifiable composition, process, or performance data. The temperature parameter was forcibly set to 0 in the API call to ensure the certainty and repeatability of the judgment results. The dual judgment protocol improved the consistency rate of manual annotation to 93.2%, effectively eliminating noisy literature and selecting approximately 3,600 highly relevant articles.

[0024] S2. Based on the selected literature, a high-chromium cast iron structured feature dataset is obtained through multidimensional structured feature extraction. Specifically, such as Figure 3 The flowchart shown in the embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention illustrates the process of obtaining a structured feature dataset of high-chromium cast iron. In step S2, based on the screened literature, a structured feature dataset of high-chromium cast iron is obtained through multi-dimensional structured feature extraction, including: S21. Based on the filtered literature, the first-stage dataset is obtained by discarding data that lacks basic skeleton elements; S22. Based on the first-stage dataset, filter out abnormal data through metallurgical prior interval constraints to obtain the second-stage dataset. S23. Based on the second-stage dataset, the hardness data is uniformly converted and duplicate entries are eliminated through the joint primary key deduplication mechanism to obtain the high-chromium cast iron structured feature dataset.

[0025] Furthermore, considering the highly diverse data representation formats in materials science academic papers, this embodiment designs a multidimensional structured data parsing framework covering 22 subdivided characterization fields across three dimensions: composition, process, and performance. By constraining the output of the large language model with JSON format and employing a three-stage processing flow for longer documents—segmented input, local extraction, and cross-segment merging—it avoids confusion in parameter serial correspondence and errors in cross-segment correspondence. First, data entries lacking basic framework elements such as carbon content, chromium content, and at least one performance indicator are discarded. Second, physically meaningless anomalous data are filtered out using metallurgical prior interval constraints (e.g., carbon content interval [1.5, 4.5] wt%, chromium content interval [10, 35] wt%). Third, hardness data undergoes unified unit conversion (Rockwell HRC, Vickers HV, and Brinell HB conversion), and duplicate entries are eliminated through a joint primary key deduplication mechanism. After a comprehensive and meticulous data cleaning pipeline process, the HCCI-3500 dataset, containing 3,568 complete and compliant data entries with high signal-to-noise ratio, was finally constructed, with a comprehensive field-weighted extraction accuracy of 90.1%.

[0026] S3. Based on the high-chromium cast iron structured feature dataset, obtain the fully connected weighted component topology map and the structured process feature vector; Specifically, based on the structured feature dataset of high-chromium cast iron, the compositional features of each sample are reconstructed into a fully connected weighted compositional topology graph. Each actual alloying element contained in the sample (mainly including Fe, C, Cr, Mn, Si, Mo, Ni, etc.) is treated as a graph node, and the node feature vector consists of four intrinsic physicochemical properties: atomic number, Pauling electronegativity, atomic radius (unit pm), and number of valence electrons. An edge is established between any two nodes, and the edge weight is defined as the normalized product of the mass fractions of the elements at both ends. (2) in, This is the normalized product of the mass fractions of the elements at both ends. x i For elements i Percentage quality score, x j For elements j The percentage mass fraction. This weight explicitly encodes the combined thermodynamic probability and interaction strength of the metallurgical reaction between the two in the liquid melt or solid solution.

[0027] Key parameters covering the complete heat treatment process were systematically extracted from the HCCI-3500 dataset to construct a fixed 10-dimensional structured process feature vector, mainly including core scalars such as casting temperature, austenitizing temperature, quenching holding time, tempering temperature, and tempering holding time. For categorical feature variables (such as quenching cooling media), ordered numerical encoding was performed based on the physical nature of their cooling rate. Cooling rate gradients were set, with 0.0 for furnace cooling, 1.0 for natural air cooling, 2.0 for forced air cooling, 3.0 for oil cooling, and 4.0 for water cooling, ensuring that the relative magnitudes of each value corresponded to the physical magnitude of the actual cooling rate.

[0028] S4. Based on the fully connected weighted component topology graph and the structured process feature vector, obtain the low-dimensional component process features through a variational autoencoder. Specifically, such as Figure 4 The flowchart shown in the embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention obtains low-dimensional compositional process features. In step S4, based on the fully connected weighted compositional topology graph and the structured process feature vector, the low-dimensional compositional process features are obtained through a variational autoencoder, including: S41. Based on the fully connected weighted component topology and the structured process feature vector, the first stage features are obtained through three fully connected dense layers. S42. Based on the characteristics of the first stage, the characteristics of the second stage are obtained through reparameterization techniques. S43. Based on the characteristics of this second stage, the low-dimensional composition process characteristics are obtained through a progressive KL annealing strategy.

[0029] Furthermore, based on the fully connected weighted component topology and the structured process feature vector, the statistical bias characteristics of the HCCI-3500 dataset are systematically analyzed, revealing the multi-peak distribution of chromium content, the heavy-tailed distribution of wear rate, and the significant scarcity of impact toughness labels, thus clarifying the inherent limitations of direct modeling.

[0030] The variational autoencoder takes an 8-dimensional normalized feature tensor as input, progressively increases its dimensionality through three fully connected dense layers (dimensions increasing sequentially from 8 to 64 to 128 to 256), and outputs a 4-dimensional latent vector with a mean vector μ and a log-variance vector log σ². The latent vector is sampled using a reparameterization technique. (3) (4) Where E[log p(x|z)] is the expected reconstruction loss, KL(·‖·) is the KL divergence, and β is the weight coefficient.

[0031] To prevent the risk of posterior probability collapse, a progressive KL annealing strategy was adopted. During training, the beta coefficient was linearly increased from near 0 to a target value of 0.5, successfully compressing the high-dimensional, dispersed material features into a 4-dimensional continuous Gaussian manifold. The normalized MAE of each field was below 0.05, and the reconstruction accuracy met the tolerance requirements of industrial engineering. t-SNE visualization revealed that the latent space has a spontaneous ordered encoding capability for key parameters such as hardness and carbon-chromium ratio, and linear interpolation verified the physical interpolability of the latent space.

[0032] The quality of the latent space is systematically evaluated through the following three dimensions: (1) Reconstruction accuracy evaluation. Encoding-decoding round-trip tests are performed on all samples in the test set, and the normalized mean square absolute error (MAE) of each feature field is statistically analyzed. The results show that the normalized MAE of core component fields such as carbon content and chromium content is less than 0.03, and the normalized MAE of process fields such as austenitizing temperature is less than 0.05. The overall reconstruction accuracy meets the industrial engineering tolerance requirements, proving the lossless compression capability of the VAE latent space for the original feature information. (2) Latent space continuity verification. Two latent vectors z1 and z2 corresponding to different performance ranges (low hardness / high toughness formulation region and high hardness / low toughness formulation region) are randomly selected in the latent space. Ten intermediate latent vectors are generated between the two in the form of linear interpolation with equal step size. After being restored to the physical formulation by the decoder, they are input into CGP-Net for performance prediction. The results show that the predicted performance values ​​change monotonically and smoothly with the interpolation parameters, proving that the latent space has good physical interpolation properties. There are no physical forbidden zones (i.e., combinations of components that are physically meaningless) on the interpolation path, which verifies the effectiveness of the KL divergence regularization constraint. (3) Latent space semantic visualization. The t-SNE dimensionality reduction algorithm is used to map the 4-dimensional latent space to a 2D plane. After coloring with hardness labels, it can be clearly observed that the latent space has a spontaneous ordered encoding ability for Rockwell hardness. High hardness samples and low hardness samples form obvious spatial clustering regions in the latent space, which further verifies the effective characterization and organization ability of the latent space for material properties.

[0033] S5. Based on the characteristics of this low-dimensional component process, positive performance prediction data is obtained by fusing a graph attention network and a multilayer perceptron dual-branch composite prediction network. Specifically, such as Figure 5 The flowchart shown in the embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention illustrates the process of obtaining positive performance prediction data. In step S5, based on the low-dimensional composition process characteristics, positive performance prediction data is obtained through a dual-branch composite prediction network that fuses a graph attention network and a multilayer perceptron, including: S51. Based on the characteristics of the low-dimensional component process, the graph attention network branch and the multilayer perceptron branch are obtained; S52. Based on the attention network branches in the graph, feature extraction is performed through a multi-head self-attention node aggregation mechanism to obtain the output data of the first branch; S53. Based on the multilayer perceptron branch, perform deep encoding to obtain the output data of the second branch; S54. Based on the output data of the first branch and the output data of the second branch, the fused features are obtained by late-stage fusion splicing and nonlinear fusion through a fully connected layer. S55. Based on this fusion feature, positive performance prediction data is obtained by using a mask-aware multi-task weighted loss function.

[0034] Furthermore, this positive performance prediction data includes: Rockwell hardness and impact toughness.

[0035] Furthermore, a late-stage fusion deep neural network architecture is adopted, comprising two parallel branches: the first branch outputs data from a graph attention network (GAT) branch, which extracts features from the composition topology graph through a multi-head self-attention node aggregation mechanism, adaptively learning the asymmetric importance weights of each neighbor node to the central node, effectively capturing the differentiated interaction relationships between elements in a specific chemical ratio system; the second branch outputs data from a multilayer perceptron (MLP) branch, which performs deep encoding of the process feature vectors. The outputs of the two branches are concatenated by the late-stage fusion module and then nonlinearly fused through a fully connected layer. A mask-aware multi-task weighted loss function is then used to simultaneously predict Rockwell hardness (HRC) and impact toughness (J / cm²).

[0036] To address the issue that the impact resilience label coverage in the HCCI-3500 dataset is only about 41.2%, a mask-aware multi-task weighted loss function is introduced: (5) Among them, mask k w serves as the validity flag (0 or 1) for the k-th performance indicator. k L represents the weighting coefficient for the corresponding task. k The loss is the prediction loss for the corresponding task. This strategy effectively solves the problem of misaligned performance labels in the training data and improves the generalization robustness of the model under sparse performance label conditions. Experimental results show that CGP-Net achieves an R² coefficient of 0.921 in hardness prediction, which is superior to traditional machine learning methods such as baseline random forest and XGBoost.

[0037] The dual-branch composite prediction network, which integrates graph attention networks and multilayer perceptrons, is trained using the Adam optimizer with an initial learning rate of 0.001. A cosine annealing (Cosine AnnealingLR) learning rate scheduling strategy is employed to gradually reduce the learning rate to 1×10⁻⁶ over 200 training epochs. -5To avoid oscillations in the later stages of training, the batch size was set to 64. A random partitioning strategy was used to divide the HCCI-3500 dataset into training, validation, and test sets in an 8:1:1 ratio, and an early stopping mechanism (patience=20) was implemented to prevent overfitting. The GAT branch uses a two-layer graph attention convolution. The first layer employs a four-head parallel attention mechanism, concatenating and fusing the outputs of each head. The second layer uses single-head attention to further filter redundant information, outputting a 128-dimensional component feature vector. The MLP process branch uses a two-layer fully connected network with 64 and 64 neurons respectively, projecting the 10-dimensional process input vector into a 64-dimensional latent space. Each layer is configured with batch normalization (BatchNorm) and Dropout (dropout rate 0.2). The late-stage fusion module concatenates the GAT output (128-dimensional) and MLP output (64-dimensional) into a 192-dimensional feature. This feature is then transformed nonlinearly through a fully connected layer and ReLU activation, and configured with a dropout rate of 0.2. Finally, it is mapped to the multi-task prediction head to output the predicted values ​​of Rockwell hardness (HRC) and impact toughness (J / cm²).

[0038] S6. Based on the positive performance prediction data and the target performance data, the reverse optimization iteration is performed with high hardness and high impact toughness as dual objective constraints, and the high chromium cast iron material formula that meets the target data requirements is obtained through variational self-decoder.

[0039] Specifically, such as Figure 6 The flowchart shown in the embodiment of the intelligent collaborative R&D method for high-chromium cast iron materials of the present invention obtains a high-chromium cast iron material formulation that meets the target data requirements. In step S6, based on the positive performance prediction data and the target performance data, a reverse optimization iteration is performed with high hardness and high impact toughness as dual objective constraints, and a variational self-decoder is used to obtain a high-chromium cast iron material formulation that meets the target data requirements, including: S61. Based on the positive performance prediction data and the target performance data, the discriminant loss function is obtained by using the Wasserstein distance and gradient penalty mechanism. S62. Based on the discriminant loss function, perform inverse optimization iteration with high hardness and high impact toughness as dual objective constraints to obtain low-dimensional optimal data; S63. Based on the low-dimensional optimal data, a high-chromium cast iron material formula that meets the target data requirements is obtained through a variational self-decoder.

[0040] Furthermore, the Wasserstein distance and gradient penalty (WGAN-GP) mechanism is employed for training, effectively overcoming the gradient vanishing and mode collapse problems of standard GANs. The discriminator loss is: (6) in, This is a random interpolation between real samples and generated samples. is the gradient penalty coefficient. The generator simultaneously constrains two opposing performance indicators, hardness (target HRC≥62) and impact toughness (target ≥6.5 J / cm²), in the form of Hinge loss. The total loss is: (7) in, and These are the weighting coefficients for hardness constraints and toughness constraints, respectively. and These are the corresponding Hinge constraint loss terms.

[0041] Specifically, it also includes: Based on the target performance data and the high-chromium cast iron material formula that meets the target data requirements, a knowledge question answering system is obtained by accelerating semantic retrieval through HNSW graph indexing and using a two-layer prompt word constraint and source citation mechanism.

[0042] Furthermore, based on the target performance data and the high-chromium cast iron material formulation that meets the target data requirements, the collected full-text literature on high-chromium cast iron was recursively segmented into blocks using a sliding window (maximum block length 600 characters, overlap 60 characters). The text blocks were vectorized using a 768-dimensional text embedding model, and after L2 normalization, stored in a vector database, generating approximately 62,000 valid text blocks. Secondly, HNSW (Hierarchical Navigable Small Text Array) was used... The World graph index structure accelerates semantic retrieval, compressing the single retrieval response time to less than 50ms. An adaptive Top-K recall strategy achieves a dynamic balance between precision and coverage, automatically adjusting the number of recalls based on query complexity. Then, a multi-hop referencing resolution mechanism addresses the contextual gaps in multi-turn dialogues. By tracking entity referencing relationships in the dialogue history, the system automatically expands the semantic scope of subsequent queries, ensuring the coherence of multi-turn question-and-answer sessions. Finally, a two-layer prompt constraint framework and a source citation mechanism fundamentally suppress the illusionary output of large language models. System responses must include traceable literature sources, reducing the illusionary rate from approximately 38% at the baseline to approximately 9%.

[0043] like Figure 7 The diagram shown is a system block diagram of an embodiment of the intelligent collaborative R&D system for high-chromium cast iron materials according to the present invention. The present invention provides an intelligent collaborative R&D system for high-chromium cast iron materials, which is applied to an intelligent collaborative R&D method for high-chromium cast iron materials. The system includes: a literature screening module, a data acquisition module, a first feature acquisition module, a second feature acquisition module, a forward prediction module, and a formula acquisition module. Specifically, The literature screening module is used to screen literature based on academic databases through high-throughput literature collection and a dual-judgment protocol using a large language model. The data acquisition module is used to obtain a high-chromium cast iron structured feature dataset by extracting multidimensional structured features from the filtered literature. The first feature acquisition module is used to obtain a fully connected weighted component topology map and a structured process feature vector based on the high-chromium cast iron structured feature dataset. The second feature acquisition module is used to obtain low-dimensional component process features by using a variational autoencoder based on the fully connected weighted component topology graph and the structured process feature vector. The positive prediction module is used to obtain positive performance prediction data based on the low-dimensional component process characteristics by fusing a graph attention network and a multilayer perceptron dual-branch composite prediction network. The formula acquisition module is used to perform reverse optimization iteration based on the positive performance prediction data and target performance data, with high hardness and high impact toughness as dual objective constraints, and obtain a high-chromium cast iron material formula that meets the target data requirements through a variational self-decoder.

[0044] This invention provides a digital and intelligent collaborative research and development method and system for high-chromium cast iron materials. Firstly, it utilizes a high-throughput literature acquisition and structured feature extraction pipeline driven by a large language model. A dual-judgment protocol effectively eliminates noisy literature, laying a solid data foundation for subsequent data-driven research. Secondly, by fusing a graph attention network and a multilayer perceptron's bi-branch composite prediction network, it models the alloy composition as a fully connected weighted topological graph, introducing intrinsic physicochemical properties of elements as node features and using the product of composition concentrations as edge weights. This overcomes the limitation of traditional machine learning methods that treat chemical composition as isolated numerical vectors, improving... The system improves the accuracy of alloy property prediction. Secondly, a progressive KL annealing strategy is used to train a variational autoencoder, compressing the high-dimensional compositional and process characteristics of high-chromium cast iron, providing a physically legitimate exploration space for inverse optimization. Then, a Wasserstein distance and gradient penalty mechanism is used to ensure training stability, achieving inverse formula optimization with high Rockwell hardness and high impact toughness as dual objectives. Finally, a retrieval-enhanced knowledge-generating question-answering system, using a vector database as the retrieval engine and a large language model as the reasoning core, fundamentally suppresses the illusionary output of the large language model through a source citation mechanism, providing researchers with reliable domain knowledge query services.

[0045] It is understood that the present invention has been described through the above embodiments and should not be construed as limiting the implementation and scope of the present invention. Those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A digital and intelligent collaborative research and development method for high-chromium cast iron materials, characterized in that, The method includes: S1. Based on academic databases, high-throughput literature collection and dual-judgment protocol screening using large language models are used to obtain the screened literature. S2. Based on the selected literature, a high-chromium cast iron structured feature dataset is obtained through multi-dimensional structured feature extraction. S3. Based on the high-chromium cast iron structured feature dataset, obtain the fully connected weighted component topology graph and the structured process feature vector; S4. Based on the fully connected weighted component topology graph and the structured process feature vector, obtain low-dimensional component process features through a variational autoencoder. S5. Based on the low-dimensional component process characteristics, positive performance prediction data is obtained by fusing a graph attention network and a multilayer perceptron dual-branch composite prediction network. S6. Based on the positive performance prediction data and target performance data, reverse optimization iteration is performed with high hardness and high impact toughness as dual objective constraints, and a variational self-decoder is used to obtain a high-chromium cast iron material formula that meets the target data requirements.

2. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, In step S1, based on academic databases, high-throughput literature collection is performed, and a dual-judgment protocol using a large language model is used for screening to obtain the screened literature, including: S11. Based on academic databases, relevant literature on high-chromium cast iron materials was obtained through a multi-threaded concurrent download architecture and a high-fault-tolerant queue mechanism based on the exponential backoff algorithm. S12. Based on the relevant literature on the high-chromium cast iron material, the literature is screened through a dual-judgment protocol at the abstract level to obtain the screened literature.

3. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 2, characterized in that, The dual-determination protocol includes: Agreement 1: The main body of the research on the relevant literature on the high-chromium cast iron material is high-chromium cast iron as the core material; Agreement 2: The relevant literature on the high-chromium cast iron material contains quantifiable data on composition, process or performance.

4. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, In step S2, based on the screened literature, a high-chromium cast iron structured feature dataset is obtained through multi-dimensional structured feature extraction, including: S21. Based on the filtered literature, the first-stage dataset is obtained by discarding data with missing basic skeleton elements; S22. Based on the first stage dataset, filter out abnormal data through metallurgical prior interval constraints to obtain the second stage dataset; S23. Based on the second stage dataset, the hardness data is uniformly converted to units and duplicate entries are eliminated through the joint primary key deduplication mechanism to obtain a high-chromium cast iron structured feature dataset.

5. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, In step S4, based on the fully connected weighted component topology graph and the structured process feature vector, a low-dimensional component process feature is obtained through a variational autoencoder, including: S41. Based on the fully connected weighted component topology graph and the structured process feature vector, the first stage features are obtained through three fully connected dense layers. S42. Based on the features of the first stage, the features of the second stage are obtained through reparameterization techniques; S43. Based on the characteristics of the second stage, the low-dimensional composition process characteristics are obtained through a progressive KL annealing strategy.

6. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, In step S5, based on the low-dimensional component process characteristics, positive performance prediction data is obtained through a dual-branch composite prediction network that fuses a graph attention network and a multilayer perceptron, including: S51. Based on the low-dimensional component process characteristics, obtain the graph attention network branch and the multilayer perceptron branch; S52. Based on the graph attention network branches, feature extraction is performed through a multi-head self-attention node aggregation mechanism to obtain the output data of the first branch; S53. Based on the multilayer perceptron branches, perform deep encoding to obtain the second branch output data; S54. Based on the output data of the first branch and the output data of the second branch, late-stage fusion splicing and nonlinear fusion through a fully connected layer are performed to obtain fused features; S55. Based on the fusion features, positive performance prediction data is obtained by predicting using a mask-aware multi-task weighted loss function.

7. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, The positive performance prediction data includes: Rockwell hardness and impact toughness.

8. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, In step S6, based on the positive performance prediction data and the target performance data, a reverse optimization iteration is performed with high hardness and high impact toughness as dual objective constraints. Through a variational self-decoder, a high-chromium cast iron material formulation that meets the target data requirements is obtained, including: S61. Based on the positive performance prediction data and target performance data, the discriminant loss function is obtained by using the Wasserstein distance and gradient penalty mechanism. S62. Based on the discriminant loss function, perform inverse optimization iteration with high hardness and high impact toughness as dual objective constraints to obtain low-dimensional optimal data; S63. Based on the low-dimensional optimal data, a high-chromium cast iron material formula that meets the target data requirements is obtained through a variational self-decoder.

9. The intelligent collaborative R&D method for high-chromium cast iron materials according to claim 1, characterized in that, Also includes: Based on the target performance data and the high-chromium cast iron material formula that meets the target data requirements, a knowledge question answering system is obtained by accelerating semantic retrieval through HNSW graph indexing and using a two-layer prompt word constraint and source tracing reference mechanism.

10. A digital collaborative R&D system for high-chromium cast iron materials, used to implement the digital collaborative R&D method for high-chromium cast iron materials as described in any one of claims 1-9, characterized in that, The system includes: The literature screening module is used to screen literature based on academic databases through high-throughput literature collection and a dual-judgment protocol using a large language model. The data acquisition module is used to obtain a high-chromium cast iron structured feature dataset by extracting multidimensional structured features from the filtered documents. The first feature acquisition module is used to obtain a fully connected weighted component topology map and a structured process feature vector based on the high-chromium cast iron structured feature dataset. The second feature acquisition module is used to obtain low-dimensional component process features by using a variational autoencoder based on the fully connected weighted component topology graph and the structured process feature vector. The positive prediction module is used to obtain positive performance prediction data based on the low-dimensional component process characteristics by fusing a graph attention network and a multilayer perceptron dual-branch composite prediction network. The formula acquisition module is used to perform reverse optimization iteration based on the positive performance prediction data and target performance data, with high hardness and high impact toughness as dual objective constraints, and obtain a high-chromium cast iron material formula that meets the target data requirements through a variational self-decoder.

Images