A potential diffusion generation method for alloy material reverse design

Abstract

The application discloses a potential diffusion generation method for alloy material reverse design, and the method comprises the following steps: after pre-processing alloy historical sample data, a joint latent space representation is obtained based on a double-flow heterogeneous variational autoencoder mapping; a conditional diffusion generation model is constructed and trained based on the representation, and the target generation model is obtained by jointly optimizing the target performance condition vector, the retrieval enhancement mechanism and the reinforcement learning fine-tuning strategy during the training process; the target performance condition input by a user is acquired, the target generation model is used for performing conditional sampling, and multiple candidate latent variables are generated; screening is performed on the multiple candidate latent variables, and the effective candidate latent variables obtained after the screening are restored to alloy component distribution ratios and process step sequences. The application can improve the diversity, stability and reducibility of generated samples while keeping the target orientation, and solves the technical problems of long design cycle, high cost and great difficulty in multi-target consideration of traditional alloy design.

Description

Technical Field

[0001] This invention relates to the field of intelligent material design and generation model technology, and more specifically to a potential diffusion generation method for reverse design of alloy materials. Background Technology

[0002] Alloys have irreplaceable application value in many high-end manufacturing fields such as aerospace, biomedicine, chemical industry, and shipbuilding. The design process of alloys is extremely complex, usually requiring the simultaneous consideration of multiple interdependent performance objectives such as strength, plasticity, and corrosion resistance. Furthermore, there is a close coupling relationship between the alloy's composition and processing technology; even slight changes in composition or adjustments to process parameters can lead to significant differences in the final performance.

[0003] Traditional alloy design relies heavily on the experience of researchers, employing a trial-and-error approach with repeated experiments. This method is not only time-consuming and costly, but also struggles to achieve optimal balance across multiple performance objectives, failing to quickly respond to the demands of industrial production for novel alloy materials. With the application of artificial intelligence in materials science, generative model-based alloy design methods have emerged, attempting to shorten design cycles and reduce R&D costs by generating candidate samples. However, existing methods still have many limitations in practical applications, failing to meet the needs of industrial applications. Specifically, they suffer from the following three main problems:

[0004] First, data representation is inconsistent, and cross-modal information is underutilized. Existing methods typically process alloy composition, processing, and performance data separately, lacking an effective and unified characterization approach. This results in the inability to fully capture the coupling relationship between composition and processing. For example, some methods only model composition data, ignoring the impact of process parameters on performance; while other methods consider both composition and processing, they fail to achieve deep integration, causing the model to be unable to learn the intrinsic correlation between composition, processing, and performance, resulting in significant deviations between the generated sample performance and the target requirements.

[0005] Second, the model training is disconnected from the target constraints, resulting in poor target hit stability and insufficient multi-objective balance optimization capabilities. Although some generative models can generate a large number of alloy samples, they lack effective target performance guidance during training, leading to strong randomness in the generated samples and difficulty in consistently meeting the user-defined multi-objective performance requirements. At the same time, there are significant mutual constraints among the multi-objective properties of alloys (such as strength-ductility-brittleness), and existing methods lack targeted balance optimization mechanisms, making it difficult to achieve optimal balance of multi-objective performance. Even when some methods introduce target constraints, they often use simple conditional input methods, lacking precise guidance mechanisms for target performance, resulting in large fluctuations in the target hit rate of generated samples, which cannot meet the stability and balance requirements of industrial design.

[0006] Third, existing models suffer from a lack of physical logic, resulting in insufficient engineering reliability of the generated samples. Some existing generative models rely solely on data-driven training, lacking the integration of the inherent physical logic and processing constraints between alloy composition, process, and performance. This leads to generated samples that may contain logical contradictions in process steps, poor matching between composition and process, and do not conform to the physical laws of actual alloy processing.

[0007] Therefore, in view of the shortcomings of existing technologies, there is an urgent need in this field for an integrated end-to-end method that integrates data representation, model training, conditional sampling and result restoration, which can realize unified data representation, goal-oriented training of models and readable restoration of generated results, thereby improving the efficiency and engineering usability of alloy reverse engineering and solving the pain points of traditional design methods. Summary of the Invention

[0008] In view of the above problems, the present invention proposes a potential diffusion generation method for reverse design of alloy materials in order to overcome the above problems or solve at least some of the above problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a potential diffusion generation method for reverse design of alloy materials, comprising: The acquired alloy historical sample data is preprocessed, and the preprocessed alloy historical sample data is mapped to the joint latent space based on the dual-stream heterogeneous variational autoencoder to obtain the joint latent space characterization. Based on the joint latent space representation, a conditional diffusion generation model is trained in the joint latent space, and a target performance conditional vector is introduced during the training process. The conditional diffusion generation model is then jointly optimized by a retrieval enhancement mechanism and a reinforcement learning fine-tuning strategy to obtain the target generation model. Obtain the target performance conditions input by the user, perform condition sampling using the target generation model, and generate multiple candidate latent variables containing alloy composition latent variables and process latent variables; The multiple candidate latent variables are screened, and the effective candidate latent variables obtained after screening are input into the decoder to restore the alloy composition ratio and process step sequence, which serves as the output result of alloy reverse design.

[0010] Furthermore, the preprocessing of the acquired alloy historical sample data specifically includes: The structured and semi-structured data sources of alloy materials are acquired to form a unified historical sample field; the historical sample field includes alloy composition vector, process step sequence, and multi-target performance label; The historical sample fields are processed for missing values, normalized in terms of units, and truncated at the boundaries. The alloy composition vector is encoded into an alloy composition ratio vector of a preset dimension; The process step sequence is encoded into a preset process step sequence structure, which is process unit-process method-process parameter; The multi-objective performance labels are encoded into target performance condition vectors.

[0011] Furthermore, the missing value processing includes: generating a binary step mask vector for marking the validity of the process steps and a binary parameter mask vector for marking the validity of the process parameters for the process step sequence; no mask vector is generated for the alloy composition vector.

[0012] Furthermore, the dual-stream heterogeneous variational autoencoder includes a component coding branch and a process coding branch; The composition coding branch uses a fully connected neural network to encode the alloy composition ratio vector and map it to the joint latent space to form latent variables of alloy composition. The process coding branch employs slicing, embedding, and feature projection processing of the process step sequence, and combines the binary step mask vector and the binary parameter mask vector to reset the weights of invalid process step bits and invalid process parameter bits to 0, respectively. After encoding through a Transformer network, it is mapped to the joint latent space to form process latent variables. The latent variables of alloy composition and the latent variables of process together constitute a joint latent space characterization.

[0013] Furthermore, the loss function of the dual-stream heterogeneous variational autoencoder during training includes reconstruction loss and distribution constraint loss; The reconstruction loss includes component reconstruction loss and process-related reconstruction loss; The composition reconstruction loss is used to calculate the distance between the original alloy composition ratio vector and the decoded alloy composition ratio vector; The process-related reconstruction loss is used to calculate the restoration loss of process units, process methods and process parameters, and the process-related reconstruction loss is only calculated at the corresponding positions where the binary step mask vector and the binary parameter mask vector are marked as valid. The distribution constraint loss is used to constrain the distributions of the alloy composition latent variable and the process latent variable to a standard Gaussian distribution.

[0014] Furthermore, based on the joint latent space representation, a conditional diffusion generation model is trained in the joint latent space, and a target performance conditional vector is introduced during the training process. The conditional diffusion generation model is then jointly optimized using a retrieval enhancement mechanism and a reinforcement learning fine-tuning strategy to obtain the target generation model. Specifically, this includes: Construct a conditional diffusion generation model; use the binary step mask vector and binary parameter mask vector as auxiliary constraint information, and input them together with the joint latent space representation into the conditional diffusion generation model to complete model pre-training; The target performance condition vector is injected into the denoising process of the conditional diffusion generation model to establish a conditional control channel; A retrieval-enhanced generation memory is constructed. During generation, the top K prior sample latent vectors that are closest to the target performance condition vector in Euclidean distance are retrieved and incorporated into the denoising process through an attention mechanism. A multi-task performance predictor is used to predict the performance of the latent variables generated during the training of the conditional diffusion generative model, and outputs the performance estimate and the corresponding prediction uncertainty. A multi-task reward function is constructed based on the performance estimate and the corresponding prediction uncertainty, and the conditional diffusion generative model is fine-tuned using a reinforcement learning algorithm. A training stabilization strategy is adopted, and the final model parameters are selected and saved based on the performance on the validation set.

[0015] Furthermore, the multi-task reward function Represented as:

[0016]

[0017]

[0018] in, For performance bonus items; As a performance reward weight; Awarded for process feasibility; Weighting for process rewards; As a latent variable in alloy composition; As a potential variable in the process; The target performance condition vector; Performance estimates for the log-scaled space; Target performance conditions in log-scaled space; Penalize bandwidth for performance deviations; To predict the uncertainty penalty coefficient; The standard deviation of the performance estimate; This is the decoded process parameter vector; This represents the engineering feasible region for process parameters. For indicator functions; These are the nominal values ​​of the process parameters; Penalty bandwidth for deviation of process parameters; The weight of hard constraints for process compliance.

[0019] Furthermore, the step of obtaining the target performance conditions input by the user, and performing conditional sampling using the target generation model to generate multiple candidate latent variables including alloy composition latent variables and process latent variables, specifically includes: The target performance conditions input by the user are converted into a condition vector, and the condition vector is logarithmically transformed to construct the condition input; Initialize the Gaussian noise latent variables of alloy composition and process Gaussian noise, and perform a reverse time step denoising process; in each time step, simultaneously perform conditional prediction and unconditional prediction based on the condition vector, and obtain the final noise prediction value through a classifier-free fusion strategy, and update the current alloy composition latent variable and the current process latent variable according to the noise scheduler parameters, until all reverse time steps are completed, generating multiple candidate latent variables.

[0020] Furthermore, the process of filtering the multiple candidate latent variables and inputting the filtered valid candidate latent variables into the decoder to restore the alloy composition ratio and process parameter sequence as the alloy reverse design result specifically includes: The main task performance predictor is used to predict the multiple candidate latent variables to obtain the performance estimate and corresponding prediction uncertainty of each candidate latent variable; candidate latent variables with prediction uncertainty higher than the preset uncertainty quantile threshold are screened out to generate the first candidate set; Temporarily decode the candidate latent variables in the first candidate set, and construct a process quality penalty term based on structural artifacts and process parameter interval violation counts; The target performance normalization error of each candidate latent variable in the first candidate set is calculated as a base score. The process quality penalty term is weighted according to a preset weight and incorporated into the base score to obtain a comprehensive score. The candidate latent variables in the first candidate set are sorted according to the comprehensive score, and the top preset number of candidate latent variables are selected to form an effective candidate set. The valid candidate latent variables in the valid candidate set are input into the decoder, and are restored to the alloy composition ratio by the composition head and the process step sequence by the process head.

[0021] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a potential diffusion generation method for reverse design of alloy materials, which has the following beneficial effects: This invention addresses the problems of inconsistent data representation and insufficient utilization of cross-modal information in existing technologies. By combining latent space modeling with an interactive attention mechanism, it achieves deep integration of component and process information. Simultaneously, it unifies coding standards and tensor protocols to resolve the heterogeneity and inconsistency issues between data. This not only solves the problem of disconnected data processing in existing methods but also improves the utilization rate of cross-modal information, laying the foundation for the accuracy of generated samples.

[0022] This invention addresses the problems of "disconnect between model training and target constraints, and insufficient multi-objective balanced optimization" in existing technologies by implementing a full-process target-guided mechanism: Target performance conditions are injected during the model training phase, constructing a reward function that includes the achievement of multi-objective performance and process feasibility. Combined with RL fine-tuning optimization strategies, this enhances the balanced optimization capability of alloy multi-objective performance (such as strength-plasticity-brittleness). Simultaneously, a CFG fusion strategy is introduced during the sampling phase to ensure that generated samples can stably hit the target performance while achieving optimal balance of multi-objective performance, thus completely resolving the pain points of insufficient multi-objective consideration and large fluctuations in target hit rate in existing methods.

[0023] This invention addresses the problems of "missing physical logic in the model and insufficient engineering reliability of generated samples" in existing technologies. It ensures practicality through multiple constraints: a process quality penalty term is incorporated into the model training, and a priori physical logic sample reference is introduced in conjunction with the RAG retrieval enhancement mechanism to avoid generating invalid samples with contradictory process steps or mismatched composition and process; at the same time, a sample readability restoration step is added to convert the decoding results into human-readable and experimentally verifiable composition ratios and process routes without additional post-processing. This reduces the workload of R&D personnel and ensures that the output samples conform to the physical laws of actual alloy processing, thereby improving the engineering usability and practicality of the method. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the potential diffusion generation method for reverse design of alloy materials provided in an embodiment of the present invention. Detailed Implementation

[0026] 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.

[0027] This invention discloses a potential diffusion generation method for reverse design of alloy materials, such as... Figure 1 As shown, it includes the following steps: S1. The acquired alloy historical sample data is preprocessed, and the preprocessed alloy historical sample data is mapped to the joint latent space based on the dual-stream heterogeneous variational autoencoder to obtain the joint latent space characterization. S2. Based on the joint latent space representation, a conditional diffusion generation model is trained in the joint latent space. During the training process, a target performance conditional vector is introduced. The conditional diffusion generation model is then jointly optimized by a retrieval enhancement mechanism and a reinforcement learning fine-tuning strategy to obtain the target generation model. S3. Obtain the target performance conditions input by the user, perform condition sampling using the target generation model, and generate multiple candidate latent variables containing alloy composition latent variables and process latent variables; S4. Perform screening on multiple candidate latent variables, and input the effective candidate latent variables obtained after screening into the decoder to restore the alloy composition ratio and process step sequence as the output result of alloy reverse design.

[0028] The aforementioned potential diffusion generation method for reverse design of alloy materials aims to solve the technical problems of inconsistent data representation, disconnect between model training and target constraints, and lack of physical logic in existing technologies. This method is based on three main parts: data processing and characterization, model training and optimization, and sampling and generation processes. It forms a complete processing chain from raw data input to candidate sample output, and terminates the process after sample generation and composition / process reconstruction. This achieves full automation of the alloy reverse design process, eliminating the need for manual switching between multiple stages and effectively improving design efficiency.

[0029] It should be noted that the labels S1-S4 above are only for ease of explanation and do not limit the execution order of the steps. Next, each of the above steps will be explained in detail.

[0030] In step S1 above, the acquired alloy history sample data is preprocessed, and the preprocessed alloy history sample data is mapped to the joint latent space based on a two-stream heterogeneous variational autoencoder to obtain the joint latent space characterization; specifically: S11. Preprocess the acquired historical alloy sample data, specifically including: (1) Obtain structured and semi-structured data sources for alloy materials; among them, structured data sources include tabular data in the enterprise's internal experimental database and alloy industry standard database, and semi-structured data sources include technical documents and historical process ledgers in XML and JSON formats, to ensure that the data sources cover the core data in the alloy research and development and production process, and provide comprehensive data support for subsequent model training; (2) Based on the above structured data source and semi-structured data source, a unified historical sample field is formed; wherein, the historical sample field includes alloy composition vector, process step sequence and multi-target performance label; (3) Perform missing value processing, dimension normalization processing, and boundary truncation processing on historical sample fields to eliminate unusable records; specifically: Missing value handling: To avoid introducing data distribution inconsistencies, missing values ​​are not imputed. Instead, a masking mechanism is used to mark valid information bits. Specifically, for the process step sequence only, binary step mask vectors are generated to mark the validity of process steps, and binary parameter mask vectors are generated to mark the validity of process parameters. A mask value of 1 indicates that the corresponding information is valid, and a mask value of 0 indicates that the corresponding information is missing or invalid. The dimension of the mask vector is consistent with the dimension of the corresponding process coding feature. No mask vector is generated for alloy composition vectors.

[0031] Dimensional normalization: Dimensional normalization adopts minimum-maximum normalization or standard score normalization to uniformly map performance indicators and process parameters of different dimensions to the [0,1] interval, so as to avoid the interference of dimensional differences on model training.

[0032] Boundary truncation processing: Boundary truncation is carried out in accordance with the technical specifications of the alloy field to truncate abnormal data that exceeds the reasonable range. Unusable records that deviate significantly from the normal range are directly rejected.

[0033] (4) Encode the historical sample fields; specifically: The alloy composition vector is encoded as an alloy composition ratio vector with a preset dimension (e.g., 33-dimensional). Specifically, fixed dimensions are set according to the types of common alloy elements, with each dimension corresponding to a common alloy element, and the content of each element is complete. The vector value of the alloy composition ratio vector is the mass fraction or atomic fraction of the alloy element, ensuring the standardized expression of composition information. The process step sequence is encoded into a preset process step sequence structure, which is a process unit (such as melting, forging, solution treatment, aging) - process method (such as vacuum melting, hot forging, water quenching) - process parameter (such as temperature, time, speed), which facilitates the model's learning and modeling of process information. Multi-objective performance labels are encoded into target performance condition vectors.

[0034] S12. Based on a dual-flow heterogeneous variational autoencoder, the preprocessed alloy historical sample data is mapped to the joint latent space to obtain the joint latent space characterization; specifically including: (1) The dual-stream heterogeneous variational autoencoder includes a component coding branch and a process coding branch, which are encoded independently; wherein: The composition coding branch uses a fully connected neural network to encode the alloy composition ratio vector and map it to the joint latent space to form the latent variable of the alloy composition; this composition coding branch has no mask constraints throughout. The process coding branch slices, embeds, and projects the process step sequence, and combines it with binary step mask vector and binary parameter mask vector to reset the weights of invalid process step bits and invalid process parameter bits to 0. After encoding by the Transformer network, it is mapped to the joint latent space to form process latent variables. This process coding branch can ensure that the model accurately identifies the valid process information bits and avoids missing / invalid information interfering with feature extraction. The aforementioned latent variables of alloy composition and process constitute a joint latent space representation. It should be noted that although the latent variables of alloy composition and process are generated independently, joint loss optimization is required during training.

[0035] (2) The loss function of the above-mentioned dual-stream heterogeneous variational autoencoder during training includes reconstruction loss and distribution constraint loss, so that the latent vectors simultaneously retain effective information about components and processes; where: The reconstruction loss includes composition reconstruction loss and process-related reconstruction loss. The composition reconstruction loss is used to calculate the distance between the original alloy composition ratio vector and the decoded alloy composition ratio vector, without introducing mask constraints. The process-related reconstruction loss is used to calculate the restoration loss of process units, process methods, and process parameters. The process-related reconstruction loss is only calculated at the corresponding positions where the binary step mask vector and the binary parameter mask vector are marked as valid (i.e., the mask value is 1). Invalid information bits with a mask value of 0 do not participate in the loss calculation to avoid invalid information interfering with model training and to ensure that the latent vector can accurately restore the original composition and effective process features. Distribution constraint loss is used to constrain the distribution of alloy composition latent variables and process latent variables to a standard Gaussian distribution, thereby improving the continuity and sampleability of the latent space and providing a good latent space foundation for the training of subsequent generative models. Furthermore, during the training process, the gradient descent algorithm is used to iteratively optimize the parameters of the variational autoencoder until the reconstruction loss and distribution constraint loss tend to stabilize, thus obtaining the optimal latent space mapping model.

[0036] In step S2 above, based on the joint latent space representation, a conditional diffusion generation model is trained in the joint latent space. During training, a target performance conditional vector is introduced, and the conditional diffusion generation model is jointly optimized using a retrieval enhancement mechanism and a reinforcement learning fine-tuning strategy to obtain the target generation model. Specifically, this includes: S21. Construct a conditional diffusion generation model; use the binary step mask vector and binary parameter mask vector as auxiliary constraint information, and input them into the conditional diffusion generation model together with the joint latent space representation to complete the model pre-training; during the pre-training process, constrain the weight of the effective process information bits in the denoising process through the binary step mask vector and binary parameter mask vector, so that the model focuses on learning the distribution characteristics of the effective process information bits, and has the basic ability to gradually denoise from Gaussian noise and generate candidate latent vectors that conform to the latent space distribution and effective information constraints, laying the foundation for subsequent controllable generation and optimization; S22. Inject the target performance condition vector into the denoising process of the conditional diffusion generation model, establish a dedicated conditional control channel, so that the generation direction of the model can be explicitly guided by the target performance, ensure that the generated latent variables can correspond to alloy samples that meet the target performance requirements, and improve the controllability of the generation process. S23. A RAG retrieval enhancement mechanism is introduced. Based on alloy composition latent variables and process latent variables, and combined with composition, process, and performance data, a retrieval enhancement generation memory is constructed after multi-dimensional filtering. The multi-dimensional filtering includes: effective process step filtering, process method validity filtering, process parameter abundance filtering, process step logic compliance filtering, core performance threshold filtering, and parameter violation rate filtering. Based on the constructed retrieval enhancement generation memory, during generation, the top K prior sample latent vectors with the closest Euclidean distance to the target performance condition vector are retrieved. After projecting the retrieved prior sample latent vectors, an attention mechanism is used to integrate them into the denoising process to assist model generation, effectively improving the stability, process feasibility, and target fit of the generated samples. S24. Use a multi-task performance predictor to predict the performance of the latent variables generated by the conditional diffusion generation model during training, and output the performance estimate and the corresponding prediction uncertainty. Based on the performance estimate and the corresponding prediction uncertainty, construct a multi-task reward function, and use reinforcement learning algorithm to fine-tune the conditional diffusion generation model to further improve the model's target alignment ability and engineering feasibility, and ensure that the generated samples stably hit the multi-target performance requirements. The aforementioned multi-task reward function simultaneously considers the achievement of multiple objectives (i.e., the deviation between the performance estimate and the target performance condition vector), performance prediction uncertainty, and process feasibility (process step logic, parameter compliance), etc.; this multi-task reward function Represented as:

[0037]

[0038]

[0039] in, For performance bonus items; As a performance reward weight; Awarded for process feasibility; Weighting for process rewards; As a latent variable in alloy composition; As a potential variable in the process; The target performance condition vector; Performance estimates for the log-scaled space; Target performance conditions in log-scaled space; Penalize bandwidth for performance deviations; To predict the uncertainty penalty coefficient; The standard deviation of the performance estimate; This is the decoded process parameter vector; This represents the engineering feasible region for process parameters. For indicator functions; These are the nominal values ​​of the process parameters; Penalty bandwidth for deviation of process parameters; The weight of hard constraints for process compliance.

[0040] S25. Training stabilization strategies such as learning rate scheduling, gradient pruning, and parameter moving averages are employed. Combined with real-time monitoring of the model's convergence status using a validation set, the final optimized model parameters are selected and saved based on the convergence trend of the validation set loss, the target hit rate of generated samples, and stability, for use in subsequent sampling and generation processes. This training stabilization strategy effectively avoids problems such as gradient explosion and oscillations during model training, ensuring stable model convergence.

[0041] In step S3 above, the target performance conditions input by the user are obtained, and conditional sampling is performed using the target generation model to generate multiple candidate latent variables containing alloy composition latent variables and process latent variables; specifically including: S31. Convert the target performance conditions input by the user into a condition vector, and perform a logarithmic transformation on the condition vector to construct the condition input; S32. Initialize the Gaussian noise latent variables of alloy composition and process Gaussian noise, and perform the reverse time step denoising process. In each time step, perform conditional prediction and unconditional prediction based on conditional vectors, and obtain the final noise prediction value through the Classifier-Free Guidance (CFG) fusion strategy. Update the current alloy composition latent variable and the current process latent variable according to the noise scheduler parameters, until all reverse time steps are completed, and generate multiple candidate latent variables.

[0042] The core denoising formula is:

[0043] in, This is the final noise prediction value; This is the predicted value for unconditional noise. This is the predicted value for conditional noise; This is the guiding coefficient.

[0044] In step S4 above, multiple candidate latent variables are screened, and the effective candidate latent variables obtained after screening are input into the decoder to restore the alloy composition ratio and process step sequence, which serves as the alloy reverse design result; specifically including: S41. Use the main task performance predictor to predict multiple candidate latent variables and obtain the performance estimates and corresponding prediction uncertainties of each candidate latent variable. The performance estimates include yield strength (YS), tensile strength (UTS), and elongation (EL). If an impact energy (Charpy) predictor is loaded, the Charpy performance is predicted simultaneously to provide data support for subsequent screening and ranking.

[0045] S42. First-level screening, uncertainty gating: Candidate latent variables with prediction uncertainty higher than the preset uncertainty quantile threshold (usually set to 0.95) are screened out to generate the first candidate set, ensuring that the retained candidate samples have high credibility; at the same time, a fallback logic is set so that if all candidate latent variables are screened out, it will fall back to retaining all samples to avoid an empty set.

[0046] S43. Second-level screening: Temporarily decode the candidate latent variables in the first candidate set to identify structural artifacts (i.e., structured anomalies in the process sequence that do not conform to the actual processing logic, resulting in process infeasibility, such as rolling speed of 0, number of passes of 0, etc.) and violations of process parameter ranges in the process sequence. Construct a process quality penalty term based on the count of structural artifacts and violations of process parameter ranges. Calculate the target performance normalization error of each candidate latent variable in the first candidate set as the base score. Incorporate the process quality penalty term into the base score according to a preset weight to correct the base score and obtain a comprehensive score. Sort the candidate latent variables in the first candidate set according to the comprehensive score, and select the top preset number of candidate latent variables (e.g., K) to form an effective candidate set. This screening method ensures that the finally selected candidate samples have good process feasibility.

[0047] After obtaining the valid candidate set, you can choose to perform deduplication and basic feasibility screening on the valid candidate set. The basic feasibility screening includes at least field integrity checks and composition range constraint checks, and retains the valid candidate set. The deduplication process uses a vector nearest neighbor matching algorithm to calculate the cosine similarity between candidate latent vectors. Latent vectors with similarity higher than a preset threshold (usually set to 0.95) are considered duplicate vectors, and only one of them is retained. In the basic feasibility screening, the field integrity check is used to ensure that the candidate latent vectors can obtain complete composition and process information after decoding, without missing key fields. The composition range constraint check checks whether the decoded composition ratio is within a reasonable range according to the technical specifications in the alloy field, and removes candidate latent vectors with abnormal composition ratios, retaining the final valid candidate set. S44. Input the valid candidate latent variables from the valid candidate set into the decoder. The composition head is used to restore them to alloy composition ratios, and the process head is used to restore them to a sequence of process steps (including process units, process methods, and process parameters). Specifically, the decoder adopts the same network structure as the decoding end of the variational autoencoder, and restores the valid candidate latent vectors to structured composition ratios and process sequences. For continuous process parameters such as temperature, time, and speed, inverse normalization is performed to map them from the [0,1] interval back to the actual physical quantity range, ensuring the practicality of the process parameters. After decoding, the decoded component ratios and process sequences are converted into human-readable text and structured records, specifying the component strings and process JSON format data for each candidate sample. Specifically, the decoding results are mapped to human-readable sample descriptions, which include at least element content, process unit order, and key process parameters. According to predefined field templates, the decoded component ratios are converted into standard component expression strings (e.g., Ti-6.2Al-4.1V-0.2Fe-0.03C), and the process sequences are converted into clear process path descriptions, specifying the order of process units, the method of each process unit, and key parameters, ensuring that the sample descriptions are concise, standardized, and easy to understand, facilitating subsequent manual review and experimental verification.

[0048] Finally, export two files, top_candidates.csv and summary.json, to complete the reverse design results output. Top_candidates.csv records detailed information of all Top-K candidate samples, while summary.json records the core summary information of the candidate samples, which is convenient for researchers to quickly view, review and verify in subsequent experiments.

[0049] The aforementioned invention provides a potential diffusion generation method for reverse design of alloy materials. First, it unifies and optimizes multi-source alloy data, resolving issues of data fragmentation and inconsistent formats. Then, it constructs a goal-oriented generation model to ensure that generated samples meet multi-objective performance requirements. Finally, it completes a closed loop of "parameter injection - model loading - conditional sampling - attribute prediction - screening and ranking - Top-K truncation - decoding and restoration - readable output," outputting usable reverse design solutions and effectively improving the efficiency and quality of alloy reverse design. The core feature of this sampling and generation chain is that it does not directly output after sampling, but forms a complete closed loop of "sampling + gating + multi-objective ranking + feasibility rearrangement + decoding and export." Candidate screening and basic filtering are key steps, enabling the selection of candidate samples whose "appearance meets the target" as "more credible, more aligned with the target, and more feasible in terms of process." Simultaneously, the process quality penalty is not a post-hoc score but is directly integrated into the comprehensive score, having a substantial impact on the final Top-K candidate sample selection. Furthermore, the core of this process is screening and ranking, without employing explicit deduplication algorithms.

[0050] Based on the same inventive concept, embodiments of the present invention also provide a potential diffusion generation system for reverse design of alloy materials, comprising: The data processing and characterization module is used to preprocess the acquired alloy history sample data and map the preprocessed alloy history sample data to the joint latent space based on the dual-stream heterogeneous variational autoencoder to obtain the joint latent space characterization. The model training and optimization module is used to train the conditional diffusion generation model in the joint latent space based on the joint latent space representation, introduce the target performance condition vector during the training process, and combine the retrieval enhancement mechanism and reinforcement learning fine-tuning strategy to jointly optimize the conditional diffusion generation model to obtain the target generation model. The conditional sampling module is used to obtain the target performance conditions input by the user, and to perform conditional sampling using the target generation model to generate multiple candidate latent variables containing alloy composition latent variables and process latent variables. The filtering and decoding output module is used to filter multiple candidate latent variables and input the effective candidate latent variables obtained after filtering into the decoder to restore the alloy composition ratio and process step sequence as the output result of alloy reverse design.

[0051] Since the principle behind the problem solved by this diffusion generation system is similar to that of the aforementioned diffusion generation method, the implementation of this diffusion generation system can be found in the aforementioned method, and the repetitions will not be repeated.

[0052] The inventive principle and specific implementation process of the present invention will be further described in detail below with reference to specific embodiments. The various embodiments can be combined with each other to form new embodiments. All modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0053] Example 1: Complete Implementation of the End-to-End Reverse Engineering Process: This embodiment details the complete implementation process of the generation method described in this invention, covering all aspects from data input to sample output, ensuring the operability and repeatability of the method.

[0054] 1. Data processing and representation: (1) Data access: Read sample records from the materials database, which includes the company's internal experimental database, historical process ledgers, publicly available materials datasets (such as the MatWeb Materials Properties Database), and the company's internal production records, to ensure the comprehensiveness and diversity of data sources.

[0055] (2) Field uniformity: Each sample is standardized into "composition field, process field, and performance field" (i.e., gold composition vector, process step sequence, and multi-objective performance label), where: The composition field is represented in the form of "element-content pairs", clearly recording the name of each element and its corresponding mass fraction or atomic fraction, such as "Ti: balance, Al: 6.2%, V: 4.1%, Fe: 0.2%, C: 0.03%"; The process field is represented in the form of a step sequence, recording each process unit, process method and key parameters in the process execution order, such as "melting → forging → solution treatment → aging; melting method: vacuum melting, forging temperature: 1050℃, solution treatment temperature: 920℃, aging temperature: 550℃, aging time: 4h"; The performance field is represented in the form of multi-objective numerical values, including at least core performance indicators such as yield strength, tensile strength, and elongation, for example, "Yield strength: 860MPa, tensile strength: 950MPa, elongation: 15%".

[0056] At the same time, a unified field dictionary is established to map synonymous fields from different data sources in a unified manner. For example, “YS” and “yield strength” are mapped to the same target field, and “UTS” and “tensile strength” are mapped to the same target field to avoid data confusion caused by inconsistent field naming.

[0057] (3) Data cleaning: A tiered cleaning process is employed to thoroughly clean the sample data, ensuring its accuracy and usability. This process includes the following four tiers: The first layer is format validation. The type of each field is checked to ensure that numeric fields (such as component content, process parameters, and performance indicators) can be parsed as floating-point numbers or integers, and that category fields (such as process methods and cooling methods) are in a predefined thesaurus. Records that do not meet the format requirements are marked and corrected; those that cannot be corrected are removed. The second layer is boundary verification. According to the technical specifications in the alloy field, reasonable value ranges are set for key parameters. For example, the forging temperature range is set to 800-1200℃, the aging time range is set to 1-10h, and the composition content range is set according to the alloy type. For records that exceed the reasonable range, they are marked. If the deviation is within the allowable range (e.g., ±5%), it is corrected. If the deviation exceeds the allowable range, it is rejected. The third layer is consistency verification. The total component sum is verified. If the difference between the total component sum and the theoretical sum (usually 100%) exceeds a threshold (e.g., ±0.5%), normalization correction is performed to ensure the total component sum is 100%. If duplicate and conflicting records appear in the same sample, a strategy of "source priority (enterprise experimental database > industry standard database > academic paper) + time priority (latest record preferred)" is used to select and retain the best record. The fourth layer involves missing information handling. Samples with severely missing essential fields such as components and core performance indicators, as well as completely empty process sequences, are directly removed to avoid affecting model training performance. For samples with missing non-essential fields such as individual steps or process parameters in the process sequence, a masking mechanism is used to mark the valid information bits.

[0058] In addition, duplicate records are merged and deduplicated, and units from different sources are standardized. For example, temperature units are standardized to degrees Celsius (°C), time units to minutes (min) or hours (h), and component units to mass fraction (%). This ensures that the same field is trained under the same unit and avoids interference from differences in units on model training.

[0059] (4) Data encoding A targeted coding strategy is employed to convert composition, process, and performance data into vector or sequence forms that the model can compute, as follows: Composition coding: The composition information is mapped to a fixed-length vector. Each dimension of the vector corresponds to a common element in the alloy (such as Ti, Al, V, Fe, C, etc.), and the vector value is the mass fraction of that element. For elements not included, the vector value is set to 0 to ensure the standardized expression of composition information and facilitate unified processing by the model. Process coding: Map the process flow to a combination of "process unit, method label, and parameter vector". The process unit uses one-hot coding, the method label uses embedded coding, and the parameter vector uses a standardized numerical vector. The coding results of each process step are concatenated in sequence to form a process sequence tensor. The sequence dimension corresponds to the number of process steps, and the feature dimension includes process unit coding, method coding, and parameter coding. Performance encoding: The target performance is mapped to a condition vector. After standardizing each performance metric, the vectors are concatenated to form a fixed-dimensional condition vector, which is used to guide the target of the subsequent model.

[0060] The encoded sample data can be directly input into the neural network, providing a foundation for subsequent latent space construction and model training.

[0061] (5) Latent space representation: A dual-stream heterogeneous variational autoencoder (Hetero-VAE) is used to independently compress the encoded component features and process features, constructing two independent low-dimensional latent spaces. Joint optimization is then performed during the loss function calculation stage. The specific process is as follows: First, the encoded component features and process features are processed using independent encoding methods: the component features are input into the component stream encoding end and encoded using an MLP fully connected structure, with no mask constraints throughout the process, fully preserving the 33-dimensional component information; the process features are input into the process stream encoding end and encoded using a Transformer structure, combined with a process mask to mask invalid information bits, ensuring that only valid process information participates in the encoding; the two encoding processes are independent of each other, obtaining independent representations of the component features and process features respectively.

[0062] Then, the two independent encoders output their respective mean vectors and variance vectors, and the component latent vectors are obtained by sampling them separately using reparameterization techniques. and process potential vector The decoding end also adopts a two-channel independent structure, with the split-stream decoding end only receiving... It also restores the component characteristics, and the process flow decoding end only receives... The process characteristics are restored; the training constraints of the two latent vectors are achieved through a joint loss function, that is, the independent losses of the component flow and the process flow are calculated separately and then weighted and summed to complete the joint training of the model, ensuring that the two latent vectors retain the core information of their respective features.

[0063] Finally, the latent vectors output from the encoder are input to the decoder of the two-stream variational autoencoder. The decoder adopts a two-stream structure that corresponds one-to-one with the encoder to restore the component features and process sequence features respectively. The entire training process is constrained by the process mask mechanism generated above, which constrains the encoding, decoding and loss calculation process. The detailed implementation method is as follows: 1) Mask pre-constraint in the encoding stage: For the process flow encoding branch, the process step mask and the process sequence encoding features are weighted bit by bit. The feature weight corresponding to invalid process steps with a mask value of 0 is reset to 0, ensuring that invalid process information and missing parameter information do not participate in the latent vector encoding process. Only the features of valid process steps are compressed into the latent space, avoiding data distribution deviations introduced by missing values ​​from the source of encoding. The component flow encoding branch has no mask constraint and fully retains all information of the 33-dimensional fixed component vector to participate in the latent space encoding, ensuring the integrity of component information.

[0064] 2) Precise masking constraints for multi-task loss during the decoding phase: During training, a multi-task joint loss function is used, combined with a masking mechanism to weight the loss of each branch bit by bit, specifically including: Component reconstruction loss: Directly calculate the mean square error between the original 33-dimensional fixed component vector and the component vector output by the decoder, without introducing mask constraints throughout the process, and fully constrain the accuracy of component information reconstruction; Process unit classification loss and process method classification loss: The cross-entropy loss function is used. The loss value is calculated only for the valid steps with a mask of 1. The loss weight of the invalid steps with a mask of 0 is reset to 0 and they do not participate in the loss calculation and gradient backpropagation. Process label BCE loss: The binary cross-entropy loss function is used. First, the effective process steps are filtered out by the process step mask, and then the loss is calculated only for the effective label bits where the parameter mask is 1. Process parameter MSE loss: The mean square error loss function is adopted. First, the effective process steps are filtered out by the process step mask. Then, the loss is calculated only for the effective parameter bits with a parameter mask of 1. Missing parameter bits with a mask of 0 are not included in the loss calculation and model parameter update. 3) Cooperative optimization of distribution constraint loss: Simultaneously optimize the KL divergence distribution constraint loss to constrain the distribution of latent vectors to near the standard Gaussian distribution, thereby improving the continuity and sampleability of the latent space; During training, the AdamW optimizer is used to iteratively optimize the joint loss until the losses of both the training set and the validation set tend to stabilize and there is no overfitting trend, thus obtaining the optimal latent space mapping model.

[0065] 2. Generative model training and optimization: This step strictly follows the process of "Conditional Diffusion Model Construction and Pre-training → Target Condition Injection and Controllable Generation → RAG Retrieval Enhancement → RL Fine-tuning → Training Stabilization and Model Optimization" to ensure that the generated model has controllability, stability, and target alignment capabilities, as detailed below: (1) Construction and pre-training of the conditional diffusion model: Based on the joint latent space, a conditional diffusion generation model is constructed. The model adopts a two-stream structure, modeling component latent vectors and process latent vectors separately to ensure that the component-process coupling relationship can be fully captured. In the model pre-training stage, the process mask vector corresponding to the sample is used as auxiliary constraint information and input into the denoising network along with the latent vector. The weight of the effective process information bits in the denoising process is constrained by the process mask vector, so that the model focuses on learning the distribution characteristics of the effective process information bits. The pre-training process does not introduce target condition constraints, but only trains the model's ability to gradually denoise from Gaussian noise and restore the true latent vector based on the distribution of real samples in the latent space. This enables the model to have basic latent vector generation capabilities. The pre-training process focuses on optimizing the denoising loss combined with mask constraints to ensure that the model can learn the effective information distribution characteristics of the latent space, laying the foundation for subsequent controllable generation and optimization.

[0066] (2) Target condition injection and controllable generation mechanism: The user-defined multi-objective performance conditions (such as yield strength, tensile strength, elongation, etc.) are encoded and converted into fixed-dimensional condition vectors, which are then injected into the denoising network of the diffusion model through a conditional embedding layer. A dedicated conditional control channel is established so that the model can receive target condition guidance in each step of the denoising process, clarify the generation direction, and achieve explicit control of the generation process. This ensures that the generated latent vectors can correspond to alloy samples that meet the target performance requirements, reduce generation randomness, and improve target orientation.

[0067] (3) Access to RAG search enhancement mechanism: During the model generation phase, the RAG (Retrieval Enhanced Generation) mechanism is integrated, calling upon the high-quality latent vector memory built in the early stages. For the target performance conditions of the current input, the latent vector of the most similar prior sample in the memory is retrieved and incorporated as reference information into the diffusion denoising process to assist the model in noise prediction and latent vector generation. By introducing similar prior sample references, the stability and process feasibility of the generated samples are effectively improved, reducing the generation of samples that do not conform to physical logic or process violations, while further improving the fit between the generated samples and the target performance.

[0068] (4) RL fine-tuning (multi-task reward function): Based on the output of a multi-task performance predictor, a multi-task reward function is constructed, incorporating multi-objective performance achievement and process feasibility into a unified reward evaluation system. Multi-objective performance achievement is quantified by calculating the deviation between predicted and target performance; the smaller the deviation, the higher the reward value. Process feasibility is quantified by temporarily decoding latent vectors and checking the compliance of process step logic and parameters; higher compliance results in a higher reward value. Based on this reward function, a reinforcement learning algorithm is used to fine-tune the diffusion model. The reward signal guides the model to adjust parameters, further improving the model's target alignment ability and ensuring that generated samples consistently meet multi-objective performance requirements while also considering process feasibility.

[0069] (5) Training stabilization and model optimization: A series of training stabilization strategies are employed to ensure stable model convergence: Learning rate scheduling uses a "warm-up + segmented decay" model, with a small learning rate for initial warm-up and then a segmented reduction in the learning rate based on loss convergence during the middle stages; gradient thresholds are set, and gradients exceeding the threshold are pruned to prevent gradient explosion; moving averages are applied to model parameters to improve inference stability; mixed precision training is used to improve training efficiency while maintaining accuracy. Simultaneously, the model's convergence status, target hit rate, and stability are monitored in real-time using a validation set. When the validation set loss stabilizes and the target hit rate reaches a preset threshold, the final optimized model parameters are selected and saved for subsequent sampling and generation processes.

[0070] 3. Sampling and Generation Process: (1) Parameter and target injection: Global configuration initialization is completed by entering parameters through the command line. Configuration parameters include, but are not limited to: target performance (such as yield strength, tensile strength, elongation, etc.), sampling scale (number of candidate latent vectors generated), top-K number (number of candidate samples in the final output), uncertainty quantile (used for gating screening), whether to enable Pareto sorting, and process quality weight (used for penalty term weighting).

[0071] (2) Model loading: Load all trained models and components, including: Variational Autoencoder (VAE) for decoding and restoring candidate latent vectors; Diffusion Model for latent vector generation during the conditional sampling phase; Noise Scheduler for controlling the noise addition and denoising rhythm during the diffusion process; Main Task Predictor for predicting the core properties of candidate latent vectors, such as yield strength (YS), tensile strength (UTS), and elongation (EL); If the study includes Charpy performance prediction, load the Charpy predictor simultaneously.

[0072] (3) Conditional sampling: First, the target performance value input by the user is converted into a standard condition vector. A logarithmic transformation is then performed on the condition vector to construct a conditional input that meets the model's input requirements. Then, two Gaussian noise latent variables z are initialized. c z p The initial input for diffusion denoising is used; subsequently, a reverse time-step denoising process is performed. In each time step, conditional prediction (combined with target performance conditions) and unconditional prediction (without combining target performance conditions) are performed simultaneously. The dual-branch prediction results are fused using the CFG (Classifier-Free Guidance) fusion strategy to obtain the final noise prediction value, which is used to adjust the strength of conditional guidance. After each denoising step, the latent variables are updated according to the preset parameters of the noise scheduler until all reverse time steps are completed, generating a batch of candidate latent vectors.

[0073] (4) Candidate attribute prediction: For all candidate latent vectors obtained from conditional sampling, the main task predictor batches and predicts their corresponding core performance indicators, including yield strength (YS), tensile strength (UTS), and elongation (EL). If a Charpy predictor is loaded, the Charpy performance of the candidate latent vectors is predicted simultaneously. At the same time, an evaluation score is constructed, and factors such as target performance error and process feasibility are considered to provide a quantitative basis for subsequent candidate selection and ranking.

[0074] (5) Candidate screening and basic filtering: This step is the core of the sampling and generation process. It is used to filter candidate samples that "meet the target in appearance" into candidate samples that are "more credible, more closely aligned with the target, and more feasible in terms of process." It is executed in two layers: The first layer involves uncertainty gating and pre-screening: First, multi-objective performance tolerance hard filtering is performed, eliminating candidate samples whose target performance deviation exceeds the tolerance range based on a preset performance tolerance threshold. Then, based on a globally configured uncertainty quantile threshold, the performance prediction uncertainty of the remaining candidate latent vectors is evaluated, eliminating candidate latent vectors with uncertainties higher than the threshold, and retaining a candidate set with lower uncertainty and higher confidence. A fallback logic is also included: if the candidate set is empty after screening, it automatically reverts to retaining all candidate latent vectors to avoid process interruption.

[0075] The second layer, sorting and filtering: Further optimization and filtering are performed on the candidate set retained after the initial screening. First, a comprehensive score is constructed, calculated with the target performance normalized error as the core. Optional process quality constraints are implemented. First, the candidate latent vectors are temporarily decoded, and hard filtering is performed through process step logic verification to remove invalid samples whose process step order violates the processing logic. Then, structural artifacts (i.e., invalid core process parameter values, such as rolling speed of 0, pass number of 0, etc.) and violations of process parameter ranges are identified in the process sequence. Violation counts are statistically analyzed, and penalty terms are constructed. These penalty terms are weighted and incorporated into the comprehensive score according to preset process quality weights to correct the comprehensive score. Then, a sorting strategy is selected based on the global configuration: if Pareto sorting is enabled and the Charpy predictor is loaded, the candidate samples are first non-dominated sorted using the normalized target performance error as the optimization dimension, and then lexicographically sorted by the comprehensive score; if Pareto sorting is not enabled or the Charpy predictor is not loaded, only the comprehensive score is sorted in descending order.

[0076] (6) Top-K cutoff: For the sorted candidate latent vectors, based on the globally configured top-K number, the top K candidate latent vectors are truncated as the effective candidate set, and candidate samples ranked lower are removed to ensure that the final output candidate samples are the optimal and most closely related to the target. The truncation process strictly follows the sorting result and does not change the relative order of the candidate samples.

[0077] (7) Decoding and restoration: The Top-K valid candidate latent vectors are input into the trained decoder (with the same structure as the variational autoencoder decoder) for batch decoding and reconstruction. This process consists of two parts: first, decoding through the component head restores the latent vectors to standardized component ratios, ensuring that the total composition conforms to alloy industry standards; second, decoding through the process head restores the latent vectors to structured process sequences, including process units, process methods, and key process parameters. Continuous process parameters such as temperature and time undergo inverse normalization, mapping them from the [0,1] interval back to the actual physical quantity range; discrete process parameters such as process methods and cooling methods are reverse-looked up using a category mapping table, converting the encoded values ​​output by the model into corresponding textual descriptions to ensure the practicality and standardization of the decoding results.

[0078] (8) Sample readability and output: Sample readability: The decoded and restored component ratios and process sequences are converted into a human-readable format. The component ratios are converted into standard alloy composition strings (such as Ti-6.2Al-4.1V-0.2Fe-0.03C), and the process sequences are converted into JSON-formatted structured data. At the same time, the performance prediction value and uncertainty of each candidate sample are associated. If Pareto sorting is enabled, the Pareto level is recorded synchronously to form a complete and clear sample record, which is convenient for R&D personnel to review and evaluate.

[0079] Output: Export two files in a preset format: top_candidates.csv and summary.json. top_candidates.csv records detailed information for all Top-K candidate samples, including component strings, process JSON data, performance predictions, uncertainties, and Pareto grades (if applicable). summary.json records core statistical information for the candidate samples, such as the total number of candidate samples, the number of Top-K samples, the mean performance prediction, and the process compliance rate, providing a reference for subsequent experimental verification and scheme optimization. This completes the entire closed loop of "parameter injection - model loading - conditional sampling - attribute prediction - screening and sorting - Top-K truncation - decoding and restoration - readable output".

[0080] Example 2: Modular System Implementation Corresponding to the above generation method, this embodiment provides a modular system implementation scheme, which divides the system into four collaborative modules, facilitating independent upgrades of each sub-module while maintaining consistency throughout the entire process, as detailed below: 1. Data Processing and Representation Module: This module is responsible for accessing historical alloy data, standardizing fields, cleaning data, encoding, and representing latent space data. It transforms raw data into joint representation inputs that can be trained by the model, providing high-quality data support for subsequent model training. This module can independently upgrade data cleaning rules, encoding strategies, and latent space construction methods to adapt to the needs of different data sources and data scales.

[0081] 2. Model Training and Optimization Module: Responsible for training and optimizing the latent space generation model (conditional diffusion model), introducing a multi-task performance predictor for joint constraints to ensure that the model can generate candidate latent vectors that meet the target performance requirements; this module can independently upgrade the model structure, training strategy and predictor algorithm to improve the generation effect and stability of the model.

[0082] 3. Conditional sampling module: responsible for receiving the target performance conditions input by the user and performing conditional sampling to generate candidate latent vectors.

[0083] 4. The screening and decoding output module is responsible for deduplicating candidate latent vectors and performing basic feasibility screening, retaining the valid candidate set. This module can independently adjust sampling parameters, deduplication thresholds, and screening rules to balance sample diversity and effectiveness. It also decodes valid candidate latent vectors back into component ratios and process sequences, converting them into human-readable sample descriptions and outputting candidate sample results. This module can independently optimize the decoding algorithm and sample output template to improve the readability and usability of the results.

[0084] The modules communicate with each other through a unified data interface, which includes: a representation interface (outputting standardized samples and latent vectors), a model interface (outputting generated candidate latent vectors), and a reconstruction interface (outputting readable composition and process samples), ensuring smooth and consistent data transmission between modules.

[0085] Example 3: Specific application case of multi-objective alloy reverse design: Based on the above-mentioned end-to-end method, this embodiment provides a complete set of alloy reverse design application examples, which clearly demonstrate the implementation effect of the entire process from target performance input to final composition and process output, and verify the feasibility and effectiveness of this method.

[0086] 1. Input performance requirements (reverse design goals): The four target alloy performance requirements set in this case are as follows: (1) Yield strength YS = 1200 MPa (2) Tensile strength UTS = 1280 MPa (3) Elongation after fracture EL = 12% (4) Charpy impact work = 40 J 2. Generation Process (from Requirements to Candidates): Based on the aforementioned performance targets, reverse engineering is performed using the end-to-end method of this invention. The core process is as follows: (1) Map the target performance to the scaled_log condition space as the diffusion sampling condition vector.

[0087] (2) Perform 100 steps of back sampling on the two-stream diffusion model to generate 10,000 initial latent vector candidates.

[0088] (3) By using the main performance proxy model and the Charpy proxy model, the performance mean and uncertainty of all candidates are predicted, and filtered according to the preset uq quantile threshold, and finally 5003 valid candidates are retained.

[0089] (4) Pareto non-dominated sorting + process quality penalty is used to rearrange the retained set to finally select the top_k=100 best candidates.

[0090] (5) Use the trained Hetero-VAE to batch decode the top_k candidates and export the standardized component strings and process JSON structured data.

[0091] 3. Optimal candidate sample: (1) Sample A: High strength and toughness route of TC4 (Ti-6Al-4V) microalloying: 1) Candidate component ratio: Ti(Bal)-5.61Al-3.51V-1.39Mo-0.45Cr-0.33Fe-0.28Zr-0.27Sn-0.03Mn-0.01W; 2) Performance prediction results: Yield strength YS_pred = 1218.28 MPa (target value 1200 MPa); Tensile strength UTS_pred = 1276.05 MPa (target value 1280 MPa); Post-fracture elongation EL_pred = 12.13% (target value 12%); Charpy impact energy Charpy_pred = 41.30 J (target value 40 J); This sample is a qualified candidate that hits all four targets simultaneously, fully meeting the preset reverse engineering design objectives.

[0092] 3) Corresponding process path (decoding output): Powder metallurgy: temperature 1624 K, holding time 906.9 s, vacuum degree 9.382e-05 Pa, forming pressure 6.851e+08 Pa; Multi-directional forging: forging temperature 1153 K, deformation amount 2.721, strain rate 0.02831 s^-1, number of passes 7.378; Solution treatment: solution temperature 1260 K, holding time 1407 s, cooling rate 0.0195 K / s; Two-stage aging process: Stage 1: 881.3 K / 4.121e+04 s; Stage 2: 937.2 K / 1.129e+05 s; 4.) Project Interpretation: Compositional aspect: This candidate is a microalloying design in the neighborhood of the classic Ti-6Al-4V system. By adding trace amounts of Mo, Cr, Zr and Sn, it can ensure high strength while taking into account plasticity and toughness targets. This is in line with the industry general rules of alloy microalloying design and has the basis for experimental verification.

[0093] On the process side: The process path strictly follows the main chain of alloy processing, namely "powder metallurgy forming → plastic processing → solution treatment → aging treatment". The heat treatment sequence conforms to industry standards, there is no logical contradiction, and all process parameters are within the reasonable range of alloy processing, which is feasible for engineering.

[0094] Screening side: The sample comes from the Pareto optimal frontier after uncertainty filtering (Pareto_rank=0), which fully reflects the joint screening mechanism of "multi-objective performance hit + prediction risk constraint + process feasibility guarantee" of this invention, and verifies the engineering practicality of this method.

[0095] (2) Sample B: Low-alloy simplified formulation (TC4 neighborhood "slight microalloying"): 1) Candidate component ratio: Ti(Bal)-5.60Al-3.76V-0.31Mo-0.17Fe-0.11Cr-0.06Sn-0.04Zr; 2) Performance prediction results: Yield strength YS_pred = 1185.61 MPa (target 1200); Tensile strength UTS_pred = 1245.55 MPa (target 1280); Post-fracture elongation EL_pred = 11.57% (Target 12); Charpy impact energy Charpy_pred = 40.99 J (target 40); Conclusion: The simultaneous hit on all four targets demonstrates the reachability of "a small number of enhancement elements".

[0096] 3) Corresponding process path (decoding output): Powder metallurgy: temperature 1118 K, time 7608 s, vacuum degree 0.004725 Pa, pressure 3.32e+07 Pa; Annealing: temperature 1278 K, time 7403 s, cooling rate 0.01815 K / s; Two-stage lag: 1081 K / 1.437e+04 s + 1049 K / 1.087e+04 s; 4) Project Interpretation: On the composition side: the total amount of strengthening elements in this sample is relatively low, which is closer to the simplified modification route of the classic Ti-6Al-4V, making it easier to establish a low-cost control group.

[0097] From a process perspective: the process is short, the steps are clear, and the threshold for experimental implementation is low, making it suitable as a starting point for "rapid verification".

[0098] Comparison side: The impact of "belonging to the same TC4 neighborhood but with different enhancement depths" on the performance window can be compared with sample A.

[0099] (3) Sample C: Co-strengthened by β-rich stabilizing elements and multi-element microalloying (containing Cu / Nb): 1) Candidate component ratio: Ti(Bal)-5.17Al-4.32V-2.50Mo-0.97Cr-0.65Sn-0.60Zr-0.50Fe-0.06W-0.03Mn-0.03Cu-0.02Nb; 2) Performance prediction results: Yield strength YS_pred = 1200.57 MPa (target 1200); Tensile strength UTS_pred = 1285.20 MPa (target 1280); Post-fracture elongation EL_pred = 11.29% (Target 12); Charpy impact energy Charpy_pred = 40.21 J (target 40); Conclusion: All four targets were hit simultaneously, and the composite microalloy exhibited the characteristics of coupled strengthening.

[0100] 3) Corresponding process path (decoding output): Rolling: Rolling temperature 668.8 K, finishing temperature 736.4 K, rolling speed 0.9535 m / s, total reduction 0.8227 Annealing: Temperature 1180 K, Time 1729 s, Cooling rate 0.5649 K / s Annealing: Temperature 1127 K, Time 2799 s, Cooling rate 0.01097 K / s 4) Project Interpretation: On the compositional side: Mo / Cr / V are significantly increased, and trace elements Cu / Nb are introduced. The compositional style is significantly different from the previous two examples, making it suitable for verifying multi-element coupling strengthening pathways.

[0101] On the process side: the high-pressure reduction rolling + double annealing route emphasizes microstructure control and can be regarded as a representative of the "deformation-dominated" process paradigm.

[0102] Comparison side: Suitable for a horizontal comparison with sample B on the theme of "low alloy simplification vs. multi-element strengthening".

[0103] (4) Sample D: High-strength near-β multi-component system (high Mo / V / Cr / Zr): 1) Candidate component ratio: Ti(Bal)-5.01Al-3.60Mo-3.09V-1.19Cr-0.71Fe-0.54Sn-0.53Zr-0.02Mn-0.02W-0.02Cu-0.02Nb-0.01Si; 2) Performance prediction results: Yield strength YS_pred = 1282.12 MPa (target 1200); Tensile strength UTS_pred = 1340.23 MPa (target 1280); Post-fracture elongation EL_pred = 11.71% (Target 12); Charpy impact energy Charpy_pred = 39.78 J (target 40); Conclusion: All four targets were hit simultaneously, and the intensity index was significantly higher than the target value.

[0104] 3) Corresponding process path (decoding output): Forging: Forging temperature 1317 K, strain rate 0.0524 s^-1, number of passes 6.91; Rolling: Rolling temperature 1300 K, finishing temperature 896.3 K, number of passes 2.779, total reduction 0.4282; Solution treatment: temperature 1033 K, time 2941 s, cooling rate 0.02333 K / s; Two-stage slack time: 983.1 K / 6158 s + 978.3 K / 1.003e+05 s; 4) Project Interpretation: Compositional aspects: This sample is significantly biased towards a near-β high-strength system, with both the types and contents of alloying elements being higher than those of the aforementioned samples, representing the direction of "exploring the upper limit of high strength".

[0105] On the process side: the combined forging and rolling process, solution treatment, and long aging process aligns with common microstructural strengthening strategies for high-strength alloys.

[0106] The contrast side shows the greatest difference from the low-alloy system of sample B, which can be used to construct a "composition complexity-performance gain" curve.

[0107] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the aforementioned potential diffusion generation method for reverse design of alloy materials. The computer-readable storage medium can be a common storage medium such as a read-only memory (ROM), random access memory (RAM), hard disk, optical disk, or USB flash drive. The computer program stored in it can be called and executed by a processor to implement the diffusion generation method of the present invention, which facilitates the promotion and application of the method.

[0108] Based on the same inventive concept, this embodiment of the invention also provides a processing terminal, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to realize the above-mentioned potential diffusion generation method for reverse design of alloy materials; the processing terminal can be a server, desktop computer, laptop computer or other device with data processing capabilities, the memory is used to store the computer program and related data, and the processor is used to execute the computer program to realize the whole-link generation of alloy reverse design and meet the application needs in different scenarios.

[0109] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0110] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A potential diffusion generation method for reverse design of alloy materials, characterized in that, include: The acquired alloy historical sample data is preprocessed, and the preprocessed alloy historical sample data is mapped to the joint latent space based on the dual-stream heterogeneous variational autoencoder to obtain the joint latent space characterization. Based on the joint latent space representation, a conditional diffusion generation model is trained in the joint latent space, and a target performance conditional vector is introduced during the training process. The conditional diffusion generation model is then jointly optimized by a retrieval enhancement mechanism and a reinforcement learning fine-tuning strategy to obtain the target generation model. Obtain the target performance conditions input by the user, perform condition sampling using the target generation model, and generate multiple candidate latent variables containing alloy composition latent variables and process latent variables; The multiple candidate latent variables are screened, and the effective candidate latent variables obtained after screening are input into the decoder to restore the alloy composition ratio and process step sequence, which serves as the output result of alloy reverse design.

2. The potential diffusion generation method for reverse design of alloy materials as described in claim 1, characterized in that, The preprocessing of the acquired historical alloy sample data specifically includes: The structured and semi-structured data sources of alloy materials are acquired to form a unified historical sample field; the historical sample field includes alloy composition vector, process step sequence, and multi-target performance label; The historical sample fields are processed for missing values, normalized in terms of units, and truncated at the boundaries. The alloy composition vector is encoded into an alloy composition ratio vector of a preset dimension; The process step sequence is encoded into a preset process step sequence structure, which is process unit-process method-process parameter; The multi-objective performance labels are encoded into target performance condition vectors.

3. The potential diffusion generation method for reverse design of alloy materials as described in claim 2, characterized in that, The missing value processing includes: generating a binary step mask vector for marking the validity of the process steps and a binary parameter mask vector for marking the validity of the process parameters for the process step sequence; no mask vector is generated for the alloy composition vector.

4. The potential diffusion generation method for reverse design of alloy materials as described in claim 3, characterized in that, The dual-stream heterogeneous variational autoencoder includes a component coding branch and a process coding branch; The composition coding branch uses a fully connected neural network to encode the alloy composition ratio vector and map it to the joint latent space to form latent variables of alloy composition. The process coding branch employs slicing, embedding, and feature projection processing of the process step sequence, and combines the binary step mask vector and the binary parameter mask vector to reset the weights of invalid process step bits and invalid process parameter bits to 0, respectively. After encoding through a Transformer network, it is mapped to the joint latent space to form process latent variables. The latent variables of alloy composition and the latent variables of process together constitute a joint latent space characterization.

5. The potential diffusion generation method for reverse design of alloy materials as described in claim 4, characterized in that, The loss function of the two-stream heterogeneous variational autoencoder during training includes reconstruction loss and distribution constraint loss; The reconstruction loss includes component reconstruction loss and process-related reconstruction loss; The composition reconstruction loss is used to calculate the distance between the original alloy composition ratio vector and the decoded alloy composition ratio vector; The process-related reconstruction loss is used to calculate the restoration loss of process units, process methods and process parameters, and the process-related reconstruction loss is only calculated at the corresponding positions where the binary step mask vector and the binary parameter mask vector are marked as valid. The distribution constraint loss is used to constrain the distributions of the alloy composition latent variable and the process latent variable to a standard Gaussian distribution.

6. The potential diffusion generation method for reverse design of alloy materials as described in claim 3, characterized in that, The step involves training a conditional diffusion generation model within the joint latent space representation, introducing a target performance conditional vector during training, and jointly optimizing the conditional diffusion generation model using a retrieval enhancement mechanism and a reinforcement learning fine-tuning strategy to obtain the target generation model. Specifically, this includes: Construct a conditional diffusion generation model; use the binary step mask vector and binary parameter mask vector as auxiliary constraint information, and input them together with the joint latent space representation into the conditional diffusion generation model to complete model pre-training; The target performance condition vector is injected into the denoising process of the conditional diffusion generation model to establish a conditional control channel; A retrieval-enhanced generation memory is constructed. During generation, the top K prior sample latent vectors that are closest to the target performance condition vector in Euclidean distance are retrieved and incorporated into the denoising process through an attention mechanism. A multi-task performance predictor is used to predict the performance of the latent variables generated during the training of the conditional diffusion generative model, and outputs the performance estimate and the corresponding prediction uncertainty. A multi-task reward function is constructed based on the performance estimate and the corresponding prediction uncertainty, and the conditional diffusion generative model is fine-tuned using a reinforcement learning algorithm. A training stabilization strategy is adopted, and the final model parameters are selected and saved based on the performance on the validation set.

7. The potential diffusion generation method for reverse design of alloy materials as described in claim 6, characterized in that, The multi-task reward function Represented as: in, For performance bonus items; As a performance reward weight; Awarded for process feasibility; Weighting for process rewards; As a latent variable in alloy composition; As a potential variable in the process; The target performance condition vector; Performance estimates for the log-scaled space; Target performance conditions in log-scaled space; Penalize bandwidth for performance deviations; To predict the uncertainty penalty coefficient; The standard deviation of the performance estimate; This is the decoded process parameter vector; This represents the engineering feasible region for process parameters. For indicator functions; These are the nominal values ​​of the process parameters; Penalty bandwidth for deviation of process parameters; The weight of hard constraints for process compliance.

8. The potential diffusion generation method for reverse design of alloy materials as described in claim 1, characterized in that, The process involves obtaining the target performance conditions input by the user, performing condition sampling using the target generation model, and generating multiple candidate latent variables that include alloy composition latent variables and process latent variables, specifically including: The target performance conditions input by the user are converted into a condition vector, and the condition vector is logarithmically transformed to construct the condition input; Initialize the Gaussian noise latent variables of alloy composition and process Gaussian noise, and perform a reverse time step denoising process; in each time step, simultaneously perform conditional prediction and unconditional prediction based on the condition vector, and obtain the final noise prediction value through a classifier-free fusion strategy, and update the current alloy composition latent variable and the current process latent variable according to the noise scheduler parameters, until all reverse time steps are completed, generating multiple candidate latent variables.

9. The potential diffusion generation method for reverse design of alloy materials as described in claim 1, characterized in that, The process of filtering the multiple candidate latent variables and inputting the filtered valid candidate latent variables into the decoder to restore the alloy composition ratio and process parameter sequence as the alloy reverse design result specifically includes: The main task performance predictor is used to predict the multiple candidate latent variables to obtain the performance estimate and corresponding prediction uncertainty of each candidate latent variable; candidate latent variables with prediction uncertainty higher than the preset uncertainty quantile threshold are screened out to generate the first candidate set; Temporarily decode the candidate latent variables in the first candidate set, and construct a process quality penalty term based on structural artifacts and process parameter interval violation counts; The target performance normalization error of each candidate latent variable in the first candidate set is calculated as a base score. The process quality penalty term is weighted according to a preset weight and incorporated into the base score to obtain a comprehensive score. The candidate latent variables in the first candidate set are sorted according to the comprehensive score, and the top preset number of candidate latent variables are selected to form an effective candidate set. The valid candidate latent variables in the valid candidate set are input into the decoder, and are restored to the alloy composition ratio by the composition head and the process step sequence by the process head.

Images