Method, device and equipment for predicting performance of rubber formulation, and storage medium

Abstract

The application discloses a kind of rubber formula performance prediction method, device, equipment and storage medium, the rubber formula performance prediction method includes: student large model is obtained by training using large-scale sample set in advance respectively, and at least one teacher small model is obtained by training using small-scale sample set;Wherein, the sample quantity in large-scale sample set is greater than the sample quantity of small-scale sample set;And each group of sample data in large-scale sample set and small-scale sample set includes formula sample, process condition sample and performance sample;Student large model is trained using teacher small model to reverse distillation, and obtains rubber performance prediction model;The rubber formula to be predicted is input into rubber performance prediction model, and performance prediction result is obtained.In the present application, based on the rubber performance prediction model obtained by pre-distillation training, the expected performance of the rubber formula can be quickly evaluated without relying on a large number of physical experiments, the development cycle of the rubber product is shortened, and the development cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of rubber technology, and in particular to a method, apparatus, device, and computer-readable storage medium for predicting the performance of rubber formulations. Background Technology

[0002] In the rubber industry, formulation design is a core element determining product performance, and its efficiency and accuracy directly impact a company's R&D cycle and cost control. Traditional rubber formulation design heavily relies on expert experience and extensive trial-and-error experiments, resulting in long development cycles, high costs, and a strong dependence on the experience level of formulation designers. In recent years, with the rapid development of computer technology, applying it to the preliminary prediction of rubber formulation performance has been significant in shortening the development cycle and reducing costs, reducing reliance on physical experiments. However, this also depends on the computer's learning and analysis of large-scale historical rubber formulation and performance data to establish a mapping relationship between the two, thereby enabling the prediction of the performance corresponding to the rubber formulation.

[0003] However, in practical industrial applications, predicting rubber formulation performance using computer technology still faces numerous challenges. On the one hand, the large-scale historical formulation performance data accumulated in industrial scenarios are often collected under conditions of severe lack or high heterogeneity in process conditions, with only a small, strictly controlled dataset available under specific processes. On the other hand, there are significant differences in the combination of test conditions, aging conditions, and test temperatures corresponding to different processes. Traditional methods often model each process separately, making it difficult to effectively transfer and share knowledge between data from different processes. Therefore, establishing a reliable rubber formulation performance prediction model under the simultaneous existence of large-scale incomplete process condition data and small-scale, precisely recorded process data, thereby achieving accurate and reliable prediction of rubber formulation performance, is crucial for reducing the cost and cycle time of rubber product development. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, device, and computer-readable storage medium for predicting the performance of rubber formulations, which can achieve more accurate and reliable prediction of the corresponding performance of rubber formulations, and provide reliable support for reducing the R&D cost of rubber products and improving R&D efficiency.

[0005] To address the aforementioned technical problems, this invention provides a method for predicting the performance of rubber formulations, comprising:

[0006] A large student model is obtained by training a large sample set in advance, and at least one small teacher model is obtained by training a small sample set in advance; wherein the number of samples in the large sample set is greater than the number of samples in the small sample set; and each set of sample data in both the large sample set and the small sample set includes a formula sample, a process condition sample, and a performance sample.

[0007] The teacher's small model is used to perform reverse distillation training on the student's large model to obtain a rubber performance prediction model;

[0008] The rubber formulation to be predicted is input into the rubber performance prediction model to obtain the performance prediction results.

[0009] In one embodiment, a large student model is obtained by training using a large-scale sample set, including:

[0010] Construct a large neural network model, which includes a neural network prediction layer, a performance correlation monitor, a first GNN refinement layer, and a second GNN refinement layer;

[0011] The neural network prediction layer performs learning operations on the formula sample and the process condition sample, and outputs an initial prediction vector; the initial prediction vector includes the initial prediction performance and the predicted process conditions.

[0012] The performance correlation monitor identifies the strength of correlation between various performance indicators for each performance sample and obtains a performance correlation topology map.

[0013] Under the set decoupling constraints, the first GNN refinement layer corrects the initial prediction performance in the initial prediction vector according to the performance correlation topology graph; wherein the set decoupling constraints are that the process conditions retain self-loops and do not participate in the information transmission between performance indicators.

[0014] The second GNN refinement layer, based on the causal hierarchical prior rules and in combination with the formulation sample and the process condition sample, performs a secondary correction on the initial prediction performance in the initial prediction vector to obtain the corrected prediction vector; wherein, the causal hierarchical prior rules characterize the causal hierarchical topological relationship between the formulation, process conditions and performance indicators in rubber.

[0015] The predicted vector is subjected to a loss function operation, and the large neural network model is optimized and iterated based on the loss function value obtained from the operation until a large neural network model that satisfies the function threshold is obtained, which is then used as the student large model.

[0016] In one embodiment, the teacher-based small model is used to perform reverse distillation training on the student-based large model to obtain a rubber performance prediction model, including:

[0017] Large-scale sample data is selected from the large-scale sample set. The student model performs performance prediction based on the formula samples and process condition samples in the large-scale sample data. The prediction results and the performance samples in the large-scale sample data are subjected to MSE loss calculation to obtain the basic true value loss.

[0018] The small-scale sample set selects small-scale sample data, and performs zero-padding or truncation dimension alignment processing on the process condition samples in the small-scale sample data to make the dimension of the process condition samples the same as the set dimension, thus obtaining small-scale aligned sample data; the set dimension is the dimension required by the student large model for the input process condition parameters.

[0019] The student large model performs performance prediction based on the formula samples and process condition samples in the small-scale aligned sample data to obtain the large model prediction results.

[0020] The MSE loss is calculated by performing MSE loss on the prediction results of the large model and the performance samples in the small-scale sample data to obtain the true value loss.

[0021] The teacher mini-model performs performance prediction based on the formula samples and process condition samples in the small-scale sample data to obtain the mini-model prediction results.

[0022] The MSE loss is calculated by performing MSE loss on the prediction results of the large model and the prediction results of the small model to obtain the output spatial distillation loss.

[0023] The feature extraction backbone network of the student large model extracts features from the recipe samples in the small-scale aligned sample data to obtain the feature vector extracted by the large model.

[0024] The feature extraction backbone network of the teacher mini-model extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the mini-model; wherein, the feature vector extracted by the large model and the feature vector extracted by the mini-model have the same dimension;

[0025] MSE loss calculation is performed on the feature vectors extracted from the large model and the feature vectors extracted from the small model to obtain the distillation loss of the character space.

[0026] The total loss value is obtained by weighted summing of the basic truth loss, the truth loss, the output space distillation loss, and the characterization space distillation loss.

[0027] The network parameters of the student model are iteratively optimized based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

[0028] In one embodiment, MSE loss calculation is performed on the large model prediction result and the small model prediction result to obtain the output spatial distillation loss, including:

[0029] The performance indicators are divided into three levels, A, B and C, according to their importance, and a corresponding weighting coefficient is determined for each level.

[0030] Based on the levels corresponding to each prediction performance in the large model prediction results and the small model prediction results, and the level weight coefficients, the output space distillation loss formula is applied. The output space distillation loss is obtained through calculation; wherein, The output space distillation loss, These are the grade weighting coefficients corresponding to grade A, grade B, and grade C, respectively. These are the prediction results of the large model and the small model, respectively, for the first item in the A-level category. Predictive performance of each performance metric; These are the prediction results of the large model and the small model, respectively, for the first B level. Predictive performance of each performance metric; The prediction results of the large model and the small model are respectively the first prediction result of the C level. Predictive performance of each performance metric; All of these are performance prediction masks for the teacher's small model, and all have values ​​of 0 or 1; It is a very small positive number.

[0031] In one embodiment, after obtaining the small model prediction result, the method further includes:

[0032] The MSE loss is calculated by performing MSE loss on the prediction results of the small model and the performance samples in the small-scale sample data to obtain the true value loss of the small model.

[0033] Accordingly, the weighted summation of the basic truth loss, the output space distillation loss, and the characterization space distillation loss yields the total loss value, including:

[0034] The total loss value is obtained by weighted summing of the basic truth loss, the output space distillation loss, the representation space distillation loss, and the small model truth loss; wherein the weight coefficient corresponding to the small model truth loss is less than the weight coefficients corresponding to the basic truth loss, the output space distillation loss, and the representation space distillation loss, respectively.

[0035] In one embodiment, a small teacher model is obtained by training with a small sample set, including:

[0036] The small sample set is divided into two independent subsets: a first small sample set and a second small sample set.

[0037] The first small-scale sample subset is input into a lightweight GNN mini-model for neural network model training to obtain the first teacher mini-model.

[0038] The second small sample subset is input into a lightweight GNN mini-model for neural network model training to obtain the second teacher mini-model.

[0039] In one embodiment, the teacher-based small model is used to perform reverse distillation training on the student-based large model to obtain a rubber performance prediction model, including:

[0040] The student large model performs performance prediction on the large-scale sample data selected from the large-scale sample set, and obtains the basic true loss based on the prediction results.

[0041] The first small-scale sample data selected from the first small-scale sample subset is subjected to dimension alignment processing, such as zero-padding or truncation of process condition samples, to obtain the first small-scale aligned sample data.

[0042] The student large model performs performance prediction based on the formula samples and process condition samples in the first small-scale aligned sample data to obtain the prediction result of the first large model.

[0043] The first true value loss is obtained by performing MSE loss calculation on the prediction results of the first large model and the performance samples in the first small-scale aligned sample data.

[0044] The first teacher mini-model performs performance prediction based on the formula samples and process condition samples in the first small-scale sample data, and obtains the prediction result of the first mini-model.

[0045] Perform MSE loss calculation on the prediction results of the first large model and the prediction results of the first small model to obtain the first output space distillation loss;

[0046] The feature extraction backbone network of the student large model extracts features from the recipe samples in the first small-scale aligned sample data to obtain the feature vector extracted by the first large model.

[0047] The feature extraction backbone network of the first teacher mini-model extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the first mini-model; wherein, the feature vector extracted by the first large model and the feature vector extracted by the first mini-model have the same dimension;

[0048] MSE loss is calculated by performing MSE loss operation on the feature vectors extracted from the first large model and the feature vectors extracted from the first small model to obtain the first representation space distillation loss.

[0049] The second small-scale sample data selected from the second small-scale sample subset is subjected to dimension alignment processing, such as zero-padding or truncation of process condition samples, to obtain the second small-scale aligned sample data.

[0050] The student large model performs performance prediction based on the formula samples and process condition samples in the second small-scale aligned sample data to obtain the prediction results of the second large model.

[0051] The second true value loss is obtained by performing MSE loss calculation on the prediction results of the second large model and the performance samples in the second small-scale aligned sample data.

[0052] The second teacher mini-model performs performance prediction based on the formula samples and process condition samples in the second small-scale sample data to obtain the prediction results of the second mini-model.

[0053] The MSE loss is calculated by performing MSE loss on the prediction results of the second large model and the prediction results of the second small model to obtain the second output space distillation loss.

[0054] The feature extraction backbone network of the student large model extracts features from the recipe samples in the second small-scale aligned sample data to obtain the feature vector extracted by the second large model.

[0055] The feature extraction backbone network of the second teacher mini-model extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the second mini-model; wherein, the feature vector extracted by the second large model and the feature vector extracted by the second mini-model have the same dimension;

[0056] The MSE loss is calculated by performing MSE loss operation on the feature vectors extracted from the second large model and the feature vectors extracted from the second small model to obtain the distillation loss of the second representation space.

[0057] According to the total loss formula Perform a weighted summation to obtain the total loss value; where, These are the total loss values, This is the basic truth loss; These are the first truth loss and the second truth loss, respectively; These are the distillation losses in the first output space and the distillation losses in the second output space, respectively. These are the first characterization space distillation loss and the second characterization space distillation loss, respectively. All are loss weighting coefficients;

[0058] The network parameters of the student model are iteratively optimized based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

[0059] A rubber formulation performance prediction device, comprising:

[0060] The model training module is used to pre-train a large student model using a large sample set and at least one teacher model using a small sample set; wherein the number of samples in the large sample set is greater than the number of samples in the small sample set; and each set of sample data in both the large sample set and the small sample set includes a formula sample, a process condition sample, and a performance sample.

[0061] The distillation training module is used to perform reverse distillation training on the student large model using the teacher small model to obtain a rubber performance prediction model.

[0062] The performance prediction module is used to input the rubber formulation to be predicted into the rubber performance prediction model to obtain the performance prediction result.

[0063] A rubber formulation performance prediction device, comprising:

[0064] Memory, used to store computer programs;

[0065] A processor for executing the computer program to implement the steps of the rubber formulation performance prediction method as described in any of the preceding claims.

[0066] A computer-readable storage medium storing a computer program that is executed to implement the steps of the rubber formulation performance prediction method as described in any of the preceding claims.

[0067] The present invention provides a method, apparatus, device, and computer-readable storage medium for predicting the performance of rubber formulations. The method includes: pre-training a large student model using a large sample set and at least one small teacher model using a small sample set; wherein the number of samples in the large sample set is greater than the number of samples in the small sample set; and each set of sample data in both the large and small sample sets includes formulation samples, process condition samples, and performance samples; using the small teacher model to perform reverse distillation training on the large student model to obtain a rubber performance prediction model; and inputting the rubber formulation to be predicted into the rubber performance prediction model to obtain the performance prediction result.

[0068] This application fully utilizes the simultaneous existence of large-scale incomplete process condition data and small-scale accurate process condition data in industrial scenarios. Specifically, it trains large-scale and small-scale sample sets to obtain student large models and teacher small models respectively. Furthermore, through knowledge distillation, the student large model is further optimized using the teacher small model. The refined knowledge learned by the teacher small model based on the small-scale sample set is transferred to the student large model, resulting in a final rubber performance prediction model with both broad knowledge coverage and accurate prediction performance. It can uniformly handle formulation performance prediction tasks under various process conditions without requiring separate modeling and maintenance for different processes. Therefore, the rubber performance prediction model in this application, based on pre-trained data, can quickly evaluate the expected performance of rubber formulations without relying on extensive physical experiments, shortening the R&D cycle of rubber products and reducing experimental material and labor costs. Attached Figure Description

[0069] To more clearly illustrate the technical solutions of 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 A schematic flowchart illustrating the rubber formulation performance prediction method provided in this application embodiment;

[0071] Figure 2 This is a structural block diagram of the rubber formulation performance prediction device provided in the embodiments of this application. Detailed Implementation

[0072] The core of this invention is to provide a method, apparatus, device, and computer-readable storage medium for predicting the performance of rubber formulations. It aims to utilize knowledge distillation technology to transfer knowledge about specific processes learned from one or more precise small datasets to a large student model trained on massive, mixed data. This unlocks the value of historical big data, ultimately resulting in a model capable of accurately predicting the performance of different processes across a broad formulation space. This effectively improves the prediction accuracy of the model and reduces reliance on high-quality labeled data. Based on this prediction model, the expected performance of rubber formulations can be quickly evaluated, shortening the R&D cycle of rubber products and reducing experimental material and labor costs.

[0073] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely 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.

[0074] like Figure 1 As shown, Figure 1 This is a flowchart illustrating the rubber formulation performance prediction method provided in this application embodiment.

[0075] In one specific embodiment of this application, the method for predicting the performance of rubber formulations includes:

[0076] S1: Train a large student model using a large sample set in advance, and train at least one teacher model using a small sample set.

[0077] In this embodiment, the number of samples in the large-scale sample set is greater than the number of samples in the small-scale sample set.

[0078] Understandably, in the vast amount of historical R&D and production data related to rubber formulation development in the rubber industry, most of the data often has severely missing or highly mixed records of process parameters, testing conditions, etc., making it difficult to use directly for supervised learning. There are also a small number of small-scale orthogonal experimental data records completed under specific and strictly controlled processes that are accurate, but they suffer from insufficient sample size and cannot cover the broad formulation space.

[0079] The large-scale sample set in this embodiment is obtained based on a large amount of historical R&D data with missing or mixed process parameters and test conditions. The small-scale sample set, on the other hand, is obtained based on a small amount of data recorded from experiments conducted under specific, strictly controlled processes. The large-scale sample set corresponds to scenarios where process conditions are not explicitly modeled or are only represented by coarse labels, while the small-scale sample set corresponds to scenarios where process conditions are controlled and fully recorded.

[0080] Therefore, the number of sample data contained in the large-scale sample set in this embodiment is much greater than the number of sample data in the small-scale sample set. For example, the large-scale sample set may contain tens of thousands or even hundreds of thousands of data points from historical R&D records or production lines, while the small-scale sample set may only contain a few hundred data points measured under strictly controlled laboratory conditions.

[0081] Furthermore, large-scale sample sets involve more complex and diverse process conditions, and there may be instances where process conditions are missing from some sample data. In short, the process conditions covered in large-scale samples are more complex. On the other hand, the process condition parameters involved in small-scale sample sets are more specific and singular, meaning they are less complex.

[0082] Based on this, each sample data set in both the large-scale and small-scale sample sets includes formulation samples, process condition samples, and performance samples. Formulation samples record the types and amounts of various raw materials, such as 100 parts natural rubber, 50 parts carbon black, and 1 part accelerator. Process condition samples actually include samples of process parameters for preparing the rubber and samples of test conditions for testing rubber properties. For example, a vulcanization temperature of 150℃ and a vulcanization time of 30 minutes are process parameters, while a test environment temperature of 23℃ is a test condition sample. In this application, process parameters and test conditions are considered together as the same type of sample to comprehensively consider their impact on rubber properties. Performance samples record the test results of various physicochemical properties of the corresponding formulation, such as Mooney viscosity of 65 and tensile strength of 28 MPa.

[0083] Before starting model training, you can configure the files uniformly, as follows:

[0084] 1) Configuration of the complete set of rubber raw materials; enumerate all possible raw material categories in the rubber formulation R&D scenario, covering all categories of materials such as matrix rubber, various functional resins, vulcanization system additives, fillers and reinforcing agents, processing aids, protective additives, coupling modifiers, special synthetic rubbers, and homogenization modifiers; assign a unique numerical index to each raw material in the set as the basic number of the graph neural network node; synchronously configure the raw material functional category mapping relationship, divide all raw materials into 23 preset functional groups, and establish a mapping relationship table between raw material name and functional category number to generate functional category embedding features of raw material nodes; when reading any formulation data in the future, the raw material name in the table only needs to match the complete set to complete the standardized encoding, eliminating the input dimension disorder caused by the difference in raw material naming in different datasets.

[0085] 2) Complete set of process conditions and performance indicators; a complete set of rubber test performance indicators is compiled to cover all conventional performance indicators such as vulcanization performance, Mooney viscosity, Mooney scorch, tensile mechanics, hot air aging, hardness, rebound, flexural fatigue, compression heat generation, DIN (German Industrial Standard) abrasion, Akron abrasion, etc.; within this set, two types of fields are divided: one is the set of process conditions, which are only used as auxiliary features for model input and do not participate in performance prediction; the other is the set of performance indicators to be predicted, which are the real mechanical performance indicators remaining after removing all process conditions, and are the core output objects of the student's large model multi-task regression.

[0086] In addition, a fixed number of extended dimensions are reserved at the end of the process condition feature vector as process coding slots, defining a process number field to display the process category to which the rubber process belongs. If the sample data in the dataset carries a unique process code, the process number can be uniquely encoded and filled into the reserved slots. This configuration does not require modification to the overall code and feature dimension structure, and is compatible with the addition of niche process types in the future, improving the process expansion adaptability of the entire model.

[0087] 3) Network structure parameter configuration: Configure the network structure parameters of the student large model and the teacher small models of each process: including the number of GNN (Graph Neural Network) convolutional layers, node embedding dimension, global pooling method, shared multilayer perceptron size, and operation parameters of label graph refinement layer, recipe, process conditions and performance ternary graph refinement layer; record the standard process feature dimension of the large model and the native process feature dimension of each small teacher.

[0088] Based on this, in one embodiment, the training process for the large student model may include:

[0089] S11: Construct a large neural network model, which includes a neural network prediction layer, a performance correlation monitor, a first GNN refinement layer, and a second GNN refinement layer.

[0090] S12: The neural network prediction layer performs learning operations on the formula samples and process condition samples, and outputs an initial prediction vector; the initial prediction vector includes the initial prediction performance and the predicted process conditions.

[0091] S13: The performance correlation monitor identifies the strength of correlation between various performance indicators for each performance sample and obtains a performance correlation topology map.

[0092] S14: Under the set decoupling constraints, the first GNN refinement layer corrects the initial prediction performance in the initial prediction vector according to the performance correlation topology graph; the set decoupling constraints are that the process conditions retain self-loops and do not participate in the information transmission between performance indicators.

[0093] S15: The second GNN refinement layer, based on the causal hierarchical prior rules and combined with the formulation samples and process condition samples, performs a secondary correction on the initial prediction performance in the initial prediction vector to obtain the corrected prediction vector; whereby the causal hierarchical prior rules characterize the causal hierarchical topological relationship between the formulation, process conditions and performance indicators in rubber.

[0094] S16: Perform loss function calculation on the predicted vector, and optimize and iterate the large neural network model based on the loss function value obtained from the calculation until a large neural network model that satisfies the function threshold is obtained, which serves as the student model.

[0095] In this embodiment, the large neural network model, which is the neural network architecture that is finally trained to form the student large model, can be composed of a GNN backbone network layer, a global mean pooling layer, a shared multilayer perceptron, and a multi-task prediction head.

[0096] It should be noted that before the large neural network model learns and trains on the input sample data, it first actively performs upward operations on the sample data, converting it into an embedded vector form that can be directly computed by the various structural layers of the large neural network model. The formula samples are constructed as graph data of formula raw material vectors, with the vector dimensions set in the configuration file process described above. Each raw material is represented by a raw material node, and the features of each raw material node include the raw material name code, functional category code (one-hot encoding for 23 functional categories), and feed quantity (e.g., number of parts), concatenated together. The edges between raw material nodes are constructed as fully connected layers, representing the blending relationships of all raw materials in the formula sample. The process condition samples are processed into a conditional feature vector (cond_feat); discrete process parameters (e.g., vulcanization temperature level) are one-hot encoded; continuous process parameters (e.g., vulcanization time) are normalized.

[0097] In addition, for performance samples, in order to avoid large differences in the performance data corresponding to different types of performance indicators, all performance samples can be normalized first. For example, normalization can be performed using the mean or standard deviation of all performance samples, or other normalization methods can be used, as long as the impact of performance data differences between different performance indicators on subsequent model training can be avoided.

[0098] The GNN backbone network can be constructed by stacking multiple GCNConv layers (graph convolutional layers). The GNN backbone network performs multi-layer graph convolution operations (such as GCNConv) on the ingredient embedding vectors generated from the recipe samples, performing message passing on each ingredient node and aggregating information from its neighboring nodes to update the node's feature representation. Subsequently, global mean pooling is used to aggregate the features of all nodes into a fixed-length graph embedding vector, which serves as a comprehensive representation of the entire recipe.

[0099] The physics-assisted branch can be embedded in the large neural network model or used as a computational module independent of the neural network model, which can be called when needed. This physics-assisted branch is used to calculate the interpretable physical feature vector. According to the 23 functional categories defined above, each raw material in the formulation sample is grouped according to its functional category. The amount of all raw materials in the same functional category is summed to obtain a 23-dimensional interpretable physical feature vector. This vector represents the total content of each functional component in the formulation. As a priori physical feature, it is used to assist the training of the main network and enhance the physical consistency of the model.

[0100] The Shared Multilayer Perceptron (SMP) concatenates the graph embedding vector, the process condition vector (cond_feat), and the interpretable physical feature vector, and then inputs them into a shared multilayer fully connected network for deep fusion. The output of the Shared MLP connects to two independent branches: a Performance Head, which outputs the initial predicted performance vector; and a Process Head, which outputs a predicted process condition vector. These two vectors together constitute the initial predicted vector. This decoupled dual-output design treats performance prediction and process condition prediction as two related but distinct tasks, achieving implicit decoupling at the feature level and creating conditions for subsequent refinement and decoupling constraints.

[0101] Building upon this foundation, the performance correlation monitor dynamically or statically calculates the correlations between different performance metrics during training. For example, it calculates the Pearson correlation coefficient between the "tensile strength" and "elongation at break" performance metrics for all samples in the training data. When the absolute value of the correlation coefficient between two performance metrics is higher than a preset threshold (e.g., 0.6), they are considered strongly correlated, and an edge is established between them in the topology graph; otherwise, the edge is broken, ultimately generating an N×N adjacency matrix, where N is the total number of performance metrics. This matrix is ​​the performance correlation topology graph. This topology graph describes the mutual influence relationships between performance metrics, providing a priori structure that conforms to the statistical laws of the data for subsequent performance prediction correction based on graph networks. Dynamic calculation of performance metrics by the performance correlation monitor means updating the correlation of the learned performance metrics once after each batch of sample data is read during model iteration. Static calculation of performance metric correlation means determining the correlation between performance metrics all at once based on all sample data. It is worth noting that the performance correlation monitor only performs the correlation identification operation between performance metrics during model training. After the model training is completed, the final performance correlation topology is used as a fixed performance correlation topology and applied to the actual performance prediction.

[0102] The first GNN refinement layer, also known as the decoupled constraint performance graph refiner (LabelGraphRefiner), is essentially a GNN neural network layer. Under decoupling constraints, the first GNN refinement layer uses a performance correlation topology graph as the graph structure to correct the initial prediction vector, primarily revising the initial predicted performance. The core of this correction process lies in setting decoupling constraints: during graph message passing, process condition nodes (i.e., the data in each dimension of the predicted process condition vector) are forced to retain only self-loops and not participate in information transmission between performance indicator nodes. This ensures that the process condition feature vector is updated only based on itself, without being affected by the performance prediction value. Simultaneously, the predicted value of a performance indicator will transmit information to its related other performance indicators according to the performance correlation topology graph. This decoupling design effectively avoids the mutual confusion between process condition information and performance information during the refinement process, making the performance correction purer and more accurate.

[0103] The second GNN refinement layer, namely the domain prior layer graph refiner (FPPGraphRefiner), is essentially also a GNN neural network layer. In essence, it only corrects the initial prediction performance in the initial prediction vector after the first GNN refinement layer.

[0104] The second GNN refinement layer operates based on a predefined causal hierarchy prior rule. This rule states that the formulation determines the process conditions, and the process conditions determine the performance indicators, representing a fundamental causal relationship in rubber science: the formulation determines the material's potential properties, while the formulation and process conditions together determine the final product's performance. This second GNN refinement layer constructs a ternary graph (FPP Graph) containing three information elements: formulation, process conditions, and performance, based on the causal hierarchy prior rule, and performs message passing on this graph. The second GNN refinement layer also dynamically adjusts the correction magnitude using a gating vector. The gating vector works by comparing the initial predicted performance value with the corrected predicted performance value after one correction by the first GNN refinement layer. If the difference is small, the initial prediction is considered highly reliable, and a larger weight is given to the initial prediction; if the difference is large, the initial prediction may be inaccurate, and a larger weight is given to the corrected value. Through this dynamic weighting mechanism, fine-tuning of the prediction results is achieved, resulting in the final corrected prediction vector.

[0105] The predicted performance value in the corrected prediction vector and the true performance in the sample data can be compared using the Mean Square Error (MSE) loss function. Simultaneously, to enhance the model's physical consistency, a Physical Regularizer loss can be added, ensuring, for example, that the predicted post-aging strength is always less than or equal to the pre-aging strength—a fundamental physical fact. Using backpropagation, the gradient of the loss function value (MSE loss function value + Physical Regularizer loss) is propagated back to each network layer of the entire student model, and an optimizer (such as AdamW) is used to update the network parameters. This optimization and training process is repeated until the loss function value no longer decreases on the validation set or reaches the preset number of training epochs. The resulting neural network model is then considered the successfully trained student model.

[0106] Based on the above discussion, the Teacher Small Model can be trained using a similar graph neural network architecture to the Student Large Model, but with fewer network layers and lower hidden dimensions. For example, it contains only 2-3 GCNConv layers and 1-2 fully connected layers, resulting in a much smaller number of parameters compared to the Student Large Model. This lightweight design effectively prevents overfitting on small datasets.

[0107] Furthermore, in practical applications, a small sample set can be divided into multiple independent small sample subsets according to actual needs. Each small sample subset is independently trained to obtain a teacher small model, and the teacher small models work together to distill the student large model.

[0108] In one embodiment, the training process of the teacher mini-model may include:

[0109] The small sample set is divided into two independent subsets: the first small sample set and the second small sample set.

[0110] The first small sample subset is input into a lightweight GNN mini-model for neural network model training to obtain the first teacher mini-model.

[0111] The second small sample subset is input into the lightweight GNN mini-model for neural network model training to obtain the second teacher mini-model.

[0112] In this embodiment, the small-scale sample set is divided into multiple small-scale sample subsets according to the process conditions, and each subset corresponds to a process condition. Multiple teacher small models are trained on each small-scale sample subset, and the student large model is trained by distillation using each teacher small model to obtain multiple rubber performance prediction models corresponding to different process conditions.

[0113] To improve the robustness and diversity of the teacher models, the small sample set is divided into two independent small sample subsets. The sample data from each of these subsets is then input into two structurally identical lightweight GNN mini-models for neural network training, resulting in two teacher mini-models (a first teacher mini-model and a second teacher mini-model). Each lightweight GNN mini-model has a small number of network layers and parameters (e.g., only two graph convolutional layers and two fully connected layers) to prevent overfitting on limited sample data. This "dual-teacher" setup integrates two slightly different perspectives on small sample data, making the subsequent knowledge distillation process more stable and effective.

[0114] In this embodiment, the teacher small model is a lightweight graph neural network model, and the student large model is a graph neural network model with a larger parameter scale. During the distillation training process, only the network parameters of the student large model are updated to achieve reverse knowledge distillation.

[0115] S2: Use the teacher's small model to perform reverse distillation training on the student's large model to obtain a rubber performance prediction model.

[0116] As mentioned above, the teacher small model, trained on a high-quality, low-noise, accurate sample dataset, represents the precise mapping relationship under a specific process, but its generalization ability is limited by the size of the small dataset. The student large model, trained on a large-scale dataset, has a broader perspective and a more comprehensive understanding of the recipe space, but its prediction accuracy under a specific process is poor due to the mixed or missing process conditions in its training data.

[0117] Therefore, the training process in fusion distillation is not simply a matter of copying the parameters of the teacher model to the student model, but rather a process of knowledge transfer. By designing specific loss functions, for example, forcing the student model's large model to numerically approximate the teacher model's small model's predicted output for the same input samples from a small dataset, the student model, while maintaining its understanding of the broad formulation distribution learned from large-scale data, can absorb the implicit knowledge learned by the teacher model under specific processes (such as implicit correlations between performance indicators). Ultimately, it becomes a rubber performance prediction model with high-precision predictive ability for specific processes across a broad formulation space.

[0118] In one embodiment, the process of using a teacher-small model to perform reverse distillation training on a student-large model may include:

[0119] S200: Select large-scale sample data from a large-scale sample set. The student model performs performance prediction based on the formula samples and process condition samples in the large-scale sample data. The prediction results and the performance samples in the large-scale sample data are used to perform MSE loss calculation to obtain the basic true value loss.

[0120] S201: Select small-scale sample data from the small-scale sample set, and perform zero-padding or truncation dimension alignment on the process condition samples in the small-scale sample data to make the dimension of the process condition samples the same as the set dimension, thus obtaining small-scale aligned sample data; the set dimension is the dimension required by the student large model for the input process condition parameters.

[0121] S202: The student large model performs performance predictions based on the formula samples and process condition samples in the small-scale aligned sample data, and obtains the prediction results of the large model.

[0122] S203: Perform MSE loss calculation on the prediction results of the large model and the performance samples in the small-scale sample data to obtain the true loss;

[0123] S204: The teacher's small model performs performance predictions based on formula samples and process condition samples in a small-scale sample data, and obtains the small model prediction results;

[0124] S205: Perform MSE loss calculation on the prediction results of the large model and the small model to obtain the output spatial distillation loss;

[0125] S206: The feature extraction backbone network of the student large model extracts features from the recipe samples in the small-scale aligned sample data to obtain the feature vector extracted by the large model.

[0126] S207: The feature extraction backbone network of the teacher's small model extracts features from formula samples in a small-scale sample data to obtain the feature vector extracted by the small model; the feature vector extracted by the large model and the feature vector extracted by the small model have the same dimension.

[0127] S208: Perform MSE loss calculation on the feature vectors extracted from the large model and the feature vectors extracted from the small model to obtain the distillation loss of the representation space;

[0128] S209: The total loss value is obtained by weighted summing of the basic truth loss, the truth loss, the output space distillation loss, and the characterization space distillation loss.

[0129] S210: Based on the total loss value, perform network parameter optimization iterations on the student model until the total loss value meets the loss threshold. Then, use the optimized student model as the rubber performance prediction model.

[0130] First, a subset of sample data is selected from the large-scale sample set as the large-scale sample data. The formulation samples and process condition samples from this subset are input into the student large model to obtain the prediction results. The mean squared error (MSE) loss is then calculated between the prediction results and the true performance samples from the large-scale sample data to obtain the basic ground value loss Lo. This loss ensures that the student large model does not forget the general formulation and performance mapping knowledge learned on large-scale, mixed data during the distillation process, and serves as the anchor for the entire training process.

[0131] Secondly, a portion of the sample data from the small-scale sample set is selected as the small-scale sample data. Since the input dimensions of the student large model and the teacher small model may differ, the process condition samples in the small-scale sample data need to be dimension-aligned. Specifically: if the dimension of the process condition vector required by the student large model is greater than the dimension corresponding to the teacher small model, zeros are padded to the end of the process condition vector in the teacher model; conversely, if the dimension of the process condition vector required by the student large model is less than the dimension corresponding to the teacher small model, the dimension of the process condition vector is truncated to the dimension of the student model, that is, some process condition parameters that do not exist in the large-scale sample set are removed. Through zero-padding or truncation, small-scale aligned sample data with the same set dimension is obtained, ensuring that it can be correctly received by the student large model.

[0132] Therefore, the aligned small-scale sample data is input into the student's large model to obtain the large model's prediction results, and its MSE with the true performance samples is calculated to obtain the true loss L. B This loss measures the direct performance of the current student model on clean, small data and serves as a benchmark for the distillation effect.

[0133] Furthermore, a small-scale sample dataset (unaligned sample data, maintaining its original dimensions) is input into the teacher's small model to obtain its prediction results, which are denoted as the small model prediction results. Then, MSE loss is calculated on the large model prediction results and the small model prediction results to obtain the output space distillation loss. This loss is the core of distillation; it forces the final output of the student's large model to numerically approximate the output of the teacher's small model, thereby directly learning the teacher's accurate prediction patterns under specific processes.

[0134] In another embodiment, the process of calculating the output space distillation loss may further include:

[0135] The performance indicators are divided into three levels, A, B and C, according to their importance, and a corresponding weighting coefficient is determined for each level.

[0136] Based on the prediction performance levels and weight coefficients of the large and small models, and according to the output space distillation loss formula... The output space distillation loss is obtained through computation; where, To account for the distillation loss in the output space, These are the grade weighting coefficients corresponding to grades A, B, and C, respectively. These are the predictions from the large model and the small model, respectively, for the first item in level A. Predictive performance of each performance metric; These are the predictions for level B in the large model prediction results and the small model prediction results, respectively. Predictive performance of each performance metric; These are the predictions from the large model and the small model, respectively, for the C-level. Predictive performance of each performance metric; All are performance prediction masks for teacher small models; It is a very small positive number.

[0137] In this embodiment, based on the level of importance of performance indicators in industrial applications, all performance indicators to be predicted are divided into three levels: A, B, and C, with a corresponding weighting coefficient assigned to each level. Level A is the highest and most important, while level C is the lowest and least important. Therefore, level A has the largest weighting coefficient, and level C has the smallest. In a specific example, key indicators characterizing vulcanization properties, such as ML (dN.m), Mooney viscosity ML(1+4), resilience (%), hardness after aging (Shore A), and bottom temperature rise (°C), can be classified as level A and assigned the highest weight. The elongation at break (%), strength after aging (MPa), and elongation at break after aging (%) are classified into grade B and assigned a medium weight. Indicators that may be significantly affected by other factors or are not critical, such as strength (MPa), elongation at break (%), and 30,000 flexural cycles (grade), are classified as Grade C and assigned lower weights. For other target metrics that are not explicitly categorized, a weak distillation weight, such as 0.3, is assigned. This tiered strategy can be flexibly configured to adapt to the varying performance prediction accuracy requirements of different application scenarios.

[0138] Based on the prediction performance levels and level weight coefficients corresponding to each prediction performance in the large-model and small-model prediction results, the output space distillation loss is calculated using the following formula:

[0139] .

[0140] in, This represents the distillation loss in the output space.

[0141] These are the predictions from the large model and the small model, respectively, for the first item in level A. Predictive performance of each performance metric.

[0142] These are the predictions for level B in the large model prediction results and the small model prediction results, respectively. Predictive performance of each performance metric.

[0143] These are the predictions from the large model and the small model, respectively, for the C-level. Predictive performance of each performance metric.

[0144] All of these are performance prediction masks for the teacher's small model. If the value is 1, it means that the teacher's small model has a predicted performance value for this performance indicator. If the value is 0, it means that the teacher's small model does not output the predicted performance value corresponding to this performance indicator and does not participate in the loss calculation. It is a very small positive number, and can take the value 1e-8, to prevent the denominator from being zero.

[0145] By determining the output space distillation loss through weighted summation, the student large model will focus on performance indicators that are more important for industrial applications during the distillation process, thereby achieving a better balance between model capacity and prediction accuracy.

[0146] Based on the above discussion, this embodiment further utilizes the feature extraction backbone network of the student large model to extract features from the recipe samples in the small-scale aligned sample data, obtaining a fixed-dimensional graph-level embedding vector, denoted as the large model extracted feature vector. Simultaneously, the feature extraction backbone network of the teacher small model is used to extract features from the recipe samples in the small-scale sample data, obtaining a small model extracted feature vector. The dimensions of these two feature vectors should be the same (this can be achieved by reasonably setting the dimensions of the output features of the corresponding feature extraction backbone networks in the student large model and teacher small model, respectively). MSE loss calculation is performed on the large model extracted feature vector and the small model extracted feature vector to obtain the representation space distillation loss L. kdrepr This loss is designed to enable the student large model to learn not only the final output of the teacher small model, but also the intermediate feature representations of the teacher small model. This loss calculation method usually leads to more stable and more fundamental knowledge transfer effects.

[0147] The base true value loss, the true value loss, the output space distillation loss, and the representation space distillation loss are weighted and summed to obtain the total loss value. Based on the total loss value, the network parameters of the student large model are iteratively optimized using an optimizer (such as Adam). During the iteration process, the parameters of the teacher small model remain fixed. The process from S201 to S210 is repeated until the total loss value meets the preset loss threshold or the number of training epochs. At this point, the optimized student large model can be used as a rubber performance prediction model.

[0148] In another embodiment, a step of collaborative fine-tuning the teacher mini-model can be added to improve the distillation effect. Prior to S209 above, the following may also be included:

[0149] The MSE loss is calculated by combining the prediction results of the small model obtained from the above teacher small model on small-scale sample data with the measured performance samples in the small-scale sample data, and then obtaining the true value loss L of the small model. small This loss is used to measure the performance of the teacher model itself on a precise small dataset under the current training state;

[0150] Accordingly, the process described in S209 above may include:

[0151] The total loss value is obtained by weighting and summing the base truth loss, output space distillation loss, characterization space distillation loss, and small model truth loss. The weight coefficient corresponding to the small model truth loss is smaller than the weight coefficients corresponding to the base truth loss, output space distillation loss, and characterization space distillation loss, respectively.

[0152] For example, the weight coefficient corresponding to the ground truth loss of the small model is set to 0.1 to prevent the teacher's small model from undergoing drastic changes and forgetting its useful knowledge under the guidance of the student's large model. Then, based on the total loss value, the optimizer is used to iteratively optimize the network parameters of both the student's large model and the teacher's small model. This means that during the distillation process, the teacher's small model is not completely fixed, but can be fine-tuned at a very slow learning rate to better adapt to the learning of the student's large model, achieving teacher-student collaborative optimization, and ultimately further improving the performance of the student's large model after distillation.

[0153] Based on the above discussion, in practical applications, there may be multiple teacher mini-models. Taking an example where the teacher mini-model includes a first teacher mini-model and a second teacher mini-model, and the first and second teacher mini-models are trained on the first and second small-scale sample subsets, respectively, then the reverse distillation training process for the student large model can also include:

[0154] S2a0: The student large model performs performance prediction on large-scale sample data selected from a large-scale sample set, and obtains the basic true loss based on the prediction results.

[0155] S2b1: The first small-scale sample data selected from the first small-scale sample subset is subjected to dimension alignment processing, such as zero-padding or truncation of process condition samples, to obtain the first small-scale aligned sample data.

[0156] S2b2: The student large model performs performance predictions based on the formula samples and process condition samples in the first small-scale aligned sample data, and obtains the prediction results of the first large model.

[0157] S2b3: Perform MSE loss calculation on the prediction results of the first large model and the performance samples in the first small-scale aligned sample data to obtain the first true value loss;

[0158] S2b4: The first teacher small model performs performance prediction based on the formula samples and process condition samples in the first small-scale sample data, and obtains the prediction results of the first small model;

[0159] S2b5: Perform MSE loss calculation on the prediction results of the first large model and the prediction results of the first small model to obtain the first output space distillation loss;

[0160] S2b6: The feature extraction backbone network of the student large model extracts features from the recipe samples in the first small-scale aligned sample data to obtain the feature vector extracted by the first large model.

[0161] S2b7: The feature extraction backbone network of the first teacher small model, which extracts features from the formula samples in the first small-scale sample data to obtain the feature vector extracted by the first small model; wherein, the feature vector extracted by the first large model and the feature vector extracted by the first small model have the same dimension;

[0162] S2b8: Perform MSE loss calculation on the feature vectors extracted from the first large model and the feature vectors extracted from the first small model to obtain the first representation space distillation loss;

[0163] S2c1: The second smallest sample data selected from the second smallest sample subset is subjected to dimension alignment processing, such as zero-padding or truncation of process condition samples, to obtain the second smallest aligned sample data.

[0164] S2c2: The student large model performs performance predictions based on the formula samples and process condition samples in the second small-scale aligned sample data, and obtains the prediction results of the second large model.

[0165] S2c3: Perform MSE loss calculation on the performance samples in the second-largest model prediction results and the second-smallest scale aligned sample data to obtain the second true value loss;

[0166] S2c4: The second teacher small model performs performance prediction based on the formula samples and process condition samples in the second small-scale sample data, and obtains the prediction results of the second small model.

[0167] S2c5: Perform MSE loss calculation on the prediction results of the second largest model and the prediction results of the second smallest model to obtain the second output space distillation loss;

[0168] S2c6: The feature extraction backbone network of the student large model extracts features from the recipe samples in the second small-scale aligned sample data to obtain the feature vector extracted by the second large model.

[0169] S2c7: The feature extraction backbone network of the second teacher small model, which extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the second small model; wherein, the feature vector extracted by the second large model and the feature vector extracted by the second small model have the same dimension;

[0170] S2c8: Perform MSE loss calculation on the feature vectors extracted from the second largest model and the feature vectors extracted from the second smallest model to obtain the distillation loss of the second representation space;

[0171] S2d0: According to the total loss formula Perform a weighted summation to obtain the total loss value; where, These are the total loss values, Based on truth value loss; These are the first truth loss and the second truth loss, respectively. These are the distillation losses in the first output space and the distillation losses in the second output space, respectively. These are the first characterization space distillation loss and the second characterization space distillation loss, respectively. All are loss weighting coefficients;

[0172] S2e0: The network parameters of the student model are iteratively optimized based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

[0173] In this embodiment, processes S2b1~S2b8 and S2c1~S2c8 are two completely identical operation flows and can be completed in parallel in practical applications. Furthermore... These are all loss weighting coefficients, used to balance the importance of various losses. For example, you can set... The total loss formula symmetrically aggregates the knowledge distillation losses from all teachers, ensuring the model learns evenly from multiple processes. Finally, the network parameters of the student-scaled model are iteratively optimized based on the total loss value until the total loss value meets the loss threshold. The resulting student-scaled model integrates teacher knowledge from multiple precise processes, enabling accurate performance predictions across a wide recipe space based on different user-specified process conditions, greatly expanding the applicability of a single model.

[0174] Unlike traditional knowledge distillation methods that use a large model as the teacher and a small model as the student, this embodiment employs a reverse distillation mechanism where a lightweight teacher small model guides the student large model. The teacher small model learns cleaner formulations and the corresponding structures for performance on small-scale data with controlled process conditions, while the student large model learns extensive spatial statistical regularities of formulations on large-scale data. By outputting spatial distillation and characterizing spatial distillation, the structural knowledge of the small model is transferred to the large model, thereby enabling formulation-based performance prediction on large-scale data within a specific process reference frame.

[0175] This embodiment uses the large-scale dataset of the student model, with recipe as input and performance as output, to obtain the potential statistical regularity of recipe-performance without considering the process. Similarly, the small-scale dataset of the teacher model uses recipe as input and performance as output. However, since the process of the dataset is fixed, although the system of recipe as input and performance as output is still being learned, it actually implicitly contains process content (i.e., the small-scale dataset was originally obtained from a unified process, so the predictions achieved by the learned model are naturally predicted under this process, only the process is not reflected). The implicit process is the process corresponding to the dataset. In this single, fixed process, each teacher's small model corresponds to one process. Knowledge distillation is then performed on the student's large model using these teacher's small models. The resulting distilled student model learns the recipe-performance prediction knowledge for the specific process represented by the teacher model. The student model itself learns from a large-scale dataset with a vast recipe scale and coverage. Now, through distillation, it acquires specific process knowledge, which, combined with its previously learned statistical regularities regarding recipe-performance without considering the process, allows the distilled student model to predict recipe-performance within a broader range under the specified single process. In this process, the process is not used as input or other prerequisites; it is implicitly embedded in the learning of the teacher's small model. This achieves non-explicit process modeling—no process information is used for training, but the training result does indeed include process factors.

[0176] Based on the discussion, this application is not limited to having only one teacher model, but can have multiple teacher models, each trained on datasets B1, B2, B3...Bn, etc., thus obtaining teacher models corresponding to multiple processes. These multiple teacher models then teach the same student model, generating multiple versions of distillation student model files (it should be noted that the student model is code, while the student model file is a pre-trained model that can be directly called for prediction; integrating these together forms a selectable prediction student model with multiple model files), each corresponding to different processes (i.e., different teachers). When using prediction, different distilled student model files can be selected for prediction, thus achieving different, selectable process predictions. This allows users to choose their desired process scenario and then perform multi-process predictions based on the formulation and corresponding performance.

[0177] S3: Input the rubber formulation to be predicted into the rubber performance prediction model to obtain the performance prediction results.

[0178] Once the distillation training is completed and validated, a rubber performance prediction model tailored to a specific process can be obtained. In practical use, simply input the rubber formulation to be predicted (including the types and amounts of each raw material) into the rubber performance prediction model, and the model can quickly and accurately output the predicted performance results under that specific process condition, such as predicting its Mooney viscosity, minimum torque (ML), maximum torque (MH), tensile strength, elongation at break, and performance retention rate after aging, among other key performance indicators. This provides powerful decision support for formulation engineers, significantly reducing the number of experiments and shortening the R&D cycle.

[0179] In summary, this application fully utilizes the simultaneous existence of large-scale incomplete process condition data and small-scale accurate process condition data in industrial scenarios. Specifically, it trains student large models and teacher small models using data corresponding to large and small sample sets respectively. Furthermore, through knowledge distillation, the student large model is further optimized using the teacher small model. This transfers the refined knowledge learned by the teacher small model based on the small sample set to the student large model, resulting in a final rubber performance prediction model that possesses both broad knowledge coverage and accurate prediction performance. It can also uniformly handle formulation performance prediction tasks under various process conditions without requiring separate modeling and maintenance for different processes. Therefore, the rubber performance prediction model based on pre-trained data in this application can quickly evaluate the expected performance of rubber formulations without relying on extensive physical experiments, shortening the R&D cycle of rubber products and reducing experimental material and labor costs.

[0180] The rubber formulation performance prediction device provided in the embodiments of the present invention will be described below. The rubber formulation performance prediction device described below can be referred to in correspondence with the rubber formulation performance prediction method described above.

[0181] Figure 2 The structural block diagram of the rubber formulation performance prediction device provided in the embodiments of this application is shown below. Figure 2 The rubber formulation performance prediction device may include:

[0182] The model training module 100 is used to pre-train a large student model using a large sample set and at least one teacher model using a small sample set; wherein the number of samples in the large sample set is greater than the number of samples in the small sample set; and each set of sample data in both the large and small sample sets includes a formula sample, a process condition sample, and a performance sample.

[0183] The distillation training module 200 is used to perform reverse distillation training on the student's large model using the teacher's small model to obtain a rubber performance prediction model.

[0184] The performance prediction module 300 is used to input the rubber formulation to be predicted into the rubber performance prediction model to obtain the performance prediction results.

[0185] In one embodiment of this application, the model training module 100 includes:

[0186] The model building unit is used to build a large neural network model, which includes a neural network prediction layer, a performance correlation monitor, a first GNN refinement layer, and a second GNN refinement layer.

[0187] The preliminary prediction unit is used by the neural network prediction layer to learn and operate on the formula samples and process condition samples, and outputs an initial prediction vector; the initial prediction vector includes the initial prediction performance and the predicted process conditions.

[0188] The correlation monitoring unit is used by the performance correlation monitor to identify the strength of correlation between various performance indicators of each performance sample and obtain a performance correlation topology map.

[0189] The first refinement unit is used to modify the initial prediction performance in the initial prediction vector based on the performance correlation topology graph under the setting decoupling constraints using the first GNN refinement layer; wherein the setting decoupling constraints are that the process conditions retain self-loops and do not participate in the information transmission between performance indicators.

[0190] The second refinement unit is used to use the second GNN refinement layer to perform secondary correction on the initial prediction performance in the initial prediction vector according to the causal hierarchical prior rules, combined with the formulation samples and process condition samples, to obtain the corrected prediction vector; wherein, the causal hierarchical prior rules characterize the causal hierarchical topological relationship between the formulation, process conditions and performance indicators in rubber.

[0191] The iterative update unit is used to perform loss function calculations on the prediction vectors and optimize the large neural network model based on the loss function values ​​obtained from the calculations until a large neural network model that satisfies the function threshold is obtained, which serves as the student model.

[0192] In one optional embodiment of this application, the distillation training module 200 includes:

[0193] The basic truth unit is used to select large-scale sample data from a large-scale sample set. The student model performs performance prediction based on the formula samples and process condition samples in the large-scale sample data. The prediction results and the performance samples in the large-scale sample data are used to perform MSE loss calculation to obtain the basic truth loss.

[0194] The alignment processing unit is used to select small-scale sample data from the small-scale sample set, and perform zero-padding or truncation dimension alignment processing on the process condition samples in the small-scale sample data to make the dimension of the process condition samples the same as the set dimension, thus obtaining small-scale aligned sample data; the set dimension is the dimension required by the student large model for the input process condition parameters.

[0195] The truth loss unit is used to perform performance predictions based on formulation samples and process condition samples in small-scale aligned sample data using the student large model to obtain the large model prediction results; and to perform MSE loss calculation on the large model prediction results and the performance samples in small-scale sample data to obtain the truth loss.

[0196] The output distillation unit is used to perform performance prediction based on formula samples and process condition samples in a small-scale sample data using a teacher's small model, and obtain the small model prediction results; the MSE loss is calculated on the large model prediction results and the small model prediction results to obtain the output space distillation loss.

[0197] The characterization distillation unit is used to extract features from recipe samples in small-scale aligned sample data through the feature extraction backbone network of the student large model, obtaining the feature vector extracted by the large model; and the feature extraction backbone network of the teacher small model extracts features from recipe samples in small-scale sample data, obtaining the feature vector extracted by the small model; wherein, the feature vectors extracted by the large model and the feature vectors extracted by the small model have the same dimension; the MSE loss is calculated on the feature vectors extracted by the large model and the feature vectors extracted by the small model to obtain the characterization space distillation loss;

[0198] The loss calculation unit is used to perform a weighted summation of the basic truth loss, the truth loss, the output space distillation loss, and the characterization space distillation loss to obtain the total loss value.

[0199] The optimization iteration unit is used to iterate and optimize the network parameters of the student model based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

[0200] In an optional embodiment of this application, the distillation training module 200 is specifically used to classify each performance index into levels A, B, and C according to performance importance, and determine a corresponding level weight coefficient for each level; based on the levels and level weight coefficients corresponding to each prediction performance in the large model prediction results and small model prediction results, the distillation loss formula for the output space is applied. The output space distillation loss is obtained through computation; where, To account for the distillation loss in the output space, These are the grade weighting coefficients corresponding to grades A, B, and C, respectively. These are the predictions from the large model and the small model, respectively, for the first item in level A. Predictive performance of each performance metric; These are the predictions for level B in the large model prediction results and the small model prediction results, respectively. Predictive performance of each performance metric; These are the predictions from the large model and the small model, respectively, for the C-level. Predictive performance of each performance metric; All of these are performance prediction masks for the teacher's small model, and all have values ​​of 0 or 1. It is a very small positive number.

[0201] In an optional embodiment of this application, the distillation training module 200 is further configured to, after obtaining the prediction results of the small model, perform MSE loss calculation on the prediction results of the small model and the performance samples in the small-scale sample data to obtain the true value loss of the small model; accordingly, the basic true value loss, the output space distillation loss, the characterization space distillation loss and the true value loss of the small model are weighted and summed to obtain the total loss value; wherein, the weight coefficient corresponding to the true value loss of the small model is less than the weight coefficients corresponding to the basic true value loss, the output space distillation loss and the characterization space distillation loss respectively; and perform network parameter optimization iteration on both the student large model and the teacher small model according to the total loss value until the total loss value meets the loss threshold, and use the optimized student large model as the rubber performance prediction model.

[0202] In one optional embodiment of this application, the model training module 100 is specifically used to divide the small sample set into two independent small sample subsets; and to input the sample data from the two small sample subsets into two lightweight GNN small models for neural network training to obtain two teacher small models.

[0203] In one optional embodiment of this application, the distillation training module 200 includes:

[0204] The basic truth unit is used by the student large model to perform performance prediction on large-scale sample data selected from a large-scale sample set, and to obtain the basic truth loss based on the prediction results.

[0205] The first alignment processing unit is used to perform dimensional alignment processing on the first small-scale sample data selected from the first small-scale sample subset, such as zero-padding or truncation of process condition samples, to obtain the first small-scale aligned sample data.

[0206] The first truth loss unit is used to perform performance prediction based on the formula samples and process condition samples in the first small-scale aligned sample data using the student large model, and obtain the prediction result of the first large model; the first true loss is obtained by performing MSE loss calculation on the prediction result of the first large model and the performance samples in the first small-scale aligned sample data.

[0207] The first output distillation unit is used to perform performance prediction based on the formula samples and process condition samples in the first small-scale sample data using the first teacher small model, and obtain the prediction result of the first small model; and to perform MSE loss calculation on the prediction results of the first large model and the prediction results of the first small model to obtain the first output space distillation loss.

[0208] The first representational distillation unit is used to extract features from the formula samples in the first small-scale aligned sample data through the feature extraction backbone network of the student large model, to obtain the feature vector extracted by the first large model; the feature extraction backbone network of the first teacher small model extracts features from the formula samples in the first small-scale sample data, to obtain the feature vector extracted by the first small model; wherein, the feature vector extracted by the first large model and the feature vector extracted by the first small model have the same dimension; the MSE loss is calculated on the feature vector extracted by the first large model and the feature vector extracted by the first small model to obtain the first representational space distillation loss;

[0209] The second alignment processing unit is used to perform dimensional alignment processing on the second small-scale sample data selected from the second small-scale sample subset, such as zero-padding or truncation of process condition samples, to obtain the second small-scale aligned sample data.

[0210] The second truth loss unit is used to perform performance prediction based on the formula samples and process condition samples in the second small-scale aligned sample data using the student large model, and obtain the prediction results of the second large model; the MSE loss is calculated by performing MSE loss on the prediction results of the second large model and the performance samples in the second small-scale aligned sample data to obtain the second truth loss.

[0211] The second output distillation unit is used for the second teacher small model to perform performance prediction based on the formula samples and process condition samples in the second small-scale sample data, and to obtain the prediction results of the second small model; and to perform MSE loss calculation on the prediction results of the second large model and the prediction results of the second small model to obtain the second output space distillation loss.

[0212] The second representational distillation unit is used to extract features from the formula samples in the second small-scale aligned sample data through the feature extraction backbone network of the student large model, to obtain the feature vector extracted by the second large model; the feature extraction backbone network of the second teacher small model extracts features from the formula samples in the small-scale sample data, to obtain the feature vector extracted by the second small model; wherein, the feature vectors extracted by the second large model and the feature vectors extracted by the second small model have the same dimension; MSE loss is calculated on the feature vectors extracted by the second large model and the feature vectors extracted by the second small model to obtain the distillation loss of the second representational space;

[0213] The total loss calculation unit is used to calculate the total loss according to the total loss formula. Perform a weighted summation to obtain the total loss value; where, These are the total loss values, Based on truth value loss; These are the first truth loss and the second truth loss, respectively. These are the distillation losses in the first output space and the distillation losses in the second output space, respectively. These are the first characterization space distillation loss and the second characterization space distillation loss, respectively. All are loss weighting coefficients;

[0214] The optimization iteration unit is used to iterate and optimize the network parameters of the student model based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

[0215] The rubber formulation performance prediction device of this embodiment is used to implement the aforementioned rubber formulation performance prediction method. Therefore, the specific implementation of the device can be referred to the corresponding description in the method embodiment section above. The specific implementation can be referred to the description of the corresponding partial embodiments, which will not be repeated here.

[0216] The present invention also provides an embodiment of a rubber formulation performance prediction device. The device includes:

[0217] Memory, used to store computer programs;

[0218] A processor for executing a computer program to implement the steps of the rubber formulation performance prediction method as described in any of the preceding items.

[0219] In this embodiment, the memory can be a high-speed random access memory or a non-volatile memory, such as one or more disk storage devices, flash memory devices or other non-volatile solid-state storage devices.

[0220] The processor is used to execute computer programs. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a tensor processor (TPU), or other dedicated artificial intelligence computing chips. When the processor executes the computer program, it is able to implement all the steps of the rubber formulation performance prediction method as described in any of the preceding claims.

[0221] This device can be an industrial control computer or server integrating a high-performance GPU. By loading and executing program instructions in memory, the processor can receive user-inputted formulation data and quickly output performance prediction results, providing real-time decision support for formulation engineers.

[0222] The present invention also provides an embodiment of a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the steps of the rubber formulation performance prediction method as described in any of the preceding claims.

[0223] The computer-readable storage medium in this embodiment can be any type of volatile or non-volatile storage device, such as random access memory (RAM), read-only memory (ROM), hard disk (HDD), solid-state drive (SSD), optical disk (such as CD-ROM, DVD), or USB flash drive. When the program in the storage medium runs on a computing device, the device can execute the rubber formulation performance prediction method provided in this application, thereby achieving rapid and accurate prediction of formulation performance.

[0224] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0225] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A method for predicting the performance of rubber formulations, characterized in that, include: A large student model is obtained by training a large sample set in advance, and at least one small teacher model is obtained by training a small sample set in advance; wherein the number of samples in the large sample set is greater than the number of samples in the small sample set; and each set of sample data in both the large sample set and the small sample set includes a formula sample, a process condition sample, and a performance sample. The teacher's small model is used to perform reverse distillation training on the student's large model to obtain a rubber performance prediction model; The rubber formulation to be predicted is input into the rubber performance prediction model to obtain the performance prediction results.

2. The method for predicting the performance of rubber formulations as described in claim 1, characterized in that, Large student models are obtained by training on large-scale sample sets, including: Construct a large neural network model, which includes a neural network prediction layer, a performance correlation monitor, a first GNN refinement layer, and a second GNN refinement layer; The neural network prediction layer performs learning operations on the formula sample and the process condition sample, and outputs an initial prediction vector; the initial prediction vector includes the initial prediction performance and the predicted process conditions. The performance correlation monitor identifies the strength of correlation between various performance indicators for each performance sample and obtains a performance correlation topology map. Under the set decoupling constraints, the first GNN refinement layer corrects the initial prediction performance in the initial prediction vector according to the performance correlation topology graph; wherein the set decoupling constraints are that the process conditions retain self-loops and do not participate in the information transmission between performance indicators. The second GNN refinement layer, based on the causal hierarchical prior rules and in combination with the formulation sample and the process condition sample, performs a secondary correction on the initial prediction performance in the initial prediction vector to obtain the corrected prediction vector; wherein, the causal hierarchical prior rules characterize the causal hierarchical topological relationship between the formulation, process conditions and performance indicators in rubber. The predicted vector is subjected to a loss function operation, and the large neural network model is optimized and iterated based on the loss function value obtained from the operation until a large neural network model that satisfies the function threshold is obtained, which is then used as the student large model.

3. The method for predicting the performance of rubber formulations as described in claim 1, characterized in that, The teacher-based small model is used to perform reverse distillation training on the student-based large model to obtain a rubber performance prediction model, including: Large-scale sample data is selected from the large-scale sample set. The student model performs performance prediction based on the formula samples and process condition samples in the large-scale sample data. The prediction results and the performance samples in the large-scale sample data are subjected to MSE loss calculation to obtain the basic true value loss. The small-scale sample set selects small-scale sample data, and performs zero-padding or truncation dimension alignment processing on the process condition samples in the small-scale sample data to make the dimension of the process condition samples the same as the set dimension, thus obtaining small-scale aligned sample data; the set dimension is the dimension required by the student large model for the input process condition parameters. The student large model performs performance prediction based on the formula samples and process condition samples in the small-scale aligned sample data to obtain the large model prediction results. The MSE loss is calculated by performing MSE loss on the prediction results of the large model and the performance samples in the small-scale sample data to obtain the true value loss. The teacher mini-model performs performance prediction based on the formula samples and process condition samples in the small-scale sample data to obtain the mini-model prediction results. The MSE loss is calculated by performing MSE loss on the prediction results of the large model and the prediction results of the small model to obtain the output spatial distillation loss. The feature extraction backbone network of the student large model extracts features from the recipe samples in the small-scale aligned sample data to obtain the feature vector extracted by the large model. The feature extraction backbone network of the teacher mini-model extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the mini-model; wherein, the feature vector extracted by the large model and the feature vector extracted by the mini-model have the same dimension; MSE loss calculation is performed on the feature vectors extracted from the large model and the feature vectors extracted from the small model to obtain the distillation loss of the character space. The total loss value is obtained by weighted summing of the basic truth loss, the truth loss, the output space distillation loss, and the characterization space distillation loss. The network parameters of the student model are iteratively optimized based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

4. The method for predicting the performance of rubber formulations as described in claim 3, characterized in that, The MSE loss is calculated by performing MSE loss on the large model prediction results and the small model prediction results to obtain the output spatial distillation loss, including: The performance indicators are divided into three levels, A, B and C, according to their importance, and a corresponding weighting coefficient is determined for each level. Based on the levels corresponding to each prediction performance in the large model prediction results and the small model prediction results, and the level weight coefficients, the output space distillation loss formula is applied. The output space distillation loss is obtained through calculation; wherein, The output space distillation loss, These are the grade weighting coefficients corresponding to grade A, grade B, and grade C, respectively. These are the prediction results of the large model and the small model, respectively, for the first item in the A-level category. Predictive performance of each performance metric; These are the prediction results of the large model and the small model, respectively, for the first B level. Predictive performance of each performance metric; The prediction results of the large model and the small model are respectively the first prediction result of the C level. Predictive performance of each performance metric; All of these are performance prediction masks for the teacher's small model, and all have values ​​of 0 or 1; It is a very small positive number.

5. The method for predicting the performance of rubber formulations as described in claim 3, characterized in that, After obtaining the prediction results of the small model, the following is also included: The MSE loss is calculated by performing MSE loss on the prediction results of the small model and the performance samples in the small-scale sample data to obtain the true value loss of the small model. Accordingly, the weighted summation of the basic truth loss, the output space distillation loss, and the characterization space distillation loss yields the total loss value, including: The total loss value is obtained by weighted summing of the basic truth loss, the output space distillation loss, the representation space distillation loss, and the small model truth loss; wherein the weight coefficient corresponding to the small model truth loss is less than the weight coefficients corresponding to the basic truth loss, the output space distillation loss, and the representation space distillation loss, respectively.

6. The method for predicting the performance of rubber formulations as described in any one of claims 1 to 5, characterized in that, A small teacher model is obtained by training on a small sample set, including: The small sample set is divided into two independent subsets: a first small sample set and a second small sample set. The first small-scale sample subset is input into a lightweight GNN mini-model for neural network model training to obtain the first teacher mini-model. The second small sample subset is input into a lightweight GNN mini-model for neural network model training to obtain the second teacher mini-model.

7. The method for predicting the performance of rubber formulations as described in claim 6, characterized in that, The teacher-based small model is used to perform reverse distillation training on the student-based large model to obtain a rubber performance prediction model, including: The student large model performs performance prediction on the large-scale sample data selected from the large-scale sample set, and obtains the basic true loss based on the prediction results. The first small-scale sample data selected from the first small-scale sample subset is subjected to dimension alignment processing, such as zero-padding or truncation of process condition samples, to obtain the first small-scale aligned sample data. The student large model performs performance prediction based on the formula samples and process condition samples in the first small-scale aligned sample data to obtain the prediction result of the first large model. The first true value loss is obtained by performing MSE loss calculation on the prediction results of the first large model and the performance samples in the first small-scale aligned sample data. The first teacher mini-model performs performance prediction based on the formula samples and process condition samples in the first small-scale sample data, and obtains the prediction result of the first mini-model. Perform MSE loss calculation on the prediction results of the first large model and the prediction results of the first small model to obtain the first output space distillation loss; The feature extraction backbone network of the student large model extracts features from the recipe samples in the first small-scale aligned sample data to obtain the feature vector extracted by the first large model. The feature extraction backbone network of the first teacher mini-model extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the first mini-model; wherein, the feature vector extracted by the first large model and the feature vector extracted by the first mini-model have the same dimension; MSE loss is calculated by performing MSE loss operation on the feature vectors extracted from the first large model and the feature vectors extracted from the first small model to obtain the first representation space distillation loss. The second small-scale sample data selected from the second small-scale sample subset is subjected to dimension alignment processing, such as zero-padding or truncation of process condition samples, to obtain the second small-scale aligned sample data. The student large model performs performance prediction based on the formula samples and process condition samples in the second small-scale aligned sample data to obtain the prediction results of the second large model. The second true value loss is obtained by performing MSE loss calculation on the prediction results of the second large model and the performance samples in the second small-scale aligned sample data. The second teacher mini-model performs performance prediction based on the formula samples and process condition samples in the second small-scale sample data to obtain the prediction results of the second mini-model. The MSE loss is calculated by performing MSE loss on the prediction results of the second large model and the prediction results of the second small model to obtain the second output space distillation loss. The feature extraction backbone network of the student large model extracts features from the recipe samples in the second small-scale aligned sample data to obtain the feature vector extracted by the second large model. The feature extraction backbone network of the second teacher mini-model extracts features from the formula samples in the small-scale sample data to obtain the feature vector extracted by the second mini-model; wherein, the feature vector extracted by the second large model and the feature vector extracted by the second mini-model have the same dimension; The MSE loss is calculated by performing MSE loss operation on the feature vectors extracted from the second large model and the feature vectors extracted from the second small model to obtain the distillation loss of the second representation space. According to the total loss formula Perform a weighted summation to obtain the total loss value; where, These are the total loss values, The underlying truth loss; These are the first truth loss and the second truth loss, respectively; These are the distillation losses in the first output space and the distillation losses in the second output space, respectively. These are the first characterization space distillation loss and the second characterization space distillation loss, respectively. All are loss weighting coefficients; The network parameters of the student model are iteratively optimized based on the total loss value until the total loss value meets the loss threshold. Then, the optimized student model is used as the rubber performance prediction model.

8. A rubber formulation performance prediction device, characterized in that, include: The model training module is used to pre-train a large student model using a large sample set and at least one teacher model using a small sample set; wherein the number of samples in the large sample set is greater than the number of samples in the small sample set; and each set of sample data in both the large sample set and the small sample set includes a formula sample, a process condition sample, and a performance sample. The distillation training module is used to perform reverse distillation training on the student large model using the teacher small model to obtain a rubber performance prediction model. The performance prediction module is used to input the rubber formulation to be predicted into the rubber performance prediction model and obtain the performance prediction result.

9. A rubber formulation performance prediction device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the rubber formulation performance prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed to implement the steps of the rubber formulation performance prediction method as described in any one of claims 1 to 7.

Images