Steel thermophysical property prediction model training and steel thermophysical property prediction method and device

Abstract

The application provides a steel thermophysical property prediction model training and steel thermophysical property prediction method and device. The steel thermophysical property prediction model training method comprises the following steps: obtaining thermophysical property data calculated in batches by material calculation software; the thermophysical property data comprises steel composition, temperature and thermophysical property; a steel thermophysical property prediction model based on a bidirectional long short-term memory network is trained by using the thermophysical property data; the steel thermophysical property prediction model is configured to perform rapid reasoning on received input data to obtain a thermophysical property prediction result. The application constructs machine learning training samples based on batch calculation results of the material calculation software, which are used to train the steel thermophysical property prediction model, so that originally dispersed calculation results can form a manageable, queryable and trainable data basis. Compared with the existing separate file or manual table lookup method, the application is more conducive to batch training and version maintenance.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a method and apparatus for training a steel thermal property prediction model and predicting steel thermal properties. Background Technology

[0002] In the processes of continuous casting, rolling, heat treatment, welding, cooling control, residual stress analysis, and service performance simulation, technicians need to use a series of material thermophysical parameters such as thermal conductivity, density, Young's modulus, specific heat capacity, and coefficient of thermal expansion. These parameters typically exhibit nonlinear changes with the chemical composition and temperature of the steel, especially in the phase transformation range, where their changes are more complex and difficult to process and maintain in batches. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a steel thermal property prediction model training and steel thermal property prediction method and apparatus, which can perform batch training and version maintenance on thermal property data.

[0004] In a first aspect, embodiments of this application provide a method for training a steel thermal property prediction model, comprising: acquiring thermal property data calculated in batches by material calculation software; wherein the thermal property data includes steel composition, temperature, and thermal properties; using the thermal property data to train a steel thermal property prediction model based on a bidirectional long short-term memory network; wherein the steel thermal property prediction model is configured to perform rapid inference on the received input data to obtain thermal property prediction results.

[0005] In the above implementation process, machine learning training samples are constructed using batch calculation results from materials calculation software to train the steel thermophysical property prediction model. This transforms the previously scattered calculation results into a manageable, queryable, and trainable data foundation. Compared to existing methods involving individual files or manual table lookups, this approach is more conducive to batch training and version maintenance.

[0006] In one embodiment, the steel thermal property prediction model is a residual BiLSTM model; the BiLSTM model includes: a bidirectional LSTM layer, a LayerNorm layer, a Dropout layer, a residual connection layer, and an output fully connected layer connected in sequence.

[0007] In the above implementation process, by adopting the residual BiLSTM model and combining LayerNorm, Dropout and robust weighted loss, the model can be made to handle nonlinear thermophysical property changes under multi-element composition and wide temperature range, thereby improving the model accuracy.

[0008] In one embodiment, the material calculation software is JMatPro Dynamic; the thermal property data is configured to be obtained by batch calculation of thermal property parameters of steels with different chemical compositions at different temperatures using JMatPro Dynamic. In the above implementation process, by setting the material calculation software to JMatPro Dynamic, the calculation results of JMatProDynamic can reflect the comprehensive influence of composition, temperature and phase transformation behavior on thermal property parameters, thereby improving the accuracy of the steel thermal property prediction model.

[0009] In one embodiment, before training a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data, the method further includes: dividing the thermal property data into a training set, a test set, and a validation set; wherein the training set is used for model learning, the validation set is used for model selection and early stopping control during the training process, and the test set is used for prediction accuracy evaluation. In the above implementation process, by dividing the thermal property data into training set, test set and validation set, the training set is used for model learning, the validation set is used for model selection and early stopping control during the training process, and the test set is used for prediction accuracy evaluation. The model can be validated and tested during the model training process, the model accuracy can be understood in time, and the model training can be stopped in time when the current best model is determined, so as to avoid overfitting caused by overtraining and improve the accuracy and efficiency of model training.

[0010] In one embodiment, training a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data includes: calculating the validation set loss after each round of training of the steel thermal property prediction model; saving the steel thermal property prediction model with the lowest validation set loss as the best model; and executing an early stopping strategy to stop training the steel thermal property prediction model if the validation set loss does not decrease for a consecutive preset number of rounds.

[0011] In the above implementation process, by setting the validation set loss to be calculated after each round of model training, and saving the steel thermophysical property prediction model with the lowest validation set loss as the best model, and then executing the early stopping strategy based on the best model, the risk of overfitting can be reduced, training oscillations can be reduced, the model's generalization ability to unseen component temperature sequences can be improved, and the model's accuracy can be improved.

[0012] In one embodiment, before acquiring the thermal property data calculated in batches by the material calculation software, the method further includes: performing batch calculations on different steel compositions and temperature paths using the material calculation software to generate corresponding thermal property data; storing the thermal property data in a database in the form of structured data; wherein each data record includes at least a composition field, a temperature field, and a performance field.

[0013] In the above implementation process, the thermophysical data calculated by the material calculation software is stored in the database in the form of structured data, which facilitates batch data management, version maintenance, query filtering, training set construction and subsequent expansion.

[0014] In one embodiment, before acquiring the thermophysical property data calculated in batches by the material calculation software, the method further includes: generating a component combination identifier based on the content of each element, wherein the same component combination corresponds to multiple thermophysical property records at different temperatures; wherein the component combination identifier is configured to group database records and sort them according to a set field, so as to organize multiple records under the same component combination into a set sequence sample.

[0015] In the above implementation process, through the above organizational form, the discrete structured data calculated by the material calculation software can be converted into sequence samples suitable for training the steel thermal property prediction model, so that the steel thermal property prediction model can learn the continuous law of the change of ordinary steel thermal property parameters with temperature, which facilitates batch data management, version maintenance, query filtering, training set construction and subsequent expansion.

[0016] In one embodiment, before training a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data, the method further includes: cleaning the thermal property data; performing a log1p transformation on the cleaned thermal property data; standardizing the transformed thermal property data; grouping the standardized thermal property data by composition and sorting it by temperature to construct a composition-temperature sequence sample; and training a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data includes: training a steel thermal property prediction model based on a bidirectional long short-term memory network using the composition-temperature sequence sample. In the above implementation process, the thermal property data is preprocessed before model training to avoid the cross-occurrence of data for the same component combination in different datasets, thereby more objectively evaluating the model's predictive ability for new component combinations.

[0017] In one embodiment, the method further includes: after the steel thermal property prediction model has been trained, exporting the steel thermal property prediction model as an ONNX model and saving the input standardized parameters and output standardized parameters; wherein the ONNX model is configured for front-end loading.

[0018] In the above implementation process, by exporting the steel thermophysical property prediction model as an ONNX model for front-end loading, the model training environment and the front-end inference environment can be decoupled, enabling local fast inference and instantaneous response for steel thermophysical property prediction on the front end. Compared with directly calling JMatPro Dynamics or other complex calculation programs, ONNX Runtime inference does not require repeated execution of complex thermodynamic calculations on the front end. The front end can preload the model and standardized parameters, enabling rapid response when users adjust the composition or temperature.

[0019] Secondly, embodiments of this application also provide a method for predicting the thermal properties of steel, comprising: acquiring steel information input by a user; inputting the steel information into a steel thermal property prediction model in the first aspect or any possible implementation of the first aspect; and outputting steel thermal property prediction data through the steel thermal property prediction model.

[0020] Thirdly, embodiments of this application also provide a steel thermal property prediction model training device, comprising: a first acquisition module, used to acquire thermal property data calculated in batches by material calculation software; wherein the thermal property data includes steel composition, temperature, and thermal properties; a training module, used to train a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data; wherein the steel thermal property prediction model is configured to perform rapid inference on the received input data to obtain thermal property prediction results.

[0021] Fourthly, embodiments of this application also provide a steel thermal property prediction device, comprising: a second acquisition module for acquiring steel information input by a user; an input module for inputting the steel information into the first aspect, or a steel thermal property prediction model in any possible implementation of the first aspect; and an output module for outputting steel thermal property prediction data through the steel thermal property prediction model.

[0022] Fifthly, embodiments of this application also provide an electronic device, including: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the machine-readable instructions are executed by the processor to perform the steps of the method in the first aspect above, or any possible implementation of the first aspect, or any possible implementation of the second aspect above, or any possible implementation of the second aspect.

[0023] In a sixth aspect, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the method described in the first aspect, or any possible implementation of the first aspect, or the second aspect, or any possible implementation of the second aspect.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, specific embodiments are described below in conjunction with the accompanying drawings. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A block diagram illustrating an electronic device provided in an embodiment of this application; Figure 2 A flowchart illustrating the steel thermophysical property prediction model training method provided in this application embodiment; Figure 3 A flowchart of the steel thermophysical property prediction method provided in the embodiments of this application; Figure 4 A schematic diagram of the functional modules of the steel thermophysical property prediction model training device provided in the embodiments of this application; Figure 5 A schematic diagram of the functional modules of the steel thermal property prediction device provided in the embodiments of this application. Detailed Implementation

[0027] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0028] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] In existing technologies, the thermophysical properties of steel are mainly obtained through the following methods: First, consult material handbooks, standards databases, or publicly available literature. This method is suitable for a small number of standard steel grades, but it is insufficient for ordinary steels, low-alloy steels, micro-alloy steels, or steel grades with continuously varying compositions within a company. Material handbooks typically cannot provide complete data on thermophysical property curves for different combinations of elements such as C, Si, Mn, Cr, Ni, Mo, Nb, Ti, V, and B.

[0030] Second, thermophysical parameters can be obtained through experimental testing. For example, data such as thermal expansion, thermal conductivity, specific heat capacity, and elastic modulus can be obtained using thermal dilatometers, differential scanning calorimeters, laser thermal conductivity meters, ultrasonic methods, or mechanical testing equipment. While this method provides experimentally supported results, it is time-consuming, costly, and difficult to conduct batch tests on a large number of component combinations and over a wide temperature range.

[0031] Third, materials calculation software such as JMatPro is used to perform thermodynamic, kinetic, and physical property calculations based on the chemical composition and thermal process conditions of the steel. JMatPro calculations can obtain thermophysical parameters and phase transformation-related information at different temperatures based on material composition, temperature path, and phase transformation process. This method can obtain thermophysical property data relatively systematically, but for a large number of composition combinations and interactive real-time prediction software, directly calling the calculation software one by one still has efficiency limitations.

[0032] Fourth, existing data can be used to make predictions through interpolation, empirical formulas, or ordinary regression models. This method is simple to implement, but the effects of multi-element composition, continuous temperature changes, and nonlinear changes in phase transition intervals on thermophysical properties are complex, and simple interpolation or ordinary single-point regression models are difficult to stably represent the complete temperature curve.

[0033] However, the inventors of this application have discovered through long-term research that the prior art has the following defects: (1) Material handbooks and public databases have limited coverage. The thermophysical properties of ordinary steel are affected by the combined effects of multiple elemental compositions and temperature, especially in the phase transformation range, where the property curves may exhibit nonlinear changes. Traditional handbook data usually only cover typical grades and are difficult to cover a large number of compositional combinations.

[0034] (2) The experimental testing cost is high. If parameters such as thermal conductivity, density, and Young's modulus are obtained through experiments, it is necessary to prepare samples, set test conditions, and perform data processing, which is not suitable for large-scale component screening and rapid process design.

[0035] (3) Directly calling professional calculation software has limited efficiency. JMatPro Dynamic is suitable for generating high-quality data, but each calculation requires setting the composition and temperature path. When users frequently adjust the steel composition or temperature on the front end, directly calling complex calculation software is difficult to meet the needs of real-time interaction.

[0036] (4) Existing machine learning material performance predictions are mostly single-point performance predictions. For the thermophysical property curves of the same steel composition in a continuous temperature range, ordinary single-point models often only focus on a certain temperature point and it is difficult to utilize the continuous change relationship between points before and after the temperature sequence.

[0037] (5) Existing solutions usually only focus on one aspect of data, algorithms or engineering simulation, and lack an integrated software process from batch computing data entry, database reading, data preprocessing, sequence model training, model evaluation, ONNX export, Java front-end deployment to interactive visualization prediction.

[0038] In view of this, this application proposes a training method for a steel thermal property prediction model. By storing the batch calculation results of JMatProDynamic into a thermal property database, the residual BiLSTM sequence prediction model is trained using the composition-temperature-thermal property data in the database, and instantaneous response is achieved through ONNX and Java front-end deployment, thereby improving the above-mentioned problems.

[0039] To facilitate understanding of this embodiment, the electronic equipment for implementing the steel thermal property prediction model training method and / or steel thermal property prediction method disclosed in this application embodiment will first be described in detail.

[0040] like Figure 1 The diagram shown is a block illustration of an electronic device. The electronic device 100 may include a memory 111 and a processor 113. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device 100. For example, the electronic device 100 may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0041] The aforementioned memory 111 and processor 113 are electrically connected to each other directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The aforementioned processor 113 is used to execute executable modules stored in the memory.

[0042] The memory 111 can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory 111 stores programs, and the processor 113 executes these programs upon receiving execution instructions. The methods executed by the electronic device 100 as defined in any embodiment of this application can be applied to the processor 113, or implemented by the processor 113.

[0043] The aforementioned processor 113 may be an integrated circuit chip with signal processing capabilities. The processor 113 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a digital signal processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor.

[0044] The electronic device 100 in this embodiment can be used to execute various steps in the various methods provided in the embodiments of this application. The implementation process of the steel thermophysical property prediction model training method is described in detail below through several embodiments.

[0045] Please see Figure 2 This is a flowchart of the steel thermophysical property prediction model training method provided in the embodiments of this application. The following will describe... Figure 2 The specific process shown will be explained in detail.

[0046] Step 201: Obtain the thermal property data calculated in batches by the material calculation software.

[0047] The thermal property data includes steel composition, temperature, thermal properties, thermal conductivity, density, and Young's modulus.

[0048] In one embodiment, prior to step 201, the method further includes: performing batch calculations on different steel compositions and temperature paths using material calculation software to generate corresponding thermophysical property data. The steel may include ordinary steel and high-quality steel.

[0049] The material calculation software accepts inputs including various elements such as Fe, Al, Cr, Cu, Mn, Mo, Nb, Ni, O, Si, Ta, Ti, V, W, B, C, N, P, and S. The calculation covers a temperature range from room temperature to the liquidus temperature. This temperature range can be set according to the steel hot working and heat treatment scenarios, for example, from 1650℃ to 25℃, with step sizes set to 1℃, 2℃, 5℃, or other preset values.

[0050] The materials calculation software here is based on the theories of materials thermodynamics and phase transformation kinetics, and calculates the thermophysical parameters at different temperatures according to the chemical composition, temperature path and phase transformation process of steel.

[0051] Optionally, the material calculation software may include BiLSTM model, LSTM, GRU, Transformer, Temporal Convolutional Network, MLP, XGBoost, Random Forest, Gaussian process regression, physically constrained neural network, hybrid expert model or multi-model ensemble, etc., and the material calculation software can be configured according to the actual situation.

[0052] Step 202: Using thermal property data, train a steel thermal property prediction model based on a bidirectional long short-term memory network.

[0053] The steel thermal property prediction model is configured to perform rapid inference on the received input data to obtain the thermal property prediction results.

[0054] The inputs to the steel thermal property prediction model here are temperature and multiple elemental compositions, and the outputs are thermal conductivity, density, and Young's modulus.

[0055] In the above implementation process, machine learning training samples are constructed using batch calculation results from materials calculation software to train the steel thermophysical property prediction model. This transforms the previously scattered calculation results into a manageable, queryable, and trainable data foundation. Compared to existing methods involving individual files or manual table lookups, this approach is more conducive to batch training and version maintenance.

[0056] In one possible implementation, the steel thermal property prediction model is a residual BiLSTM model.

[0057] In one embodiment, the BiLSTM model includes: a bidirectional LSTM layer, a LayerNorm layer, a Dropout layer, a residual connection layer, and an output fully connected layer connected in sequence.

[0058] The bidirectional LSTM layer is used to extract the forward and reverse context features of the temperature sequence, the LayerNorm layer is used to stabilize the feature distribution between different component sequences, the Dropout layer is used to reduce the risk of overfitting, the residual connection layer is used to preserve the direct mapping relationship between temperature and component inputs and thermal properties, and the output fully connected layer is used to simultaneously output three thermal property parameters: thermal conductivity, density and Young's modulus.

[0059] Specifically, during BiLSTM model training, the input layer receives multiple input features: temperature and various elemental compositions. The bidirectional LSTM layer is used to learn the correlation between consecutive temperature points in the temperature sequence. Since thermophysical parameters typically change continuously with temperature, and their trends may change during phase transitions, bidirectional sequence modeling can utilize the contextual information of complete temperature curves for the same composition. During model training, the input features sorted by temperature for the same steel composition are used as the sequence input, where each time step includes a temperature value and various elemental compositions. The forward LSTM in the bidirectional LSTM calculates the hidden states sequentially in the forward direction of the temperature sequence to extract temperature change information before the current temperature point; the backward LSTM calculates the hidden states sequentially in the reverse direction of the temperature sequence to extract temperature change information after the current temperature point. At each temperature point, the system concatenates the forward and backward hidden states to form a bidirectional sequence feature, which is then output as the corresponding thermal conductivity, density, and Young's modulus through LayerNorm, Dropout, and a fully connected layer. Simultaneously, residual connections map the original input features to the output to enhance the direct expressive power of temperature and composition on the prediction results.

[0060] The LayerNorm layer normalizes the LSTM output, improving training stability. The Dropout layer reduces the risk of overfitting. The fully connected output layer maps the bidirectional LSTM output to three thermal property parameters. The residual connection layer maps the original input features at each temperature point to the same dimension as the output thermal property parameters through a linear layer and adds it to the output of the bidirectional LSTM obtained through the fully connected layer. Since the thermal conductivity, density, and Young's modulus of ordinary steel are directly correlated with temperature and elemental composition, this residual connection can directly transmit the fundamental influence of temperature and composition on thermal properties to the output, allowing the bidirectional LSTM to primarily learn the temperature sequence context and the nonlinear correction relationship within the phase transition interval. In this way, the model output can be represented as the sum of the fundamental direct mapping term and the sequence nonlinear correction term, thereby reducing the difficulty of model training, reducing information attenuation during the transmission of deep sequence features, and improving the continuity, smoothness, and fitting stability of the thermal property curve prediction.

[0061] Optionally, the input to the residual BiLSTM model can be data that can establish the correspondence between the composition, temperature and thermophysical parameters of ordinary steel, such as batch calculation results from JMatPro Dynamics, Thermo-Calc, Pandat, DICTRA, finite element simulation, experimental test data, historical test data from enterprises, material handbook data or multi-source data fusion results.

[0062] In one embodiment, during the training of the steel thermophysical property prediction model, a robust loss function can be used, with weighting applied to high-value regions. This loss function uses a squared error form when the error is small and a linear error form when the error is large, to reduce the impact of abnormal errors on model training. Simultaneously, samples with target thermophysical property values ​​exceeding a preset threshold are assigned higher weights to improve the model's predictive stability and fitting ability in key thermophysical property ranges.

[0063] For example, the AdamW optimizer can be used to train a residual BiLSTM model with an initial learning rate of 5e-4 and a weight decay coefficient of 1e-5. Weight decay is used to limit the model weights from becoming too large during parameter updates, reducing the risk of overfitting the model to local noise in the training data.

[0064] Alternatively, a cosine annealing learning rate scheduling strategy can be adopted to maintain a relatively high learning rate in the early stage of training to accelerate convergence, and gradually reduce the learning rate in the later stage of training to reduce the parameter update amplitude and improve the model convergence stability.

[0065] The residual BiLSTM model uses a robust loss function. This loss function is based on Huber loss, and the model's predicted value is... The true value is y, and the error is e= -y. When |e| is less than the preset threshold δ, the loss function adopts the squared error form of 0.5e²; when |e| is greater than or equal to δ, the loss function adopts the linear error form of δ(|e|-0.5δ) to reduce the impact of abnormal errors on model training. Furthermore, a weight function is set according to the true target value y. When y is greater than the preset high-value threshold θ, the loss of that sample is multiplied by the high-value weight w_high; when y is not greater than θ, the weight is 1. Therefore, the loss of a single output component can be expressed as L(e,y)=w(y)·L_huber(e).

[0066] In one embodiment, δ=0.5, θ=0.3, and w_high=4.0. For each component temperature sequence, the weighted robust loss of each temperature point and each thermophysical property output component is averaged to obtain the sequence loss; then, the sequence loss within the batch is averaged to obtain the model training loss. This loss function can improve the model's ability to fit the high-value range of key thermophysical properties while suppressing the influence of outliers.

[0067] In the above implementation process, by adopting the residual BiLSTM model and combining LayerNorm, Dropout and robust weighted loss, the model can be made to handle nonlinear thermophysical property changes under multi-element composition and wide temperature range, thereby improving the model accuracy.

[0068] In one possible implementation, the materials calculation software is JMatPro Dynamic.

[0069] The thermophysical data configuration was obtained by batch calculation of thermophysical parameters of steels with different chemical compositions at different temperatures using JMatPro Dynamics. The JMatPro Dynamic calculations here are based on the theories of materials thermodynamics and phase transformation kinetics, and calculate the thermophysical parameters at different temperatures according to the chemical composition of the steel, the temperature path, and the phase transformation process.

[0070] Specifically, JMatPro Dynamics can calculate thermophysical parameters by determining the phase composition at different temperatures based on the Gibbs free energy minimization relationship and the elemental mass conservation relationship. Furthermore, steel may undergo microstructural transformations such as austenite, ferrite, pearlite, bainite, and martensite during cooling; changes in the proportions of these different phases affect parameters such as thermal conductivity, density, and Young's modulus.

[0071] The minimization relation can be expressed as: ; The law of conservation of element mass can be expressed as: ; Cooling rate can be expressed as: ; Thermal conductivity can be expressed as: ; Density can be expressed as: ; Young's modulus can be expressed as: .

[0072] As can be seen from the above, the calculation results of JMatPro Dynamics can reflect the combined effects of composition, temperature, and phase change behavior on thermophysical parameters.

[0073] In the above implementation process, by setting the material calculation software to JMatPro Dynamic, the calculation results of JMatProDynamic can reflect the comprehensive influence of composition, temperature and phase transformation behavior on thermal property parameters, thereby improving the accuracy of the steel thermal property prediction model.

[0074] In one possible implementation, prior to step 202, the method further includes: dividing the thermal property data into a training set, a test set, and a validation set.

[0075] The training set is used for model learning, the validation set is used for model selection and early stopping control during the training process, and the test set is used for prediction accuracy evaluation. In one embodiment, before model training, all thermal property data is randomly shuffled, and approximately 20% of the thermal property data is designated as a test set, while the remaining approximately 80% is used as a training candidate set. The test set does not participate in parameter updates or early stopping judgment during model training and model selection; it is only used for independent performance evaluation after model training is completed.

[0076] Subsequently, during model training, the candidate training set is further divided into a training set and a validation set, with approximately 80% and 20% allocated to each. The training set is used for updating model parameters, while the validation set is used to calculate the validation loss, determine model convergence, implement early stopping strategies, and save the best-performing model.

[0077] In the above implementation process, by dividing the thermal property data into training set, test set and validation set, the training set is used for model learning, the validation set is used for model selection and early stopping control during the training process, and the test set is used for prediction accuracy evaluation. The model can be validated and tested during the model training process, the model accuracy can be understood in time, and the model training can be stopped in time when the current best model is determined, so as to avoid overfitting caused by overtraining and improve the accuracy and efficiency of model training.

[0078] In one possible implementation, step 202 includes: calculating the validation set loss after each round of training of the steel thermal property prediction model; saving the steel thermal property prediction model with the lowest validation set loss as the best model; and executing an early stopping strategy to stop training the steel thermal property prediction model if the validation set loss does not decrease for a consecutive preset number of rounds.

[0079] Specifically, after each training round of the steel thermal property prediction model, the validation set loss is calculated, and the model with the lowest validation set loss is saved as the current best model. When the validation set loss does not decrease for several consecutive rounds, the system executes an early stopping strategy to stop training and prevent the model from overfitting due to continued training. In the current embodiment, the validation set value of the best model is less than a set accuracy (e.g., 0.005), indicating that the steel thermal property prediction model has met the preset training accuracy requirement.

[0080] Once the optimal model is determined, it can be further evaluated independently using a test set. Test set evaluation metrics include RMSE, MAPE, and R², used to assess the model's predictive ability on new component sequence samples that were not involved in training and validation. If the model does not meet the preset accuracy requirements, the process returns to the model training module to continue adjusting model parameters, training epochs, loss function parameters, or data preprocessing strategies. When the model meets the preset accuracy requirements, the optimal model is saved, and the process proceeds to the subsequent model export process.

[0081] In one implementation, a validation set value less than a set precision (e.g., 0.005) is used as the criterion for determining whether the model meets the precision requirements. When the model validation set value is not less than the set precision, the system returns to the model training module to continue training or parameter tuning; when the model validation set value is less than the set precision, the system saves the best model and proceeds to the model export process.

[0082] For example, the early stopping tolerance rounds can be set to 12. By combining weight decay, learning rate scheduling, and early stopping strategies, the system can reduce the risk of overfitting, decrease training oscillations, improve the model's generalization ability to unseen component temperature sequences, and make the optimal model validation set smaller than the set accuracy.

[0083] In this embodiment of the application, the prediction of steel thermal properties is transformed from a traditional single-point regression problem into a multi-property curve prediction problem of temperature sequence under composition constraints.

[0084] Specifically, the process first reads batch calculation data from JMatPro Dynamic stored in the thermal property database. This data includes the elemental composition of ordinary steel, temperature, and corresponding thermal conductivity, density, and Young's modulus. Then, based on the elemental composition of the steel, data at different temperature points with the same composition are grouped together and arranged in temperature order to form a thermal property temperature sequence sample corresponding to the same steel composition.

[0085] Based on this, a residual BiLSTM model is constructed. This residual BiLSTM model uses temperature and multiple elemental compositions of steel as input features, and thermal conductivity, density, and Young's modulus as joint output targets. The residual BiLSTM model structure includes a bidirectional LSTM layer, a LayerNorm layer, a Dropout layer, a fully connected output layer, and a residual connection layer. Specifically, the bidirectional LSTM is used to extract the forward and reverse contextual information of the temperature sequence; LayerNorm is used to stabilize the feature distribution between different composition sequences; Dropout is used to reduce overfitting; and the residual connection layer is used to preserve the direct mapping relationship between temperature and composition inputs and thermal properties.

[0086] Meanwhile, this application embodiment employs a robust weighted loss training strategy oriented towards key thermal property ranges. This strategy combines Huber robust loss with a high-value range weighting mechanism, reducing the impact of outlier data while improving the model's fitting ability to key thermal property ranges. Through the above algorithm combination, this application embodiment can simultaneously predict three types of thermal property curves—thermal conductivity, density, and Young's modulus—in the same model, and the current model training accuracy reaches a level where the validation set accuracy is less than the set accuracy.

[0087] The batch computation data from JMatPro Dynamics is reconstructed into training samples of "composition-temperature sequence-multi-property curves", and the thermophysical property curves of ordinary steel are jointly predicted by residual BiLSTM and robust weighted loss.

[0088] In the above implementation process, by setting the validation set loss to be calculated after each round of model training, and saving the steel thermophysical property prediction model with the lowest validation set loss as the best model, and then executing the early stopping strategy based on the best model, the risk of overfitting can be reduced, training oscillations can be reduced, the model's generalization ability to unseen component temperature sequences can be improved, and the model's accuracy can be improved.

[0089] In one possible implementation, prior to step 201, the method further includes: performing batch calculations on different steel compositions and temperature paths using material calculation software to generate corresponding thermal property data; and storing the thermal property data in a database in the form of structured data.

[0090] Each data record includes at least a composition field, a temperature field, and a performance field.

[0091] The composition field here records the mass percentage of each element in the steel, including elements such as Fe, Al, Cr, Cu, Mn, Mo, Nb, Ni, O, Si, Ta, Ti, V, W, B, C, N, P, and S. The temperature field records the calculation temperature corresponding to this set of thermophysical property data. The performance field records the thermophysical property parameters calculated by material calculation software under this composition and temperature condition, including thermal conductivity, density, and Young's modulus.

[0092] In one embodiment, the data in the database can be organized according to the following fields: composition_id: Component ID; Fe, Al, Cr, Cu, Mn, Mo, Nb, Ni, O, Si, Ta, Ti, V, W, B, C, N, P, S: element content; temperature; thermal_conductivity: thermal conductivity; density; youngs_modulus: Young's modulus; data_source: Data source, such as JMatPro Dynamic; calculation_id: Calculation task number; created_time: The time when the data was generated.

[0093] In the above implementation process, the thermophysical data calculated by the material calculation software is stored in the database in the form of structured data, which facilitates batch data management, version maintenance, query filtering, training set construction and subsequent expansion.

[0094] In one possible implementation, prior to step 202, the method further includes: generating component combination identifiers based on the content of each element, with the same component combination corresponding to multiple thermophysical property records at different temperatures.

[0095] The component combination identifier is configured to group database records and sort them according to a set field, so as to organize multiple records under the same component combination into a set sequence sample.

[0096] Understandably, to facilitate subsequent model training, a component combination identifier can be generated based on the content of each element. The same component combination corresponds to multiple thermophysical property records at different temperatures. During model training, the database records can be grouped according to this component combination identifier and sorted by the temperature field, organizing multiple records under the same component combination into a single temperature sequence sample.

[0097] In other words, the original structure in the database is: elemental composition + temperature → thermal conductivity, density, Young's modulus; during model training, it is further organized into: temperature sequence under the same elemental composition → performance sequence of thermal conductivity, density, and Young's modulus as temperature changes.

[0098] Alternatively, the database can be a relational database such as MySQL, PostgreSQL, SQLite, SQL Server, or Oracle, or it can be a MongoDB, Redis, Parquet file, HDF5 file, CSV file collection, Excel data table, or special binary data file.

[0099] It should be understood that the embodiments of this application model the prediction of steel thermal properties as a sequential prediction problem of composition-temperature sequence-thermal property curve. Instead of predicting a single temperature point in isolation, it learns the curve law of steel with the same composition under continuous temperature changes, which is beneficial to improve the consistency of the prediction of thermal conductivity, density and Young's modulus curves.

[0100] In the above implementation process, through the above organizational form, the discrete structured data calculated by the material calculation software can be converted into sequence samples suitable for training the steel thermal property prediction model, so that the steel thermal property prediction model can learn the continuous law of the change of ordinary steel thermal property parameters with temperature, which facilitates batch data management, version maintenance, query filtering, training set construction and subsequent expansion.

[0101] In one possible implementation, prior to step 202, the method further includes: cleaning the thermal property data; performing a log1p transformation on the cleaned thermal property data; standardizing the transformed thermal property data; and grouping the standardized thermal property data by composition and sorting it by temperature to construct a composition-temperature sequence sample.

[0102] Specifically, cleaning thermal property data can include: for completely duplicate calculation records, retaining only one valid record and deleting duplicate records to avoid data with the same composition, temperature, and performance value appearing repeatedly in the training set, affecting the model training weights; setting data filtering windows according to the actual possible density, Young's modulus, and thermal conductivity of ordinary steel to filter outliers (for example, keeping the density within a reasonable range (2-10 g / cm3), and keeping the Young's modulus and thermal conductivity within positive values ​​and not exceeding the set upper limit (300 GPa and 100 W / (m·K) respectively)), which can reduce the impact of calculation outliers, exported outliers, or data entry anomalies on model training.

[0103] For the thermal property data after cleaning, a log1p transformation can be performed, that is, for each output value y, calculate... Because thermal conductivity, density, and Young's modulus have different numerical ranges, direct training may cause the model to be biased towards outputs with larger numerical magnitudes. The log1p transformation can reduce the magnitude difference of output parameters and improve training stability.

[0104] Then, standardization can be performed on the input and output variables separately. The process is as follows: , and These represent the mean and standard deviation of each field in the input feature matrix, respectively, and the standardized parameters are saved. The input variables include temperature and various elemental compositions, while the output variables are the standardized thermal conductivity, density, and Young's modulus after log1p transformation.

[0105] After the above processing is completed, the content of multiple elements (e.g., 19 elements) can be used as the composition grouping field to group multiple temperature point records with the same composition combination into the same group. Within each group, the records are sorted according to the temperature field, forming an input sequence tensor and an output sequence tensor. Each time step in the input sequence includes a temperature value and the corresponding multiple elemental compositions, while each time step in the output sequence includes the thermal conductivity, density, and Young's modulus at that temperature.

[0106] After generating temperature sequence samples, the dataset is divided using component sequence samples as the basic unit, rather than using individual temperature points as the unit. Specifically, all temperature points under the same elemental combination are treated as a whole sequence sample for partitioning, and partial temperature points under the same component combination are not assigned to the training set and partial temperature points are assigned to the validation set or test set.

[0107] In the above implementation process, the thermal property data is preprocessed before model training to avoid the cross-occurrence of data for the same component combination in different datasets, thereby more objectively evaluating the model's predictive ability for new component combinations.

[0108] In one possible implementation, the method further includes: after the steel thermal property prediction model has been trained, exporting the steel thermal property prediction model as an ONNX model and saving the input standardized parameters and the output standardized parameters.

[0109] The trained BiLSTM model is configured to be converted to ONNX format, while saving both the input and output normalized parameters. ONNX format facilitates model deployment in different programming languages ​​and runtime environments, reducing dependence on the training framework.

[0110] In one embodiment, the ONNX model is configured for loading by the front end.

[0111] The front-end here is used to load the ONNX model, normalize parameters, and create a visualization interface. It executes model inference through ONNXRuntime and displays thermal conductivity, density, and Young's modulus curves through the front-end graphical interface. The purpose of using ONNX and the front-end is to achieve instantaneous front-end response, meaning that after the user adjusts the composition or temperature, the software can quickly refresh the current point prediction value and complete thermophysical property curves.

[0112] Optionally, the front end may include a user-interactive input module for receiving steel composition and temperature input by the user. Users can input or drag the contents of elements such as Fe, C, Si, Mn, Cr, Ni, Mo, Nb, Ti, and V, as well as temperature points.

[0113] After the user inputs the steel composition, a temperature sequence within a set temperature range is generated, and each temperature point is concatenated with the same steel composition to form the model input. The front end loads the ONNX model file through the ONNX Runtime and reads the input mean, input standard deviation, output mean, and output standard deviation. After the user inputs the ordinary steel composition, a temperature sequence is generated, concatenated to form the model input tensor, and after input standardization, it is sent to the ONNX model for inference. The model output results are inversely standardized and inversely transformed using expm1 to obtain the actual predicted values ​​of thermal conductivity, density, and Young's modulus.

[0114] In front-end interaction, when a user adjusts only the temperature point without changing the composition, the system can directly read the predicted value of the corresponding temperature point from the cached complete prediction curve without re-executing the complete model inference. This approach can further improve the interaction response speed and achieve instantaneous front-end response.

[0115] In one embodiment, the front end is Java.

[0116] This application's embodiments transform offline complex material calculation results into interactive, deployable, and instantaneously responsive software for predicting the thermophysical properties of ordinary steel.

[0117] Specifically, in existing technologies, steel thermophysical parameters usually rely on looking up tables in material handbooks, experimental testing, or calculations by professional software such as JMatPro. When users need to frequently adjust steel composition, temperature range, or quickly view thermophysical property curves, directly calling complex calculation software or manually looking up tables results in slow response speeds and makes it difficult to meet the needs of interactive prediction.

[0118] This application embodiment uses JMatPro Dynamics to batch calculate and form a thermal property database, then trains a surrogate model, and exports the trained model in ONNX format. Local inference is then performed in the front-end using ONNX Runtime. Users can input or drag the elemental composition and temperature of ordinary steel in the front-end visual interface, and the software instantly outputs predicted values ​​of thermal conductivity, density, Young's modulus, and corresponding temperature curves.

[0119] The purpose of using ONNX and the front-end is to achieve instantaneous front-end response. ONNX models facilitate cross-platform deployment and rapid inference, while the front-end facilitates the construction of desktop interactive interfaces and engineering software integration interfaces. With the two combined, the prediction stage does not require repeated calls to JMatPro, nor does it rely on Python training environments or remote server calculations, making it suitable for high-frequency interactive material design, initial screening of process parameters, and pre-processing scenarios for engineering simulations.

[0120] In the above implementation process, by exporting the steel thermophysical property prediction model as an ONNX model for front-end loading, the model training environment and the front-end inference environment can be decoupled, enabling local fast inference and instantaneous response for steel thermophysical property prediction on the front end. Compared with directly calling JMatPro Dynamics or other complex calculation programs, ONNX Runtime inference does not require repeated execution of complex thermodynamic calculations on the front end. The front end can preload the model and standardized parameters, enabling rapid response when users adjust the composition or temperature.

[0121] Please see Figure 3 This is a flowchart of the steel thermophysical property prediction method provided in the embodiments of this application. The following will describe... Figure 3 The specific process shown will be explained in detail.

[0122] Step 301: Obtain the steel information input by the user.

[0123] Optionally, the steel information may include input variables such as temperature and various elemental compositions such as Fe, Al, Cr, Cu, Mn, Mo, Nb, Ni, O, Si, Ta, Ti, V, W, B, C, N, P, and S.

[0124] Step 302: Input the steel information into the steel thermal property prediction model in the above embodiment.

[0125] Step 303: Output steel thermal property prediction data through the steel thermal property prediction model.

[0126] Optionally, the output predicted thermal properties of steel may include thermal conductivity, density, Young's modulus, etc. This predicted thermal properties can be selected according to the actual situation.

[0127] Based on the same application concept, this application also provides a steel thermal property prediction model training device corresponding to the steel thermal property prediction model training method. Since the principle of the device in this application is similar to that of the aforementioned steel thermal property prediction model training method, the implementation of the device in this application can refer to the description in the above-mentioned method embodiments, and the repeated parts will not be described again.

[0128] Please see Figure 4 This is a functional module diagram of the steel thermal property prediction model training device provided in this application embodiment. Each module in the steel thermal property prediction model training device in this embodiment is used to execute the steps in the above method embodiments. The steel thermal property prediction model training device includes a first acquisition module 401 and a training module 402; wherein, The first acquisition module 401 is used to acquire thermal property data calculated in batches by material calculation software; wherein, the thermal property data includes steel composition, temperature, and thermal properties.

[0129] Training module 402 is used to train a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data; wherein the steel thermal property prediction model is configured to perform fast inference on the received input data to obtain thermal property prediction results.

[0130] In one possible implementation, the steel thermophysical property prediction model training device further includes: a partitioning module for dividing the thermophysical property data into a training set, a test set, and a validation set; wherein the training set is used for model learning, the validation set is used for model selection and early stopping control during the training process, and the test set is used for prediction accuracy evaluation. In one possible implementation, the training module 402 is specifically used to: calculate the validation set loss after each round of training of the steel thermal property prediction model; save the steel thermal property prediction model with the lowest validation set loss as the best model; and if the validation set loss does not decrease for a consecutive preset number of rounds, execute an early stopping strategy to stop training the steel thermal property prediction model.

[0131] In one possible implementation, the steel thermophysical property prediction model training device further includes: a storage module, used to perform batch calculations on different steel compositions and temperature paths using the material calculation software to generate corresponding thermophysical property data; and to store the thermophysical property data in a database in the form of structured data; wherein each data record includes at least a composition field, a temperature field, and a performance field.

[0132] In one possible implementation, the steel thermophysical property prediction model training device further includes: an identifier generation module, used to generate component combination identifiers based on the content of each element, wherein the same component combination corresponds to multiple thermophysical property records at different temperatures; wherein the component combination identifier is configured to group database records and sort them according to a set field, so as to organize multiple records under the same component combination into a set sequence sample.

[0133] In one possible implementation, the steel thermophysical property prediction model training device further includes: a preprocessing module for cleaning the thermophysical property data; performing a log1p transformation on the cleaned thermophysical property data; standardizing the transformed thermophysical property data; and grouping the standardized thermophysical property data by composition and sorting it by temperature to construct a composition-temperature sequence sample.

[0134] In one possible implementation, the training module 402 is specifically used to: train a steel thermal property prediction model based on a bidirectional long short-term memory network using the composition-temperature sequence samples. In one possible implementation, the steel thermal property prediction model training device further includes: an export module, used to export the steel thermal property prediction model as an ONNX model after the steel thermal property prediction model training is completed, and save the input standardized parameters and the output standardized parameters; wherein, the ONNX model is configured for front-end loading.

[0135] Based on the same application concept, this application also provides a steel thermal property prediction device corresponding to the steel thermal property prediction method. Since the principle of the device in this application is similar to that of the aforementioned steel thermal property prediction method, the implementation of the device in this application can refer to the description in the above-mentioned method embodiments, and the repeated parts will not be described again.

[0136] Please see Figure 5 This is a functional module diagram of the steel thermal property prediction device provided in this application embodiment. Each module in the steel thermal property prediction device in this embodiment is used to execute the steps in the above method embodiments. The steel thermal property prediction device includes a second acquisition module 501, an input module 502, and an output module 503; wherein, The second acquisition module 501 is used to acquire steel information input by the user.

[0137] The input module 502 is used to input the steel information into the steel thermal property prediction model in the above embodiment.

[0138] The output module 503 is used to output steel thermal property prediction data through the steel thermal property prediction model.

[0139] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the steel thermal property prediction model training method and / or steel thermal property prediction method described in the above method embodiments.

[0140] The computer program product of the steel thermal property prediction model training method and / or steel thermal property prediction method provided in the embodiments of this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the steps of the steel thermal property prediction model training method and / or steel thermal property prediction method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0141] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0142] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0143] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, 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 a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0144] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for training a steel thermophysical property prediction model, characterized in that, include: Acquire thermal property data calculated in batches by materials calculation software; wherein, the thermal property data includes steel composition, temperature, and thermal properties; Using the aforementioned thermal property data, a steel thermal property prediction model based on a bidirectional long short-term memory network is trained; wherein, the steel thermal property prediction model is configured to perform rapid inference on the received input data to obtain thermal property prediction results.

2. The method according to claim 1, characterized in that, in, The steel thermophysical property prediction model is a residual BiLSTM model. The BiLSTM model includes: a bidirectional LSTM layer, a LayerNorm layer, a Dropout layer, a residual connection layer, and an output fully connected layer connected in sequence.

3. The method according to claim 1, characterized in that, The material calculation software is JMatPro Dynamic; The thermophysical data is configured to be obtained by batch calculation of thermophysical parameters of steels with different chemical compositions at different temperatures using JMatPro Dynamics.

4. The method according to claim 1, characterized in that, Before training a steel thermal property prediction model based on a bidirectional long short-term memory network using the aforementioned thermal property data, the method further includes: The thermal property data is divided into a training set, a test set, and a validation set; The training set is used for model learning, the validation set is used for model selection and early stopping control during the training process, and the test set is used for prediction accuracy evaluation.

5. The method according to claim 1, characterized in that, The step of training a steel thermal property prediction model based on a bidirectional long short-term memory network using the aforementioned thermal property data includes: After each round of training of the steel thermophysical property prediction model, the validation set loss is calculated; Save the steel thermophysical property prediction model with the lowest validation set loss as the best model; If the loss on the validation set does not decrease for a predetermined number of consecutive rounds, an early stopping strategy is executed to stop training the steel thermal property prediction model.

6. The method according to claim 1, characterized in that, Before acquiring the thermophysical property data calculated in batches by the material calculation software, the method further includes: The material calculation software is used to perform batch calculations on different steel compositions and temperature paths to generate corresponding thermal property data. The thermophysical data is stored in a database in the form of structured data; each data record includes at least a composition field, a temperature field, and a performance field.

7. The method according to claim 1, characterized in that, Before acquiring the thermal property data calculated in batches by the material calculation software, the method further includes: Based on the content of each element, a component combination identifier is generated, and the same component combination corresponds to multiple thermophysical property records at different temperatures; The component combination identifier is configured to group database records and sort them according to a set field, so as to organize multiple records under the same component combination into a set sequence sample.

8. The method according to claim 1, characterized in that, Before training a steel thermal property prediction model based on a bidirectional long short-term memory network using the aforementioned thermal property data, the method further includes: Clean the aforementioned thermal property data; Perform log1p transformation on the thermal property data after cleaning; The transformed thermal property data are standardized. The standardized thermophysical property data are grouped by composition and sorted by temperature to construct composition-temperature sequence samples; The step of training a steel thermal property prediction model based on a bidirectional long short-term memory network using the aforementioned thermal property data includes: Using the composition-temperature sequence samples, a steel thermophysical property prediction model based on a bidirectional long short-term memory network is trained.

9. The method according to claim 1, characterized in that, The method further includes: Once the steel thermal property prediction model has been trained, export the steel thermal property prediction model as an ONNX model and save the input standardized parameters and output standardized parameters. The ONNX model is configured for loading by the front end.

10. A method for predicting the thermal properties of steel, characterized in that, Obtain the steel information input by the user; Input the steel information into the steel thermophysical property prediction model according to any one of claims 1-9; The steel thermal property prediction model outputs predicted steel thermal property data.

11. A training device for predicting the thermal properties of steel, characterized in that, include: The first acquisition module is used to acquire thermal property data calculated in batches by material calculation software; wherein, the thermal property data includes steel composition, temperature, and thermal properties; The training module is used to train a steel thermal property prediction model based on a bidirectional long short-term memory network using the thermal property data; wherein the steel thermal property prediction model is configured to perform fast inference on the received input data to obtain thermal property prediction results.

12. A device for predicting the thermal properties of steel, characterized in that, The second acquisition module is used to acquire the steel information input by the user; The input module is used to input the steel information into the steel thermophysical property prediction model according to any one of claims 1-9; The output module is used to output steel thermal property prediction data through the steel thermal property prediction model.

Images