Die steel heat treatment deformation digital twin prediction method, system, device and medium

Abstract

The application discloses a die steel heat treatment deformation digital twinning prediction method, system, device and medium, relates to the metal material heat treatment technical field, and includes the following steps: according to the stress-strain field, a deformation characteristic field evolution model is constructed, non-uniform deformation characteristics of die steel in the heat treatment process are extracted and quantified, and deformation characteristic field data are generated;Acquire the actual geometric data of die steel at the key time node in the heat treatment process, construct a geometric deviation field based on the actual geometric data, calibrate the deformation characteristic field data and the geometric deviation field, and obtain the calibrated deformation field distribution data.The application constructs a geometric deviation field by acquiring the actual geometric data of die steel at the key time node, calibrates the deformation characteristic field data and the geometric deviation field, reversely corrects the model parameters of the multi-physical field digital twinning body by using a data assimilation algorithm, and realizes dynamic matching of the simulation model and the actual heat treatment process.

Description

Technical Field

[0001] This invention relates to the field of heat treatment technology for metallic materials, and in particular to a method, system, equipment, and medium for predicting digital twin deformation of mold steel during heat treatment. Background Technology

[0002] With the rapid development of high-end equipment manufacturing and precision mold forming, increasingly stringent requirements have been placed on the dimensional accuracy and deformation control of mold steel after heat treatment. The heat treatment process for mold steel involves multiple stages, including heating, holding, and cooling, accompanied by complex multi-physics field coupling effects of temperature field, phase transformation, and stress-strain, which easily leads to non-uniform and irreversible dimensional deformation. If such deformation cannot be accurately predicted and controlled, it will directly result in repeated mold trials, repairs, or even scrapping of the mold, significantly extending the manufacturing cycle and increasing production costs. This has become one of the key bottlenecks restricting the quality and efficiency of precision mold manufacturing.

[0003] Currently, the prediction of deformation during heat treatment of mold steel mainly relies on finite element simulation technology. This involves establishing a multi-field coupled thermo-mechanical-phase transformation model and performing numerical simulations by combining material property parameters and process boundary conditions. Existing techniques have involved researchers experimentally determining the material's thermophysical parameters, phase transformation kinetic parameters, and high-temperature rheological stress parameters, and constructing deformation prediction models based on heat transfer and elastoplastic mechanics theories. Some studies have further introduced online monitoring methods to collect temperature and strain data at key points inside the mold steel, which are then used to correct local parameters in the simulation model. Furthermore, with the rise of the digital twin concept, some studies have attempted to combine physical models with real-time data to construct a digital mirror of the heat treatment process, aiming to improve the dynamic response capability of deformation prediction.

[0004] However, the aforementioned existing technologies still have significant shortcomings. First, the heat transfer boundary conditions on the surface of mold steel during heat treatment exhibit strong time-varying and spatial non-uniformity. Traditional methods typically employ constant or empirically modified heat transfer coefficients, making it difficult to accurately characterize the dynamic evolution of heat transfer capacity in different surface regions during the quenching and cooling stage, thus limiting the simulation accuracy of temperature and stress-strain fields. Second, existing simulation models are mostly open-loop structures, lacking a closed-loop mechanism for systematically correcting key model parameters based on measured deformation data. The matching degree between material constitutive parameters, phase transformation kinetic parameters, and actual working conditions cannot automatically recover after process fluctuations, leading to a continuous accumulation of prediction bias. Furthermore, the integration degree between the deformation characteristic field and the geometric deviation field is low, and there is a lack of a unified spatiotemporal calibration framework between the strain results output by simulation and the actual geometric data obtained from 3D scanning, making it difficult to fully leverage the complementary advantages of multi-source heterogeneous data.

[0005] Therefore, it is essential to invent a digital twin prediction method, system, equipment, and medium for heat treatment deformation of mold steel to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a digital twin prediction method, system, equipment and medium for heat treatment deformation of mold steel, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a digital twin prediction method for heat treatment deformation of mold steel, specifically including the following steps: S1: Obtain multi-source heterogeneous data of the target mold steel during the heat treatment process. The multi-source heterogeneous data includes real-time process parameters, real-time response data, initial material property data, and historical operating condition data of the heat treatment equipment.

[0008] S2: Based on the real-time process parameters, real-time response data and historical operating condition data in the multi-source heterogeneous data, construct a dynamic boundary condition evolution model to generate a dynamic boundary condition field that is related to time and space.

[0009] S3: Based on the initial material property data, construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model. Using the dynamic boundary condition field as the load boundary input, perform multiphysics coupling solution on the heat treatment process and output the stress-strain field.

[0010] S4: Based on the stress-strain field, construct a deformation characteristic field evolution model, extract and quantify the non-uniform deformation characteristics of the mold steel during the heat treatment process, and generate deformation characteristic field data;

[0011] S5: Obtain the actual geometric data of the mold steel at key time points during the heat treatment process, construct a geometric deviation field based on the actual geometric data, and calibrate the deformation characteristic field data with the geometric deviation field to obtain the calibrated deformation field distribution data;

[0012] S6: Based on the calibrated deformation field distribution data and the deformation feature field data of the current iteration step, calculate the deformation field residual, and reverse-correct the boundary condition parameters of the dynamic boundary condition evolution model based on the deformation field residual. Input the corrected boundary condition parameters into the dynamic boundary condition evolution model to reconstruct the model state, and update the dynamic boundary condition field using the reconstructed dynamic boundary condition evolution model. Specifically, in the first iteration, the deformation feature field data of the current iteration step is the deformation feature field data generated in step S4; in subsequent iterations, the deformation feature field data of the current iteration step is the deformation feature field data regenerated in step S4 based on the updated dynamic boundary condition field.

[0013] S7: Based on the updated dynamic boundary condition field, iteratively execute steps S3 to S6 until the preset convergence condition is met, and output the final deformation prediction result.

[0014] A digital twin prediction system for heat treatment deformation of mold steel includes the following modules: The data acquisition module is used to acquire multi-source heterogeneous data of the target mold steel during the heat treatment process; The boundary condition construction module is used to construct a dynamic boundary condition evolution model based on the multi-source heterogeneous data and generate a dynamic boundary condition field. The twin solver module is used to construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model based on the initial material property data. The dynamic boundary condition field is used as the load boundary input, and the stress-strain field is output. A feature field construction module is used to generate deformation feature field data based on the stress-strain field. The calibration module is used to acquire the actual geometric data of the mold steel at key time points during the heat treatment process, construct a geometric deviation field, and calibrate the deformation characteristic field data with the geometric deviation field to obtain the calibrated deformation field distribution data. The parameter correction module is used to calculate the deformation field residual based on the calibrated deformation field distribution data and the deformation feature field data of the current iteration step, and to reverse correct the boundary condition parameters of the dynamic boundary condition evolution model based on the deformation field residual. The corrected boundary condition parameters are then input into the boundary condition construction module to reconstruct the model state. The iterative control module is used to control each module to perform iterative calculations until the preset convergence conditions are met, and output the final deformation prediction result.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the digital twin prediction method for heat treatment deformation of mold steel as described in any of the preceding claims.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the digital twin prediction method for heat treatment deformation of mold steel as described in any of the preceding claims.

[0017] The technical effects and advantages of this invention are as follows: 1. This invention constructs a dynamic boundary condition evolution model, generates a dynamic boundary condition field that is related to time and space based on real-time process parameters, real-time response data and historical working condition data, and realizes high-precision characterization of the spatiotemporal distribution of the heat transfer coefficient on the surface of mold steel during heat treatment, effectively improving the simulation accuracy of temperature field and stress-strain field. 2. This invention constructs a multi-physics digital twin that includes a phase transformation dynamics model and a thermo-mechanical coupling model, and uses the dynamic boundary condition field as the load boundary input to perform multi-physics coupling solution. This achieves bidirectional strong coupling calculation of the temperature field, microstructure field and stress-strain field in the heat treatment process, and enhances the physical consistency of deformation prediction. 3. This invention constructs a deformation feature field evolution model, performs feature decomposition and spatial differentiation operations on the stress-strain field, and extracts non-uniform deformation features in the plastic strain tensor field and phase transformation strain tensor field, thereby achieving accurate identification and quantitative characterization of deformation concentration areas. 4. This invention constructs a geometric deviation field by acquiring the actual geometric data of the mold steel at key time nodes, calibrates the deformation characteristic field data with the geometric deviation field, and uses a data assimilation algorithm to reverse correct the model parameters of the multiphysics digital twin, thereby realizing the dynamic matching between the simulation model and the actual heat treatment process. 5. This invention calculates the residual of the deformation field and corrects the boundary condition parameters of the dynamic boundary condition evolution model in reverse based on the adjoint sensitivity analysis method. After reconstructing the model state, the dynamic boundary condition field is updated, forming a closed-loop parameter optimization mechanism driven by measured data, which improves the adaptability of boundary condition prediction. 6. This invention achieves successive approximation and stable output of deformation prediction results by iteratively performing multiphysics field solution, deformation feature extraction, geometric deviation calibration and boundary condition correction, and judging the convergence condition based on the difference between deformation feature fields in adjacent iteration steps and the difference between deformation feature fields and geometric deviation fields, which significantly improves prediction accuracy and engineering applicability. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is the process for constructing the dynamic boundary condition evolution model of the present invention; Figure 3 This is a flowchart illustrating the convergence determination process of the present invention. Figure 4 This is a system module architecture diagram of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides, for example Figure 1The digital twin prediction method for heat treatment deformation of mold steel shown includes the following steps: S1: Obtain multi-source heterogeneous data of the target mold steel during the heat treatment process. The multi-source heterogeneous data includes real-time process parameters, real-time response data, initial material property data, and historical operating condition data of the heat treatment equipment. Furthermore, in the above technical solution, the real-time process parameters include at least heating rate, holding time, cooling medium flow rate, and quenching medium temperature; the real-time response data includes at least temperature data and strain data of key points inside the mold steel; and the initial material property data includes at least thermophysical parameters, phase transformation kinetic parameters, and high-temperature rheological stress parameters.

[0021] In step S1, the multi-source heterogeneous data of the target mold steel during the heat treatment process are obtained in the following ways: Method 1: Obtaining real-time process parameters: The real-time process parameters include at least the heating rate, holding time, cooling medium flow rate, and quenching medium temperature. In this embodiment, the real-time process parameters are acquired in real time by a process parameter acquisition unit set in the vacuum heat treatment furnace. The sampling period of the process parameter acquisition unit is synchronized with that of the industrial control host computer, and the sampling frequency is set to 1Hz. All acquired real-time process parameters are accompanied by corresponding timestamps and stored in the local database of the host computer in time sequence.

[0022] Specifically, the heating rate is calculated by real-time acquisition of temperature change data during the heating phase of the vacuum heat treatment furnace through the temperature control module, resulting in the temperature rise per unit time. In this embodiment, the heating rate is acquired within a range of 5℃ / min to 20℃ / min, with an acquisition accuracy of 0.5℃ / min. The heat preservation time is calculated by acquiring the starting time when the furnace temperature reaches the preset heat preservation temperature and the ending time when the temperature remains within the preset heat preservation temperature range through the temperature control module. The time difference between the two time points is the heat preservation time. In this embodiment, the acquisition accuracy of the heat preservation time is 1 second. The flow rate of the cooling medium is collected in real time by an electromagnetic flowmeter installed at the inlet of the cooling pipe of the vacuum heat treatment furnace. The cooling medium is high-purity nitrogen, and the electromagnetic flowmeter has a sampling range of 0~500m. 3 / h, with a collection accuracy of 1%FS, and the real-time collected traffic data is synchronously uploaded to the host computer; The temperature of the quenching medium is collected in real time by a platinum resistance temperature sensor installed in the quenching oil bath. The measurement range of the platinum resistance temperature sensor is -20℃ to 200℃, and the acquisition accuracy is 0.2℃. The collected quenching medium temperature data is uploaded synchronously with the furnace temperature data to ensure consistency in the time dimension.

[0023] Method 2: Acquiring real-time response data: The real-time response data includes at least temperature and strain data of key points inside the mold steel. In this embodiment, the real-time response data is acquired in real time by a built-in sensing unit embedded inside the target mold steel. The built-in sensing unit includes a K-type armored thermocouple and a high-temperature strain gauge. The sampling frequency is consistent with the sampling frequency of the process parameter acquisition unit, which is 1Hz, to ensure time synchronization with the real-time process parameters.

[0024] Specifically, the key points inside the mold steel are selected in advance based on the three-dimensional design geometric model of the target mold steel. Locations that are prone to heat treatment deformation, such as the corners of the mold steel cavity, the places where the wall thickness changes abruptly, and the places where the maximum stress concentration occurs, are selected as key points. In this embodiment, a total of 8 key points are set, and a set of thermocouples and high-temperature strain gauges are arranged at each key point. Temperature data of key points inside the mold steel are collected in real time by K-type armored thermocouples. The thermocouples have a measurement range of 0℃ to 1200℃ and an acquisition accuracy of 1℃. The collected temperature data is transmitted to the host computer through high-temperature shielded cables, and the real-time temperature change sequence of each key point with heat treatment time is recorded accordingly. The strain data of key points inside the mold steel is collected in real time by deploying high-temperature resistant strain gauges. The strain gauges can withstand temperatures of not less than 1000℃, the strain measurement range is 5000, and the acquisition accuracy is 1. The collected strain data is synchronously transmitted to the host computer, and the real-time strain change data of each key point during the heat treatment process is recorded accordingly, serving as real-time response feedback data for the heat treatment process of the mold steel.

[0025] Method 3: Obtaining Initial Material Property Data The initial material property data includes at least thermophysical parameters, phase transformation kinetic parameters, and high-temperature rheological stress parameters. In this embodiment, the initial material property data is obtained by testing the same batch of samples of the target mold steel using a material performance testing unit based on the test methods specified in national standards. All test data are pre-stored in the material property database of the host computer and are directly called when this prediction method is executed.

[0026] Specifically, the thermal properties parameters, including the thermal conductivity, specific heat capacity, density, and coefficient of thermal expansion of the target mold steel in the temperature range of 20℃ to 1200℃, are obtained by testing with a laser thermal conductivity meter and a high-temperature dilatometer according to the test methods specified in GB / T22588-2008 and GB / T4339-2008. The thermal properties parameters corresponding to each temperature node are formed into a corresponding data table, which serves as the basic input parameters for the subsequent multiphysics digital twin. Said phase transformation kinetic parameters, including austenitizing kinetic parameters, martensitic transformation kinetic parameters, pearlite / bainite transformation kinetic parameters of the target die steel, are obtained by detection through the expansion method combined with the metallographic method according to the test method specified in GB / T50563-2010, and the correlation coefficients of the phase transformation kinetic equation are fitted based on the detection data, which is used to construct the phase transformation kinetic model in the multi-physics digital twin; Said high-temperature rheological stress parameters, including rheological stress curves of the target die steel at different temperatures and different strain rates, are obtained by detection through high-temperature tensile tests conducted on a Gleeble thermal simulation testing machine according to the test method specified in GB / T30069.2-2016, and the material constitutive equation is constructed based on the detection data, which is used for the solution calculation of the thermo-mechanical coupling model in the multi-physics digital twin.

[0027] Mode 4: Acquisition of historical working condition data of heat treatment equipment: Said historical working condition data of the heat treatment equipment is obtained by retrieving from the equipment working condition database matched with the vacuum heat treatment furnace. Said historical working condition data is the historical operation data of the vacuum heat treatment furnace implementing the heat treatment process of the same type of die steel with the same specification within the past 12 months, including furnace temperature change data, cooling medium flow historical data, equipment heating power historical data, quenching medium temperature historical data, and equipment fault record data during the historical heat treatment process. In this embodiment, the retrieved historical working condition data is classified and stored according to heat treatment batches, and each batch of data includes complete working condition parameters of the whole heat treatment process, which is used for the construction and training of the subsequent dynamic boundary condition evolution model.

[0028] In step S1 of this embodiment, all acquired multi-source heterogeneous data are subjected to standardized preprocessing by the upper computer, the unified data format is structured data with time stamp and space coordinate identification, and the data time difference of different acquisition units is eliminated to ensure the matching and availability of data in subsequent steps.

[0029] S2: Based on the real-time process parameters, real-time response data and historical working condition data in said multi-source heterogeneous data, a dynamic boundary condition evolution model is constructed, and a dynamic boundary condition field related to time and space is generated; Further, in the above technical solution, as Figure 2 shown, the specific steps of constructing the dynamic boundary condition evolution model in said S2 include: S21: Acquire historical working condition data and real-time working condition data of the heat treatment equipment, and extract the distribution characteristics of the heat transfer coefficient in spatial dimension and temporal dimension; This step is the feature construction stage of the dynamic boundary condition evolution model. The real-time operating condition data is the real-time process parameters and real-time response data collected in step S1. The historical operating condition data is the historical operating data of the vacuum heat treatment furnace for the same type of mold steel and the same specification heat treatment process in the past 12 months retrieved in step S1.

[0030] The specific implementation process is as follows: Inverse calculation of surface heat transfer coefficient: Based on the real-time temperature data of key points inside the mold steel obtained in step S1, the surface heat transfer coefficient of each surface region of the mold steel is calculated using a one-dimensional unsteady-state inverse heat transfer method. Specifically, using the measured temperature of the key points as the observed value and the Fourier heat conduction equation as the governing equation, the heat transfer coefficient value of each surface region of the mold steel at the corresponding time step is obtained through iterative solution using the conjugate gradient method, solving the problem that the surface heat transfer coefficient cannot be directly measured with high precision during heat treatment. In this embodiment, based on the three-dimensional design geometric model of the target mold steel in step S1, the surface of the mold steel is divided into 200 independent surface elements, each surface element corresponding to a calculated heat transfer coefficient value, ensuring the spatial resolution of the heat transfer coefficient matches the mesh of the subsequent finite element solution.

[0031] Distribution Feature Extraction: The entire heat treatment process is divided into five stages: heating, holding, quenching and cooling, and tempering. For each stage, the temporal and spatial distribution features of the heat transfer coefficient are extracted. The time-dimensional distribution characteristics, including the mean, peak, rate of change, and time series autocorrelation coefficient of the heat transfer coefficient at each stage, are used to characterize the evolution of the heat transfer coefficient with heat treatment time. The spatial dimension distribution characteristics include the distribution gradient, variance, and inter-regional difference coefficient of the heat transfer coefficient of each surface unit of the mold steel, which are used to characterize the distribution differences of the heat transfer coefficient at different surface locations such as the cavity surface, non-working surface, and abrupt changes in wall thickness of the mold steel.

[0032] All distribution features extracted in this step are stored in time series-standardized form as the basis for input features in subsequent time series prediction models.

[0033] S22: Based on the distribution characteristics, a time-series prediction model is adopted, using the real-time process parameters and the historical operating data as drivers, to predict the heat transfer coefficient field at the next moment. This step is the core prediction step of the dynamic boundary condition evolution model. In this embodiment, the time-series prediction model adopts the LSTM long short-term memory neural network model. The model is pre-trained based on historical working condition data and then performs real-time online prediction during the heat treatment process.

[0034] The specific implementation process is as follows: Model pre-training: 200 sets of historical operating data of the same specification H13 mold steel and the same heat treatment process obtained in step S1 are selected as the dataset and divided into training set and validation set in an 8:2 ratio. The input feature vector of the model includes the heat transfer coefficient values ​​of the historical 30 time steps, the real-time process parameters (heating rate, cooling medium flow rate, quenching medium temperature, furnace temperature) of the corresponding time steps, the temperature data of key points of the mold steel, and the heat transfer coefficient distribution features extracted in step S21. The output of the model is the heat transfer coefficient value corresponding to the 200 surface units of the next time step, that is, the heat transfer coefficient field covering the entire surface of the mold steel. The optimizer used for model training is the Adam optimizer, with a learning rate of 0.001 and 500 iterations. The mean square error (MSE) is used as the loss function. When the loss value of the validation set converges to below 0.005, the model pre-training is completed, and the initial dynamic boundary condition evolution model is obtained.

[0035] Real-time online prediction: During the real-time operation phase of the heat treatment process, the real-time process parameters, real-time response data, and calculated heat transfer coefficient values ​​collected in step S1 (current time step and the previous 29 time steps), along with the real-time distribution features extracted in step S21, are input into the pre-trained LSTM model. The model outputs the heat transfer coefficient values ​​corresponding to each surface unit in the next time step, forming a complete heat transfer coefficient field. Simultaneously, after every 100 time steps of prediction, the model is fine-tuned online using real-time measured heat transfer coefficient data, updating the model weights to further improve prediction accuracy and achieve real-time adaptive optimization of the dynamic boundary condition evolution model.

[0036] S23: Obtain the geometric contour data of the mold steel during the heat treatment process, and spatially map the geometric contour data with the heat transfer coefficient field to generate the dynamic boundary condition field corresponding to the grid nodes on the surface of the mold steel.

[0037] This step is the final generation stage of the dynamic boundary condition field. The generated dynamic boundary condition field is directly used as the load boundary input of the multiphysics digital twin in the subsequent step S3.

[0038] The specific implementation process is as follows: Geometric contour data and finite element mesh acquisition: The geometric contour data of the mold steel during the heat treatment process is the three-dimensional design geometric model (STL format) of the target mold steel in step S1. The three-dimensional design geometric model is divided into tetrahedral meshes using the finite element preprocessing software ABAQUS to obtain the finite element mesh model of the mold steel. The total number of surface mesh nodes is 12,560, and each surface mesh node has a unique spatial three-dimensional coordinate identifier, which forms a spatial correspondence with the 200 surface elements divided in step S21.

[0039] Spatial mapping processing: The heat transfer coefficient field (heat transfer coefficient values ​​of 200 surface elements) output in step S22 is spatially interpolated and mapped to calculate the heat transfer coefficient value corresponding to each surface mesh node in the finite element model, thereby realizing the spatial matching between the heat transfer coefficient field and the geometric contour of the mold steel.

[0040] Dynamic boundary condition field generation: The heat transfer coefficient values ​​of each surface grid node corresponding to each time step are integrated according to the time sequence to generate a dynamic boundary condition field that is related to time and space. The dynamic boundary condition field is a convective heat transfer boundary condition (third type of thermal boundary condition) that is updated in real time with the heat treatment time, which fully matches the boundary input format requirements of the multiphysics coupling solution in the subsequent step S3.

[0041] In an optional embodiment of the present invention, if real-time geometric contour data of the mold steel is obtained by a three-dimensional scanning device during the heat treatment process (which is from the same source as the data collected in the subsequent step S5), the geometric contour data and the corresponding finite element mesh are updated, spatial mapping is re-executed, the dynamic boundary condition field is updated, and the boundary condition changes caused by geometric deformation during the heat treatment of the mold steel are further matched to improve the accuracy of deformation prediction.

[0042] S3: Based on the initial material property data, construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model. Using the dynamic boundary condition field as the load boundary input, perform multiphysics coupling solution on the heat treatment process and output the stress-strain field. In step S3, the process of constructing a multiphysics digital twin and completing the multiphysics coupling solution is implemented through the following steps: Section 1: The Basic Framework for Building Multiphysics Digital Twins In this embodiment, the multiphysics digital twin is constructed based on the finite element analysis platform ABAQUS. Custom solutions for phase transition dynamics and thermo-mechanical coupling are achieved using the user-defined subroutines UMAT and FILM. The geometric model and finite element mesh of the digital twin are completely consistent with the three-dimensional design geometric model of the target mold steel used in step S2. Specifically, the finite element mesh model is divided using tetrahedral elements, with a total of 286,500 mesh elements, of which 12,560 are surface mesh nodes. These correspond one-to-one with the spatial mapping nodes of the dynamic boundary condition field in step S2, ensuring accurate loading of the load boundaries.

[0043] The multiphysics digital twin contains two core sub-models: a phase transition dynamics model and a thermo-mechanical coupling model. The two sub-models achieve bidirectional data transfer through the temperature field and the phase volume fraction field, forming a closed-loop coupled digital twin solution system.

[0044] Step 2: Construction of the Phase Transition Dynamics Model This model is constructed based on the phase transformation kinetic parameters obtained in step S1. It is used to solve the austenitization and cooling phase transformation process inside the mold steel during heat treatment. It outputs the phase composition (austenite, martensite, pearlite, bainite) and phase volume fraction of each mesh element of the mold steel at each time step, providing basic data for phase transformation strain and material property updates for the thermo-mechanical coupling model.

[0045] The specific implementation process is as follows: Austenitizing kinetic sub-model: The JMAK (Johnson-Mehl-Avrami-Kolmogorov) equations are used as the governing equations for the austenitizing process, and the equations are as follows: f A =1-exp(-k(t-t0) n ), Among them, f A t represents the volume fraction of austenite phase, t is the holding time, t0 is the austenitization initiation time, and k and n are the rate constant and Avrami exponent obtained by fitting the austenitization kinetic parameters detected in step S1. k is related to the material's austenitization activation energy and heating temperature. The activation energy value is directly adopted from the austenitization activation energy of H13 mold steel obtained by the expansion method in step S1. In this model, the austenitization initiation temperature Ac1 and termination temperature Ac3 are both measured values ​​obtained in step S1. When the temperature of the mold steel element exceeds Ac1, the austenitization solution process is initiated.

[0046] Cooling phase transformation kinetics sub-models: solution models are constructed separately for martensitic phase transformation and pearlite / bainite phase transformation. The martensitic phase transformation uses the Koistinen-Marburger (KM) equation as the governing equation, and the equation takes the following form: f M =1-exp(-(M s -T)), Among them, f M M is the volume fraction of martensite phase. s The temperature at which the martensitic phase transformation begins is T, where T is the current temperature of the element, and M is the phase transformation rate constant obtained by fitting the martensitic phase transformation kinetic parameters detected in step S1. s The numerical values ​​are directly taken from the measured results of step S1.

[0047] The pearlite / bainite phase transformation also uses the JMAK equation to construct the solution model. The correlation coefficients of the equation are all obtained by fitting the pearlite / bainite phase transformation kinetic parameters detected in step S1. When the unit cooling rate is lower than the critical cooling rate and the temperature is in the corresponding phase transformation temperature range, the pearlite / bainite phase transformation solution is initiated.

[0048] In each solution time step, this model obtains the phase volume fraction of each element based on the current temperature field, and synchronously updates the thermophysical and mechanical property parameters of the corresponding elements, realizing real-time coupling between the phase transition process and the thermo-mechanical field.

[0049] Step 3: Construction of the thermo-mechanical coupling model: This model is constructed based on the thermophysical parameters and high-temperature rheological stress parameters obtained in step S1. It includes a heat conduction sub-model and an elastoplastic constitutive sub-model. The dynamic boundary condition field generated in step S2 is used as the thermal boundary load input to solve for the temperature field and stress-strain field of the mold steel during heat treatment. The specific implementation process is as follows: The heat conduction model uses Fourier's law of heat conduction as the governing equation, and the equation takes the following form: , Wherein, is the material density, c is the specific heat capacity, and is the thermal conductivity. These parameters are all thermophysical properties obtained in step S1, varying with temperature and phase composition. Q is the latent heat of phase change, calculated based on the change in phase volume fraction obtained from the phase change kinetic model. The thermal boundary conditions of this model directly adopt the dynamic boundary condition field generated in step S2, i.e., the convective heat transfer boundary conditions (third type of thermal boundary conditions) that are updated in real time and space. Through the ABAQUS FILM user subroutine, the heat transfer coefficient values ​​corresponding to each time step and each surface grid node are accurately loaded to the corresponding positions, achieving dynamic updating of the thermal boundary.

[0050] Elastoplastic constitutive model: A high-temperature elastoplastic constitutive model considering phase transition effects is adopted, and the total strain consists of four parts: total = e + p + th + tr , in, e For elastic strain, p For plastic strain, th For thermal strain, tr The phase transformation strain is calculated based on the generalized Hooke's law, with the elastic modulus and Poisson's ratio obtained from step S1 as parameters change with temperature and phase composition. The plastic strain is calculated based on the constitutive equation constructed from the high-temperature rheological stress parameters obtained from step S1, which considers the effects of temperature, strain rate, and phase composition on the rheological stress. The thermal strain is calculated based on the material's thermal expansion coefficient and temperature change. The phase transformation strain is calculated based on the change in phase volume fraction and the specific volume difference of each phase obtained from the phase transformation kinetic model.

[0051] In each solution time step, this model solves the mechanical field based on the temperature field solved by the thermal conductivity model and the phase volume fraction solved by the phase transformation dynamics model, and outputs the stress tensor and strain tensor of each element.

[0052] Step 4: Multiphysics Coupling Solution and Stress-Strain Field Output In this embodiment, the multiphysics coupling solution adopts a strong bidirectional coupling method of thermal-phase change-mechanical. The solution process for each time step is as follows: Load the dynamic boundary condition field generated in step S2 for the current time step, and solve the temperature field of the full model of the mold steel for the current time step through the heat conduction submodel. The temperature field is input into the phase transition dynamics model to solve for the phase volume fraction and phase transition variables of each element at the current time step, and the thermophysical parameters and mechanical performance parameters of each element are updated synchronously. By inputting the temperature field, phase volume fraction, and material parameters into the mechanical constitutive submodel of the thermo-mechanical coupling model, the stress field and strain field of the full model at the current time step can be obtained by solving. The mesh node coordinates of the mold steel are updated based on the strain field obtained by the solution, so as to realize the real-time update of geometric deformation. The updated geometric model is fed back to the thermal boundary loading stage, providing an updated geometric basis for the dynamic boundary condition field space mapping of the next time step.

[0053] In this embodiment, the solution at each time step is set with an iterative convergence criterion: the residual of the temperature field solution is less than 110. -6 The displacement residuals obtained from the mechanical field solution are less than 110. -3 When the iteration result satisfies the convergence criterion, the solution for the current time step is completed, and the calculation for the next time step begins.

[0054] After the solution is completed, the stress and strain field covering all mesh nodes of the entire mold steel model and all time steps of the entire heat treatment process is output. The stress and strain field includes the full strain tensor, elastic strain tensor, plastic strain tensor, phase transformation strain tensor, and stress tensor data of each node, providing complete basic data for the construction of the deformation characteristic field in the subsequent step S4.

[0055] In an optional embodiment of the present invention, the multiphysics digital twin can be constructed based on the ANSYS finite element platform, and the phase transformation dynamics and thermo-mechanical coupling model can be solved through a UPF user-defined program; alternatively, a grain growth evolution model can be added to the multiphysics digital twin to further consider the influence of grain size on material properties and deformation behavior, thereby improving the solution accuracy.

[0056] S4: Based on the stress-strain field, construct a deformation characteristic field evolution model, extract and quantify the non-uniform deformation characteristics of the mold steel during the heat treatment process, and generate deformation characteristic field data; Furthermore, in the above technical solution, the specific steps for constructing the deformation feature field evolution model in S4 include: S41: Extract the plastic strain tensor field and the phase transformation strain tensor field from the stress-strain field; This step is the basic data screening stage for deformation feature extraction. Its core is to extract the two core strain components that cause irreversible permanent deformation of the die steel from the full strain field, eliminating the interference of recoverable deformations such as elastic strain on deformation prediction. The specific implementation process is as follows: Node Mapping and Preprocessing of Tensor Data: The stress-strain field output in step S3 is the tensor calculation result of the element integration points in the finite element model. This step first uses the shape function linear interpolation method to interpolate and map the strain tensor data of the element integration points to all 12560 surface mesh nodes and internal mesh nodes of the model, establishing a one-to-one mapping relationship between each node and the corresponding strain tensor, forming a node-level strain tensor field covering the entire model. Simultaneously, outlier removal is performed on the interpolated tensor data: three criteria are used to identify strain singularities caused by mesh distortion and non-convergence in the solution. Singularities are corrected using the tensor average of five adjacent nodes to ensure the continuity and validity of the tensor data.

[0057] Extraction of the target strain tensor field: Based on the total strain composition formula of the thermo-mechanical coupling model in step S3, two types of irreversible strain tensor fields are extracted from the nodal-level strain tensor field: The plastic strain tensor field is the second-order symmetric tensor of plastic strain corresponding to each node, and is defined by the constitutive model in step S3. p Completely corresponding, each tensor is a second-order symmetric tensor of tensor 33, containing 6 independent strain components, characterizing the permanent deformation of mold steel caused by plastic flow during heat treatment; The phase transition strain tensor field is the second-order symmetric phase transition strain tensor corresponding to each node, and is defined by the constitutive model in step S3. tr Complete correspondence, characterizing the permanent deformation of mold steel caused by the change in specific volume during solid-state phase transformation.

[0058] Time-series data integration: The plastic strain tensor field and phase transformation strain tensor field corresponding to each heat treatment time step are integrated according to the time sequence to form a full-process tensor field sequence that evolves with the heat treatment process. All tensor data are accompanied by timestamps consistent with the aforementioned steps to ensure traceability in the time dimension.

[0059] S42: Perform eigendecomposition on the plastic strain tensor field and the phase transformation strain tensor field to obtain the principal strain direction and principal strain value, and calculate the deformation gradient field based on the principal strain direction and principal strain value. This step involves the quantitative calculation of deformation characteristics. It extracts the core feature parameters of the deformation through tensor feature decomposition and constructs a deformation gradient field characterizing the degree of spatial deformation of the mold steel. The specific implementation process is as follows: Synthesis of the total irreversible deformation tensor: The plastic strain tensor and the phase transformation strain tensor of each node are superimposed to obtain the total irreversible deformation strain tensor of that node. The calculation formula is: = p + tr The synthesized total irreversible deformation strain tensor is still a second-order symmetric tensor of 33, which fully characterizes the total permanent deformation of the mold steel at this node.

[0060] Tensor eigendecomposition: Perform orthogonal diagonal eigendecomposition on the total irreversible deformation strain tensor at each node, and solve the characteristic equation of the tensor. | -I∣=0, Where ...

[0061] Construction of the deformation gradient field: Based on the principal strain value and principal strain direction of each node, the deformation gradient tensor F of that node is calculated using the extreme decomposition theory of deformation gradient. The calculation formula is as follows: F=RU, Wherein, U is the right elongation tensor, constructed from the principal strain values ​​and principal strain directions, representing the pure deformation of the node; R is the rotation tensor, constructed from the rotation angle of the principal strain directions relative to the initial coordinate system, representing the rigid body rotation of the node. After calculating the deformation gradient tensors of all nodes in the entire model, they are integrated to form a deformation gradient field covering the entire space of the mold steel. This deformation gradient field fully represents the spatial distribution of the deformation degree, deformation direction, and rigid body rotation of the mold steel from the initial design geometry to the current time step.

[0062] S43: Perform spatial differentiation on the deformation gradient field, extract the region where the strain gradient exceeds a preset threshold as the deformation feature region, and generate the deformation feature field data containing the spatial distribution information and intensity information of the deformation feature region.

[0063] This step involves the identification of non-uniform deformation characteristics and the final generation of the deformation feature field. The core of this step is to quantify the degree of non-uniformity of deformation through strain gradient quantification, identify high-risk areas of concentrated deformation, and form standardized deformation feature field data. The specific implementation process is as follows: Calculation of strain gradient: Perform three-dimensional spatial differentiation on the deformation gradient field to calculate the strain gradient tensor at each node. The calculation formula is as follows: =▽F, Wherein, ▽ represents the Hamiltonian differential operator, and the strain gradient tensor is a third-order tensor, characterizing the rate of change of deformation in space. The larger its magnitude, the higher the degree of deformation non-uniformity at that location, and the more likely concentrated deformation and dimensional deviations will occur. In this embodiment, the spatial differential operation is implemented using the finite difference method. Based on the spatial coordinates of the node and the deformation gradient tensor value, the partial derivatives of each node in the three orthogonal directions of X, Y, and Z are calculated, and finally the magnitude of the strain gradient tensor is obtained, which serves as a quantitative indicator of the degree of deformation non-uniformity at that node.

[0064] Extraction of deformation feature regions: A strain gradient threshold is preset. In this embodiment, the preset threshold is determined based on the statistical results of historical heat treatment deformation data of the same batch of H13 mold steel. The 95th percentile of the strain gradient modulus in the historical data is taken as the threshold, specifically 0.02 mm. -1 The continuous node regions in the full model where the strain gradient modulus exceeds a preset threshold are marked as deformation feature regions. These deformation feature regions are the concentrated areas of non-uniform deformation during the heat treatment of the mold steel, and they spatially correspond to the easily deformable key points such as cavity corners and abrupt changes in wall thickness selected in step S1.

[0065] Generation of deformation feature field data: Integrating calculation results to generate standardized deformation feature field data, specifically including the following: The spatial distribution data of strain gradient magnitude, principal strain value, and principal strain direction of all nodes in the entire model are completely consistent with the world coordinate system of the initial design geometric model. Spatial distribution information of each deformation feature region, including boundary range, spatial coordinates, volume / area information, and number of nodes; Deformation intensity information such as the maximum principal strain value, average strain gradient, and strain distribution variance in each deformation characteristic region; All data are accompanied by timestamps for corresponding time steps, forming a sequence of deformation feature fields that evolve with heat treatment time, thus fully recording the entire evolution process of non-uniform deformation.

[0066] S5: Obtain the actual geometric data of the mold steel at key time points during the heat treatment process, construct a geometric deviation field based on the actual geometric data, and calibrate the deformation characteristic field data with the geometric deviation field to obtain the calibrated deformation field distribution data; Furthermore, in the above technical solution, the specific steps in S5 for calibrating the deformation feature field data with the geometric deviation field include: S51: Collect actual geometric data of mold steel at key time points of heat treatment using a three-dimensional scanning device, compare the actual geometric data with the design geometric data, and generate the geometric deviation field; This step is fundamental to the acquisition of measured data and the quantification of geometric deviations in the calibration process. The specific implementation process is as follows: Determination of key time nodes in heat treatment: Based on the heat treatment process stages divided in step S21, four key time nodes are determined, namely: the end node of the heating stage (the time when the furnace temperature reaches the preset holding temperature), the end node of the holding stage (the time when the preset holding time is completed), the end node of the quenching and cooling stage (the time when the mold steel cools to below 100℃), and the end node of the tempering stage (the time when the entire heat treatment process is completed). Each key time node corresponds to the deformation feature field data with the same timestamp in step S4 to ensure the matching of subsequent time alignment.

[0067] Acquisition and Preprocessing of Actual Geometric Data: In this embodiment, an industrial-grade laser 3D scanner is used to acquire the actual geometric data of the mold steel. The equipment scanning accuracy is 0.02 mm, and the scanning point cloud density is not less than 100 points / mm. At each key time point, the mold steel is removed from the vacuum heat treatment furnace and placed on the positioning fixture of the scanning platform to complete the full surface scanning and obtain the 3D point cloud data of the mold steel. The acquired point cloud data is preprocessed: statistical filtering is used to remove outliers and noise, and curvature simplification is used to retain the high-density point cloud of the feature region, finally obtaining standardized actual geometric point cloud data.

[0068] Geometric registration and deviation field generation: The design geometric data is the three-dimensional design geometric model (STL format) of the target mold steel in step S1. The Iterative Closest Point (ICP) algorithm is used to perform rigid body registration between the preprocessed actual geometric point cloud data and the design geometric model. During the registration process, the reference positioning surface of the mold steel is locked to eliminate rigid body displacement and rotation deviation caused by clamping and scanning. After the registration is completed, the normal distance from each point in the actual geometric point cloud to the corresponding surface of the design geometric model is calculated, which is the geometric deviation value of that point. A positive deviation means that the actual size is greater than the design size, and a negative deviation means that the actual size is less than the design size.

[0069] Using the Kriging space interpolation method, the geometric deviation values ​​of the point cloud are interpolated and mapped to all 12,560 surface mesh nodes of the finite element model in step S3, establishing a one-to-one mapping relationship between each surface mesh node and the corresponding geometric deviation value, ultimately forming a geometric deviation field covering the entire surface of the mold steel; the geometric deviation field is accompanied by timestamps of corresponding key time nodes, which are completely consistent with the time reference of the deformation feature field data in step S4.

[0070] S52: Align the deformation feature field data with the geometric deviation field in the time dimension to establish a mapping relationship between the deformation feature field and the geometric deviation field; This step involves matching and correlating simulation data with measured data. The specific implementation process is as follows: Time dimension alignment: Based on the timestamps of key time nodes, extract the deformation feature field data generated in step S4 that has the same timestamp as each key time node, and make a one-to-one correspondence between the deformation feature field data and the geometric deviation field at the same time node in terms of time dimension, so as to eliminate the time difference between simulation calculation and actual measurement and ensure that the two sets of data correspond to the same heat treatment state.

[0071] Spatial dimension matching: Using the surface mesh nodes of the finite element model as the spatial reference, the principal strain values ​​and strain gradient magnitudes of each node in the aligned deformation feature field data are matched one-to-one with the geometric deviation values ​​of the same node in the geometric deviation field, forming multiple sets of matching data pairs of node feature parameters and measured geometric deviations.

[0072] Mapping relationship construction: The Gaussian process regression (GPR) algorithm is adopted, with the principal strain values ​​and strain gradient magnitudes of the nodes in the matching data pairs as input features, and the measured geometric deviation values ​​of the corresponding nodes as output labels, to train and obtain a nonlinear mapping relationship between the deformation feature field and the geometric deviation field. In this embodiment, the kernel function of the Gaussian process regression adopts the squared exponential kernel function, and the hyperparameters of the kernel function are optimized through 5-fold cross-validation. The final mapping relationship can predict the geometric deviation of the corresponding nodes through the deformation feature field parameters, and at the same time quantify the deviation law between the simulation prediction and the measured results.

[0073] S53: Using a data assimilation algorithm, the mapping relationship is used as an observation constraint to reverse the parameters of the thermo-mechanical coupling model and the phase transition dynamics model in the multiphysics digital twin, and the corrected thermo-mechanical coupling model parameters and phase transition dynamics model parameters are updated in the multiphysics digital twin. This step is the core of the reverse correction of model parameters, and the specific implementation process is as follows: Determination of parameters to be corrected and observed values: The model parameters to be corrected are divided into two categories, which correspond completely to the model constructed in step S3: Phase transformation kinetic model parameters include the rate constant k and Avrami exponent n of the austenitizing JMAK equation, the phase transformation rate constant of the martensitic KM equation, and the Avrami exponent of the pearlite / bainite phase transformation. Thermo-mechanical coupling model parameters include the strain hardening coefficient and temperature softening coefficient of the high-temperature rheological stress constitutive equation, the temperature coefficient of elastic modulus of different phase structures, and Poisson's ratio.

[0074] Using the predicted geometric deviations of each node output by the mapping relationship constructed in step S52 as observation values, and the measured values ​​of the geometric deviation field as observation constraints, an objective function for parameter optimization is constructed.

[0075] Implementation of the data assimilation algorithm: In this embodiment, the Ensemble Kalman Filter (EnKF) algorithm is used as the data assimilation algorithm. The specific execution flow is as follows: Generate an initial parameter set: Using the initial values ​​of the parameters to be corrected (the material parameters measured in step S1) as the mean, generate a parameter set containing 100 samples with a coefficient of variation of 5% as the initial prior set; Ensemble forecast: Input each set of parameters in the parameter set into the multiphysics digital twin in step S3, complete the heat processing process of the corresponding key time nodes, and obtain the deformation field forecast result corresponding to each set of parameters. Then, convert it into a geometric deviation forecast set through the mapping relationship in step S52. Kalman gain update: Based on the observation constraints of the geometric bias prediction set and the measured geometric bias, the Kalman gain matrix is ​​calculated, the parameter set is updated, and the corrected parameter posterior set is obtained; Convergence criterion: When the coefficient of variation of the posterior set of parameters is less than 1%, the parameter correction is deemed to have converged, and the mean of the posterior set is taken as the final corrected model parameters; if convergence is not achieved, the set prediction and update steps are repeated until the convergence condition is met, and the maximum number of iterations is set to 50 rounds.

[0076] Model parameter update: Replace the corresponding initial parameters in the multiphysics digital twin in step S3 with the corrected phase transition dynamics model parameters and thermo-mechanical coupling model parameters obtained after convergence, and complete the parameter update of the digital twin to ensure that the updated model can more accurately match the deformation law of the actual heat treatment process.

[0077] S54: Based on the corrected thermo-mechanical coupling model parameters and phase transition dynamics model parameters, the calibrated deformation field distribution data is obtained by resolving the problem through the multiphysics digital twin.

[0078] This step is the final generation of the calibrated deformation field, and the specific implementation process is as follows: Multiphysics coupling re-solution: Using the dynamic boundary condition field generated in step S2 as the load boundary input, the multiphysics digital twin with updated parameters completed in step S53 is used to re-execute the thermal-phase change-force bidirectional strong coupling solution process in step S3. The solution time step and convergence criterion are completely consistent with step S3 to ensure the consistency of the solution process.

[0079] Generation of calibrated deformation field distribution data: After the solution is completed, the output is the deformation field distribution data covering all mesh nodes of the entire mold steel model and all time steps of the entire heat treatment process. Specifically, it includes: the three-dimensional displacement vector, total deformation, principal strain value, strain gradient modulus value, and spatial distribution data of plastic strain and phase transformation strain for each node. The spatial coordinate reference of all data is completely consistent with the initial design geometric model, and the time reference is synchronized with the entire heat treatment process, which fully characterizes the deformation evolution law of the entire process of heat treatment of mold steel after calibration.

[0080] In an optional embodiment of the present invention, the four-dimensional variational assimilation (4DVar) algorithm can be used to replace the ensemble Kalman filter algorithm. The geometric deviation fields of multiple key time nodes in the full heat treatment process are used as continuous observation constraints to achieve global optimization and correction of model parameters. Alternatively, metallographic data of mold steel can be collected at key time nodes, and the measured phase composition data can be used as supplementary observation constraints to further improve the correction accuracy of phase transformation dynamics model parameters.

[0081] S6: Based on the calibrated deformation field distribution data and the deformation feature field data of the current iteration step, calculate the deformation field residual, and reverse-correct the boundary condition parameters of the dynamic boundary condition evolution model based on the deformation field residual. Input the corrected boundary condition parameters into the dynamic boundary condition evolution model to reconstruct the model state, and update the dynamic boundary condition field using the reconstructed dynamic boundary condition evolution model. Specifically, in the first iteration, the deformation feature field data of the current iteration step is the deformation feature field data generated in step S4; in subsequent iterations, the deformation feature field data of the current iteration step is the deformation feature field data regenerated in step S4 based on the updated dynamic boundary condition field. Furthermore, in the above technical solution, the specific steps in S6 for reversely correcting the boundary condition parameters of the dynamic boundary condition evolution model include: S61: Compare the calibrated deformation field distribution data with the deformation feature field data of the current iteration step, and calculate the deformation field residual; This step involves error quantification in the reverse correction process. Its core is establishing a quantification system for the deviation between simulation prediction results and measured calibration results, providing a basis for the objective function in subsequent parameter correction. The specific implementation process is as follows: Spatiotemporal alignment preprocessing of data: Based on the key time node of heat treatment determined in step S51, two sets of data with the same timestamp as the key time node are extracted: one set is the calibrated deformation field distribution data output in step S5, and the other set is the deformation feature field data generated in the current iteration step S4. Using the 12560 surface mesh nodes of the finite element model in step S3 as the spatial reference, the node-level parameters of the two sets of data are matched one-to-one in space to eliminate the calculation errors caused by time difference and spatial misalignment. Specifically, in the first iteration, the deformation feature field data of the current iteration step is the deformation feature field data generated for the first time in step S4 based on the initial dynamic boundary condition field; in subsequent iterations, the deformation feature field data of the current iteration step is the deformation feature field data generated after re-executing steps S3-S4 in the previous iteration after updating the dynamic boundary condition field.

[0082] Node-level deformation residual calculation: For each matched surface mesh node, calculate two types of deformation residuals separately: Displacement residual: The difference between the three-dimensional displacement vector of the node in the calibrated deformation field distribution data and the predicted three-dimensional displacement vector of the node in the deformation characteristic field data is used to obtain the displacement residual vector of the node. , where i is the node number; Strain residual: The strain residual of a node is obtained by calculating the difference between the principal strain value of that node in the calibrated deformation field distribution data and the principal strain value of that node in the deformation characteristic field data. .

[0083] Construction of the overall deformation field residuals: Based on the displacement and strain residuals of all nodes, the objective function for the deformation field residuals is constructed as follows: , Where N is the total number of surface mesh nodes (N=12560 in this embodiment), w u is the weighting coefficient of the displacement residual (0.7 in this embodiment), w is the weighting coefficient of the strain residual (0.3 in this embodiment), and J is the overall deformation field residual, which characterizes the overall deviation between the simulation prediction result and the measured calibration result.

[0084] S62: Based on the deformation field residual, calculate the correction gradient of the boundary condition parameters in the dynamic boundary condition evolution model using the adjoint sensitivity analysis method; This step is the core solution process for the reverse correction. Its core is to quantify the influence of boundary condition parameters on the deformation field residuals through adjoint sensitivity analysis, thereby obtaining the direction and magnitude of the parameter correction. The specific implementation process is as follows: Determination of boundary condition parameters to be corrected: In this embodiment, the boundary condition parameters to be corrected are the heat transfer coefficient values ​​corresponding to the 200 mold steel surface units divided in step S2. The heat transfer coefficient values ​​are the core output parameters of the dynamic boundary condition evolution model and also the core input parameters of the thermal boundary of the multiphysics digital twin in step S3. The heat transfer coefficient of each surface unit corresponds to an independent parameter to be corrected, for a total of 200 parameters to be corrected, denoted as the parameter vector h=[h1, h2, ..., h 200 ] T .

[0085] Construction and solution of the adjoint equation: Based on the heat conduction control equation and thermo-mechanical coupling constitutive equation of the multiphysics digital twin in step S3, the adjoint equation of the residual objective function J of the deformation field with respect to the boundary condition parameter h is constructed; based on the finite element solution results of step S3, the adjoint equation is solved using the adjoint solution module of ABAQUS, combined with a custom user subroutine, to obtain the partial derivatives of the objective function J with respect to each heat transfer coefficient parameter to be corrected. , where j is the surface element number.

[0086] Calculation of the corrected gradient: Normalize the partial derivatives obtained from the solution to obtain the corrected gradient g=[g1, g2, ..., g...] for each boundary condition parameter. 200 ] T ,in The sign of the correction gradient indicates the direction of parameter correction, and the magnitude of the absolute value indicates the magnitude of parameter correction. The larger the absolute value, the greater the influence of the heat transfer coefficient parameter on the residual of the deformation field, and the higher the correction priority.

[0087] S63: Update the boundary condition parameters according to the corrected gradient, and input the updated boundary condition parameters into the dynamic boundary condition evolution model to reconstruct the model state and generate the corrected dynamic boundary condition field.

[0088] This step is the final stage of boundary condition updating and model reconstruction. Its core is to update parameters based on corrected gradients, reconstruct the state of the dynamic boundary condition evolution model, and generate a corrected dynamic boundary condition field adapted to the actual heat treatment process. The specific implementation process is as follows: Gradient descent update of boundary condition parameters: The heat transfer coefficient parameters are updated using a constrained gradient descent method, and the update formula is as follows: , in, The heat transfer coefficient value before the update. Here is the updated heat transfer coefficient value, g is the learning rate (0.01 in this embodiment), and g is the learning rate. jThis is the correction gradient for the corresponding parameters. Simultaneously, physical constraint boundaries for the parameters are set: based on the actual heat transfer characteristics of the H13 mold steel vacuum heat treatment process, the upper and lower limits of the heat transfer coefficient are set to 10 W / (m²). 2 K)h j 2000W / (m 2 K) ensures that the updated parameters conform to the actual physical meaning and avoids non-physical parameter values.

[0089] State reconstruction of the dynamic boundary condition evolution model: The updated heat transfer coefficient values ​​of 200 surface units are used as measured observation values ​​to supplement the input sequence of the dynamic boundary condition evolution model, and the historical input feature sequence of the LSTM model is updated. At the same time, based on the updated parameters, the hidden layer state of the LSTM model is fine-tuned online to reconstruct the model's prediction state, so that the model's prediction output is adapted to the corrected boundary condition law and the prediction system bias of the model is eliminated.

[0090] Generation of the corrected dynamic boundary condition field: Through the dynamic boundary condition evolution model after state reconstruction, the heat transfer coefficient field of 200 surface elements at each time step of the entire heat treatment process is re-predicted; the spatial mapping process of step S23 is repeated, and the updated heat transfer coefficient field is interpolated and mapped to the 12560 surface mesh nodes of the finite element model using the Kriging spatial interpolation method to generate the time- and space-related corrected dynamic boundary condition field; the corrected dynamic boundary condition field fully matches the load boundary input format requirements of the multiphysics digital twin in step S3, and can be directly used for the coupled solution of the next iteration.

[0091] In an optional embodiment of the present invention, the Adam adaptive gradient descent algorithm can be used instead of the ordinary gradient descent method to dynamically adjust the learning rate of parameter updates and improve the convergence speed and stability of parameter correction; or L2 regularization constraints can be added during the parameter update process to avoid model overfitting caused by excessive parameter correction and further improve the generalization ability of boundary condition prediction.

[0092] S7: Based on the updated dynamic boundary condition field, iteratively execute steps S3 to S6 until the preset convergence condition is met, and output the final deformation prediction result.

[0093] Furthermore, in the above technical solutions, such as Figure 3 As shown, the specific steps in S7 that iteratively execute until the preset convergence condition is met include: S71: Calculate the first difference between the deformation feature field data generated in the current iteration step and the deformation feature field data generated in the previous iteration step; This step is the quantification of iterative stability. Its core is to characterize the convergence stability of the iterative process by analyzing the differences in the deformed feature fields of adjacent iterations, providing the first criterion for convergence determination. The specific implementation process is as follows: Data alignment preprocessing: Based on the key time nodes of heat treatment determined in step S51, two sets of matching data are extracted: the first set is the deformation feature field data generated after the completion of step S4 in the current iteration step, which is completely consistent with the timestamp of the key time node; the second set is the deformation feature field data generated after the completion of step S4 in the previous iteration step, which has the same timestamp. Using the 12560 surface mesh nodes of the finite element model in step S3 as the spatial reference, the node-level parameters of the two sets of data are matched one by one in spatial position to eliminate the spatial misalignment caused by mesh updates during iteration and ensure that the two sets of data correspond to the same spatial position and the same heat treatment state.

[0094] In the first iteration, the deformation feature field data of the previous iteration step is the deformation feature field data generated for the first time based on the initial dynamic boundary condition field in step S4, and the deformation feature field data of the current iteration step is the deformation feature field data regenerated after the first boundary condition correction is completed.

[0095] Node-level difference calculation: For each matched surface mesh node, calculate the difference values ​​for the two core dimensions separately. Principal strain value difference: Calculate the absolute difference between the first principal strain value of this node in the current iteration step and the first principal strain value of the same node in the previous iteration step; Strain gradient magnitude difference: Calculate the absolute difference between the strain gradient magnitude of the node in the current iteration step and the strain gradient magnitude of the same node in the previous iteration step.

[0096] Quantitative calculation of the first degree of dissimilarity: The normalized root mean square error (NRMSE) is used to calculate the overall first degree of dissimilarity, eliminating the influence of the dimensions of different parameters. The calculation formula is as follows: , Where D1 is the first difference degree, N is the total number of surface mesh nodes (N=12560 in this embodiment), k is the current iteration step number, and k-1 is the previous iteration step number. i，k This represents the first principal strain value at the i-th node in the current iteration step. i，k-1 This represents the first principal strain value of the i-th node in the previous iteration step. i，k Let be the strain gradient magnitude of the i-th node in the current iteration step. i，k-1 This represents the strain gradient magnitude of the i-th node in the previous iteration step. The first difference degree D1 ranges from 0 to 1. The smaller the value, the smaller the difference in deformation prediction results between adjacent iteration steps, and the more stable the iteration process tends to be. In this embodiment, based on the iteration stability requirements of mold processing, the first threshold is preset to 0.005.

[0097] S72: Calculate the second difference degree between the deformation feature field data generated in the current iteration step and the geometric deviation field; This step quantifies the prediction accuracy. Its core is to characterize the degree of matching between the predicted deformation result and the actual geometric deviation by comparing the difference between the predicted result and the measured geometric deviation, thus providing a second criterion for convergence judgment. The specific implementation process is as follows: Data alignment preprocessing: Based on the key time nodes of heat treatment determined in step S51, extract the deformation feature field data generated after the completion of step S4 in the current iteration step, which is consistent with the timestamp of the key time node, and the geometric deviation field data generated in step S51 at the same time node. Using the 12560 surface mesh nodes of the finite element model in step S3 as the spatial reference, the predicted normal displacement of each node in the deformation feature field data (the total deformation calculated from the principal strain value and the initial coordinates of the node) is matched one-to-one with the measured geometric deviation value of the same node in the geometric deviation field, establishing node-level matching data pairs of predicted deformation and measured deviation values.

[0098] The second degree of dissimilarity is calculated using the same normalized root mean square error. The formula is as follows: , Where D2 is the second difference degree, N is the total number of surface mesh nodes (N=12560 in this embodiment), k is the current iteration step number, and u i，k Let be the predicted normal displacement of the i-th node in the current iteration step. i This represents the measured geometric deviation value of the i-th node in the geometric deviation field. The second difference degree D2 ranges from 0 to 1; the smaller the value, the higher the matching degree between the predicted result and the measured result, and the higher the prediction accuracy. In this embodiment, based on the 0.05mm dimensional tolerance requirement of precision mold heat treatment, the second threshold is preset to 0.01.

[0099] S73: When the first difference is less than a preset first threshold and the second difference is less than a preset second threshold, it is determined that the preset convergence condition is met, the iteration is stopped, and the calibrated deformation field distribution data of the current iteration step is output as the final deformation prediction result.

[0100] This step is the core of iterative convergence judgment and final result output. The specific implementation process is as follows: Convergence condition determination: Convergence is determined using a double threshold AND logic, meaning that the preset convergence condition is met only if both of the following conditions are simultaneously satisfied: Condition 1: The first difference degree D1 calculated in the current iteration step is less than the preset first threshold of 0.005; Condition 2: The second difference degree D2 calculated in the current iteration step is less than the preset second threshold of 0.01. When both conditions are met simultaneously, the iteration loop stops immediately; if either condition is not met, it is determined that the iteration has not converged. Based on the dynamic boundary condition field updated in step S6, the entire process from step S3 to step S6 is re-executed to enter the next iteration. To avoid an iteration dead loop, the maximum number of iterations is set to 20 rounds. When the convergence condition is not met after 20 iterations, the iteration is forcibly stopped, the deformation field distribution data with the smallest second difference degree in the current iteration is output, and a warning message indicating that the iteration has not converged is output simultaneously.

[0101] Output of the final deformation prediction result: When the convergence condition is met and the iteration stops, the calibrated deformation field distribution data output from step S5 of the current iteration step is used as the final deformation prediction result. The output result includes the following standardized content: The spatial distribution data of the three-dimensional displacement vector, normal deformation, and total deformation of all mesh nodes in the mold steel full model are provided. Each data point is accompanied by the corresponding three-dimensional spatial coordinates of the node. It can be directly imported into CAD software such as UG and SolidWorks for heat treatment deformation compensation design. Deformation field evolution data at each key time point in the entire heat treatment process, including principal strain distribution, stress distribution, and location and intensity changes of deformation characteristic regions at each time point; The predicted deformation values ​​of the 8 key points determined in step S1, the deformation trend curve of the entire heat treatment process, and the accuracy evaluation data of the prediction results, including the maximum error, average error, and root mean square error between the predicted value and the measured geometric deviation.

[0102] In an optional embodiment of the present invention, the final deformation prediction result can be back-mapped to the initial design geometric model of the mold steel to generate an optimized design model after heat treatment deformation compensation, which can be directly used for CNC machining of the mold; alternatively, the optimized dynamic boundary condition evolution model parameters and multiphysics digital twin parameters during the iteration process can be added to the historical working condition database of the heat treatment equipment in step S1 for pre-training of the prediction model for subsequent heat treatment of the same type of mold steel, further improving the initial accuracy and convergence speed of subsequent predictions.

[0103] Digital twin prediction system for heat treatment deformation of mold steel, such as Figure 4 As shown, it includes the following modules: The data acquisition module is used to acquire multi-source heterogeneous data of the target mold steel during the heat treatment process; The boundary condition construction module is used to construct a dynamic boundary condition evolution model based on the multi-source heterogeneous data and generate a dynamic boundary condition field. The twin solver module is used to construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model based on the initial material property data. The dynamic boundary condition field is used as the load boundary input, and the stress-strain field is output. A feature field construction module is used to generate deformation feature field data based on the stress-strain field. The calibration module is used to acquire the actual geometric data of the mold steel at key time points during the heat treatment process, construct a geometric deviation field, and calibrate the deformation characteristic field data with the geometric deviation field to obtain the calibrated deformation field distribution data. The parameter correction module is used to calculate the deformation field residual based on the calibrated deformation field distribution data and the deformation feature field data of the current iteration step, and to reverse correct the boundary condition parameters of the dynamic boundary condition evolution model based on the deformation field residual. The corrected boundary condition parameters are then input into the boundary condition construction module to reconstruct the model state. The iterative control module is used to control each module to perform iterative calculations until the preset convergence conditions are met, and output the final deformation prediction result.

[0104] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the digital twin prediction method for heat treatment deformation of mold steel as described above.

[0105] In this embodiment, the electronic device includes a processor, memory, input / output interfaces, and communication interfaces. These components are connected via an industrial-grade motherboard bus to ensure the real-time performance and stability of data transmission. Specifically: Processor: It adopts an Intel Xeon W-2295 industrial-grade processor, paired with an NVIDIA RTX A6000 professional graphics card. The processor has 18 cores, a base frequency of 3.0GHz, a turbo frequency of 4.8GHz, and the graphics card has 48GB of GDDR6 video memory. The processor is used to perform the core logic operations of computer programs, including data processing, model building, iterative optimization, etc. The graphics card is used to accelerate the finite element coupling solution of multiphysics digital twins, the prediction and fine-tuning of LSTM long short-term memory neural network models, and improve the overall computing efficiency.

[0106] Storage: Includes main memory and secondary storage. The main memory uses four 32GB DDR4-3200 ECC industrial-grade memory modules with a total capacity of 128GB. It is used for temporary storage of running computer programs, real-time acquired multi-source heterogeneous data, intermediate results of finite element solutions, etc. The secondary storage uses a 2TB NVMe SSD solid-state drive for long-term storage of the initial material property database of the target mold steel, the historical working condition database of the vacuum heat treatment furnace, the dynamic boundary condition evolution model, the multiphysics digital twin, the actual geometric data acquired by the 3D scanning equipment, the final deformation prediction results, and the computer programs that can run on the processor.

[0107] Input / output interfaces include a USB 3.2 Gen22 interface, a Gigabit Ethernet interface, and an RS485 industrial communication interface. The USB 3.2 Gen22 interface is used to connect to an industrial-grade laser 3D scanner to achieve high-speed acquisition of actual geometric data of the mold steel at key time points. The Gigabit Ethernet interface is used to connect to the industrial control module of the vacuum heat treatment furnace to achieve synchronous transmission of real-time process parameters and historical operating condition data. The RS485 industrial communication interface is used to connect to the K-type armored thermocouple and high-temperature strain gauge embedded inside the target mold steel to achieve stable acquisition of real-time response data.

[0108] Communication interfaces: Optional, including 5G industrial modules and Wi-Fi 6 modules, used to connect to the cloud-based industrial internet platform to realize the uploading of prediction results, cloud backup of historical operating data, and cloud update of dynamic boundary condition evolution models.

[0109] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the digital twin prediction method for heat treatment deformation of mold steel as described above.

[0110] The computer program is stored in an NVMe SSD solid-state drive on external storage. When the electronic device starts up, the processor loads the computer program from external storage into memory and runs it. When the processor executes the computer program, it sequentially implements all the steps of the digital twin prediction method for heat treatment deformation of mold steel as described in any of the foregoing embodiments: Call the input / output interface, execute step S1, obtain multi-source heterogeneous data of the target mold steel during the heat treatment process, including real-time process parameters, real-time response data, initial material property data and historical operating condition data of the heat treatment equipment, and store the data to external storage; The processor and graphics card are invoked to execute step S2, which constructs a dynamic boundary condition evolution model based on multi-source heterogeneous data and generates a dynamic boundary condition field that is related to time and space. Call the graphics card to accelerate the finite element solution, execute step S3, construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model based on the initial material property data, use the dynamic boundary condition field as the load boundary input, perform multiphysics coupling solution on the heat treatment process, and output the stress and strain field. The processor is invoked to perform tensor operations. Step S4 is executed to construct a deformation feature field evolution model based on the stress-strain field, extract and quantify the non-uniform deformation characteristics of the mold steel during the heat treatment process, and generate deformation feature field data. Call the input / output interface to obtain actual geometric data, execute step S5, construct the geometric deviation field, and calibrate the deformation feature field data with the geometric deviation field to obtain the calibrated deformation field distribution data; The processor is invoked to perform the accompanying sensitivity analysis, and step S6 is executed to calculate the deformation field residual, correct the boundary condition parameters of the dynamic boundary condition evolution model in reverse, reconstruct the model state, and update the dynamic boundary condition field. The processor is invoked to execute iterative convergence control. Step S7 is executed, and steps S3 to S6 are executed iteratively based on the updated dynamic boundary condition field until the preset convergence condition is met. The final deformation prediction result is output and stored in external memory. At the same time, it is output to the display device or cloud platform through the input / output interface or communication interface.

[0111] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A digital twin prediction method for heat treatment deformation of mold steel, characterized in that, Specifically, the following steps are included: S1: Obtain multi-source heterogeneous data of the target mold steel during the heat treatment process. The multi-source heterogeneous data includes real-time process parameters, real-time response data, initial material property data, and historical operating condition data of the heat treatment equipment. S2: Based on the real-time process parameters, real-time response data and historical operating condition data in the multi-source heterogeneous data, construct a dynamic boundary condition evolution model to generate a dynamic boundary condition field that is related to time and space. S3: Based on the initial material property data, construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model. Using the dynamic boundary condition field as the load boundary input, perform multiphysics coupling solution on the heat treatment process and output the stress-strain field. S4: Based on the stress-strain field, construct a deformation characteristic field evolution model, extract and quantify the non-uniform deformation characteristics of the mold steel during the heat treatment process, and generate deformation characteristic field data; S5: Obtain the actual geometric data of the mold steel at key time points during the heat treatment process, construct a geometric deviation field based on the actual geometric data, and calibrate the deformation characteristic field data with the geometric deviation field to obtain the calibrated deformation field distribution data; S6: Based on the calibrated deformation field distribution data and the deformation feature field data of the current iteration step, calculate the deformation field residual, and reverse-correct the boundary condition parameters of the dynamic boundary condition evolution model based on the deformation field residual. Input the corrected boundary condition parameters into the dynamic boundary condition evolution model to reconstruct the model state, and update the dynamic boundary condition field using the reconstructed dynamic boundary condition evolution model. Specifically, in the first iteration, the deformation feature field data of the current iteration step is the deformation feature field data generated in step S4; in subsequent iterations, the deformation feature field data of the current iteration step is the deformation feature field data regenerated in step S4 based on the updated dynamic boundary condition field. S7: Based on the updated dynamic boundary condition field, iteratively execute steps S3 to S6 until the preset convergence condition is met, and output the final deformation prediction result.

2. The digital twin prediction method for heat treatment deformation of mold steel according to claim 1, characterized in that, The real-time process parameters include at least heating rate, holding time, cooling medium flow rate, and quenching medium temperature; the real-time response data includes at least temperature data and strain data of key points inside the mold steel; and the initial material property data includes at least thermophysical parameters, phase transformation kinetic parameters, and high-temperature rheological stress parameters.

3. The digital twin prediction method for heat treatment deformation of mold steel according to claim 1, characterized in that, The specific steps for constructing the dynamic boundary condition evolution model in S2 include: S21: Obtain historical and real-time operating data of the heat treatment equipment, and extract the distribution characteristics of the heat transfer coefficient in the spatial and temporal dimensions; S22: Based on the distribution characteristics, a time-series prediction model is adopted, using the real-time process parameters and the historical operating data as drivers, to predict the heat transfer coefficient field at the next moment. S23: Obtain the geometric contour data of the mold steel during the heat treatment process, and spatially map the geometric contour data with the heat transfer coefficient field to generate the dynamic boundary condition field corresponding to the grid nodes on the surface of the mold steel.

4. The digital twin prediction method for heat treatment deformation of mold steel according to claim 1, characterized in that, The specific steps for constructing the deformation feature field evolution model in S4 include: S41: Extract the plastic strain tensor field and the phase transformation strain tensor field from the stress-strain field; S42: Perform eigendecomposition on the plastic strain tensor field and the phase transformation strain tensor field to obtain the principal strain direction and principal strain value, and calculate the deformation gradient field based on the principal strain direction and principal strain value. S43: Perform spatial differentiation on the deformation gradient field, extract the region where the strain gradient exceeds a preset threshold as the deformation feature region, and generate the deformation feature field data containing the spatial distribution information and intensity information of the deformation feature region.

5. The digital twin prediction method for heat treatment deformation of mold steel according to claim 1, characterized in that, The specific steps in S5 for calibrating the deformation feature field data with the geometric deviation field include: S51: Collect actual geometric data of mold steel at key time points of heat treatment using a three-dimensional scanning device, compare the actual geometric data with the design geometric data, and generate the geometric deviation field; S52: Align the deformation feature field data with the geometric deviation field in the time dimension to establish a mapping relationship between the deformation feature field and the geometric deviation field; S53: Using a data assimilation algorithm, the mapping relationship is used as an observation constraint to reverse the parameters of the thermo-mechanical coupling model and the phase transition dynamics model in the multiphysics digital twin, and the corrected thermo-mechanical coupling model parameters and phase transition dynamics model parameters are updated in the multiphysics digital twin. S54: Based on the corrected thermo-mechanical coupling model parameters and phase transition dynamics model parameters, the calibrated deformation field distribution data is obtained by resolving the problem through the multiphysics digital twin.

6. The digital twin prediction method for heat treatment deformation of mold steel according to claim 1, characterized in that, The specific steps in S6 for reverse correction of the boundary condition parameters of the dynamic boundary condition evolution model include: S61: Compare the calibrated deformation field distribution data with the deformation feature field data of the current iteration step, and calculate the deformation field residual; S62: Based on the deformation field residual, calculate the correction gradient of the boundary condition parameters in the dynamic boundary condition evolution model using the adjoint sensitivity analysis method; S63: Update the boundary condition parameters according to the corrected gradient, and input the updated boundary condition parameters into the dynamic boundary condition evolution model to reconstruct the model state and generate the corrected dynamic boundary condition field.

7. The digital twin prediction method for heat treatment deformation of mold steel according to claim 1, characterized in that, The specific steps in S7 that iterate until the preset convergence condition is met include: S71: Calculate the first difference between the deformation feature field data generated in the current iteration step and the deformation feature field data generated in the previous iteration step; S72: Calculate the second difference degree between the deformation feature field data generated in the current iteration step and the geometric deviation field; S73: When the first difference is less than a preset first threshold and the second difference is less than a preset second threshold, it is determined that the preset convergence condition is met, the iteration is stopped, and the calibrated deformation field distribution data of the current iteration step is output as the final deformation prediction result.

8. A digital twin prediction system for heat treatment deformation of mold steel, characterized in that, Includes the following modules: The data acquisition module is used to acquire multi-source heterogeneous data of the target mold steel during the heat treatment process; The boundary condition construction module is used to construct a dynamic boundary condition evolution model based on the multi-source heterogeneous data and generate a dynamic boundary condition field. The twin solver module is used to construct a multiphysics digital twin containing a phase transformation dynamics model and a thermo-mechanical coupling model based on the initial material property data. The dynamic boundary condition field is used as the load boundary input, and the stress-strain field is output. A feature field construction module is used to generate deformation feature field data based on the stress-strain field. The calibration module is used to acquire the actual geometric data of the mold steel at key time points during the heat treatment process, construct a geometric deviation field, and calibrate the deformation characteristic field data with the geometric deviation field to obtain the calibrated deformation field distribution data. The parameter correction module is used to calculate the deformation field residual based on the calibrated deformation field distribution data and the deformation feature field data of the current iteration step, and to reverse correct the boundary condition parameters of the dynamic boundary condition evolution model based on the deformation field residual. The corrected boundary condition parameters are then input into the boundary condition construction module to reconstruct the model state. The iterative control module is used to control each module to perform iterative calculations until the preset convergence conditions are met, and output the final deformation prediction result.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the digital twin prediction method for heat treatment deformation of mold steel as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the digital twin prediction method for heat treatment deformation of mold steel as described in any one of claims 1-7.

Images