A Machine Learning-Based Parameter Optimization Method for Multi-Pass Forging of Difficult-to-Deform Alloys

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and metal plastic forming, and discloses a machine learning-based method for optimizing multi-pass forging parameters of difficult-to-deform alloys. The method includes: constructing a physical constraint sub-model integrating dislocation density evolution, dynamic recrystallization, and grain growth mechanisms; acquiring multi-pass time-series process data through infrared thermography and optical strain measurement; predicting the microstructure state of each pass using a graph attention neural network; performing residual correction by combining the physical model and data prediction to generate a microstructure evolution trajectory; calculating the grain boundary weakening index and constructing a multi-objective optimization function incorporating energy consumption, grain uniformity, and grain boundary stability; and using an improved non-dominated sorting genetic algorithm to solve for the optimal forging parameter sequence. The system includes corresponding functional modules. This invention can significantly reduce cracking risk, improve microstructure uniformity, and reduce energy consumption.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and metal plastic forming, specifically involving a method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning. Background Technology

[0002] With the increasing demand for high-performance metal components in aerospace, energy equipment, and high-end manufacturing, difficult-to-deform alloys (such as nickel-based superalloys and titanium-aluminum alloys) are widely used in key load-bearing components due to their excellent high-temperature strength, creep resistance, and corrosion resistance. Multi-pass forging, as the core hot working process for controlling its macroscopic forming accuracy and microstructure properties, requires repeated plastic deformation under high temperature and high strain rate conditions.

[0003] This process involves not only complex thermo-mechanical coupling field evolution, but also significant microstructural dynamic responses, including dislocation multiplication, dynamic recrystallization and grain refinement mechanisms, which directly determine the mechanical properties and service reliability of the final product.

[0004] Machine learning-based forging parameter optimization methods have become a research hotspot in recent years, aiming to establish a mapping relationship between process parameters (such as pass reduction, temperature, and strain rate) and forming quality (such as streamline distribution and grain size) through data-driven approaches. These methods typically rely on historical experimental or simulation data to train models to predict the optimal process window. However, existing models generally treat each pass as an independent event, failing to adequately model the evolutionary continuity of microstates between passes, and particularly neglecting the accumulation of dislocation density and relaxation effects during dynamic recrystallization.

[0005] Traditional machine learning models (such as support vector machines, random forests, or shallow neural networks) can fit macroscopic input-output relationships, but they lack an embedded representation of the physical mechanisms, making it difficult to capture the nonlinear decay and reactivation behavior of dislocation density across multiple passes. When a dislocation density abruptly changes due to improper parameters in a pass, subsequent passes, if not adjusted in time, can easily lead to local grain boundary weakening, micropore aggregation, and even macroscopic cracking. Furthermore, purely data-driven methods have limited generalization ability in small sample sizes and high-dimensional parameter spaces, and cannot effectively extrapolate to unseen conditions. Summary of the Invention

[0006] This invention provides a machine learning-based method for optimizing parameters in multi-pass forging of difficult-to-deform alloys. By constructing a hybrid intelligent system that integrates physical metallurgical mechanisms and data-driven modeling, it quantifies and characterizes the cumulative effect of dynamic recrystallization during multi-pass forging, accurately capturing the cross-pass evolution of dislocation density, grain size, and grain boundary energy. This allows for simultaneous constraint of microstructure stability and macroscopic forming performance during parameter optimization, avoiding cracking defects caused by local grain boundary weakening.

[0007] This invention provides a machine learning-based method for optimizing multi-pass forging parameters of difficult-to-deform alloys, comprising:

[0008] Acquire material property data, thermodynamic boundary condition data, and geometric specification data of the target forging from the initial billet;

[0009] Based on the material property data and thermodynamic boundary condition data, a physical constraint sub-model is constructed, which includes the dislocation density evolution equation, the dynamic recrystallization volume fraction kinetic model, and the grain growth rate function. Through a high-precision infrared temperature measurement array and strain field optical measurement system, the surface temperature distribution, equivalent strain field, and strain rate field data of each forging process are collected in real time to form a multi-pass time-series process sensing dataset.

[0010] The multi-pass time-series process sensing dataset is input into a pre-trained graph attention neural network. The network uses forging passes as nodes and thermal coupling relationships between passes as edges to propagate and aggregate the microstructure state of each pass across passes, and outputs the predicted values ​​of dislocation density, average grain size, and grain boundary energy density at the end of each pass.

[0011] The output of the physical constraint sub-model and the prediction result of the graph attention neural network are subjected to residual correction to generate a micro-organism evolution trajectory that integrates physical priors and data observations.

[0012] Based on the microstructure evolution trajectory, the grain boundary weakening index corresponding to each pass is calculated. The grain boundary weakening index is defined as the ratio of the grain boundary energy density of the current pass to the critical grain boundary fracture threshold.

[0013] A multi-objective optimization function is constructed, with objective terms including total forging energy consumption, final grain uniformity index and maximum grain boundary weakening index, and constraint terms including equipment load limit, die life limit and final forging temperature window.

[0014] An improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function to generate an optimal forging parameter sequence that satisfies the stability of the microstructure. The forging parameter sequence includes the initial forging temperature, final forging temperature, reduction, strain rate and pass interval time for each pass.

[0015] The optimal forging parameter sequence is sent to the forging execution control system, which drives the hydraulic servo mechanism to perform multiple forming operations in sequence.

[0016] In one embodiment of the present invention, the material property data includes the mass percentage of alloy composition, initial grain size, initial dislocation density, activation energy parameter, and thermal conductivity temperature function; the thermodynamic boundary condition data includes ambient temperature, mold preheating temperature, lubrication condition thermal resistance coefficient, and cooling medium heat transfer coefficient.

[0017] In one embodiment of the present invention, the dislocation density evolution equation adopts the Kocks-Mecking form, and its expression is as follows: ,in For dislocation density, The rate of change of dislocation density over time. The equivalent rate of change, and For material constants, To dynamically restore activation energy, The gas constant is... The absolute temperature is used; the dynamic recrystallization volume fraction kinetic model adopts the Avrami type equation, and its expression is: ,in This refers to the volume fraction of dynamic recrystallization. For the heat preservation time of each pass, The time required for recrystallization to reach a volume fraction of 50%, The Avrami index is used; the grain growth rate function adopts the Hillert model, and its expression is: ,in This is the current grain size. This refers to the grain size at the point where recrystallization is complete. The growth rate constant is denoted by . This is the activation energy for grain boundary migration.

[0018] As one embodiment of the present invention, the high-precision infrared temperature measurement array consists of no less than 64 independent temperature measurement units, with a spatial resolution of 5 mm and a sampling frequency of 100 Hz; the strain field optical measurement system adopts digital image correlation technology, is equipped with a binocular high-speed camera, has a frame rate of 500 frames per second, and a strain measurement accuracy of 1‰.

[0019] As one embodiment of the present invention, the graph attention neural network includes three graph convolutional layers, each layer having a node feature dimension of 128. The edge weights are calculated by the temperature gradient, cumulative strain increment, and time interval between adjacent traces. The attention coefficients are generated through a learnable query-key mechanism and are used to weight and aggregate the influence intensity of historical traces on the microstate of the current trace.

[0020] In one embodiment of the present invention, the residual correction employs a weighted least squares method, where the weight matrix is ​​determined by the uncertainty covariance of the physical model in the high-temperature, high-strain region, and the corrected dislocation density... ,in Output for the physical model. For the graph attention network's predicted values, It is a diagonal weight matrix, and its elements take values ​​from 0.3 to 0.7.

[0021] As one embodiment of the present invention, the critical grain boundary fracture threshold of the grain boundary weakening index is preset according to the alloy type. For nickel-based superalloys, the threshold is 1.2 joules per square meter; for titanium-aluminum alloys, the threshold is 0.8 joules per square meter; and for austenitic stainless steel, the threshold is 1.0 joules per square meter.

[0022] In one embodiment of the present invention, the multi-objective optimization function is defined as:

[0023]

[0024] Total forging energy consumption, The final grain size standard deviation, The average grain size, The final grain size standard deviation, Let be the grain boundary weakening index of the i-th pass. , , The normalized weighting coefficients are set to 0.4, 0.3, and 0.3 respectively; the upper limit of the equipment load is set to 90% of the rated pressure of the hydraulic press; the mold life limit is set to a cumulative reduction of no more than 5,000 tons for a single mold; and the final forging temperature window is set to 30°C above and 20°C below the alloy recrystallization termination temperature.

[0025] As one embodiment of the present invention, the improved non-dominated sorting genetic algorithm adopts an elite retention strategy, with a population size of 200, a crossover probability of 0.9, a mutation probability of 0.1, and a Gaussian perturbation for the mutation operation. The standard deviation is 5% of the parameter range, and a diversity maintenance mechanism based on crowding distance is introduced to ensure the uniformity of the Pareto front solution set.

[0026] This invention provides a machine learning-based multi-pass forging parameter optimization system for difficult-to-deform alloys, comprising:

[0027] The material and boundary condition input module is used to acquire the material property data, thermodynamic boundary condition data, and geometric specification data of the target forging of the initial billet.

[0028] The physical constraint sub-model construction module is used to construct a physical constraint sub-model, including a dislocation density evolution equation, a dynamic recrystallization volume fraction kinetic model, and a grain growth rate function, based on the material property data and thermodynamic boundary condition data.

[0029] The multi-pass process sensing data acquisition module is used to collect surface temperature distribution, equivalent strain field and strain rate field data in real time during each forging process through a high-precision infrared temperature measurement array and strain field optical measurement system, forming a multi-pass time-series process sensing dataset.

[0030] The microstructure state cross-pass prediction module is used to input the multi-pass time-series process sensing dataset into a pre-trained graph attention neural network and output the dislocation density prediction value, average grain size prediction value, and grain boundary energy density prediction value at the end of each pass.

[0031] The physics-data fusion correction module is used to perform residual correction on the output results of the physical constraint sub-model and the prediction results of the graph attention neural network to generate a micro-organization evolution trajectory that integrates physical priors and data observations.

[0032] The grain boundary weakening index calculation module is used to calculate the grain boundary weakening index corresponding to each pass based on the microstructure evolution trajectory.

[0033] The multi-objective optimization function construction module is used to construct a multi-objective optimization function with total forging energy consumption, final grain uniformity index and maximum grain boundary weakening index as objective terms, and equipment load limit, die life limit and final forging temperature window as constraint terms.

[0034] The optimal parameter sequence solving module is used to solve the multi-objective optimization function using an improved non-dominated sorting genetic algorithm to generate an optimal forging parameter sequence that satisfies the stability of the microstructure.

[0035] The forging execution control module is used to send the optimal forging parameter sequence to the forging execution control system, driving the hydraulic servo mechanism to perform multiple forming operations in sequence.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention is the first to explicitly model the cross-track cumulative effect of dynamic recrystallization as a quantifiable microstructure evolution trajectory, and realizes the dynamic influence of inter-track thermal history on the current microstate through graph attention neural network, which solves the fundamental defect of traditional machine learning models that only rely on single-track input and ignore the tissue memory effect.

[0038] By introducing a physical constraint sub-model and a data-driven prediction residual correction mechanism, the physical consistency and data adaptability of the microstate prediction under extreme conditions of high temperature and large deformation are ensured. The defined grain boundary weakening index is directly related to the abrupt change in dislocation density and the change in grain boundary energy, providing a calculable early warning indicator for cracking risk.

[0039] The multi-objective optimization function simultaneously constrains energy consumption, microstructure uniformity, and grain boundary stability, ensuring that the generated forging parameter sequence not only meets forming requirements but also fundamentally suppresses the formation of locally weakened regions. Experimental verification shows that the forging process optimized by the method of this invention can reduce the cracking rate of difficult-to-deform alloy forgings by more than 65%, reduce grain size dispersion by 40%, and reduce forging energy consumption by 12%, significantly improving the yield and service reliability of high-end forgings. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall technical solution architecture of the machine learning-based multi-pass forging parameter optimization method for difficult-to-deform alloys proposed in this invention.

[0041] Figure 2 This is a schematic diagram of the core principle framework of the graph attention neural network in this invention for cross-channel propagation and aggregation of multi-channel micro-organization states;

[0042] Figure 3 This is a logical flowchart of the process of performing residual correction on the physical constraint sub-model and data-driven prediction results in this invention to generate a fused micro-organism evolution trajectory.

[0043] Figure 4 This is a flowchart illustrating the logical process of constructing and solving the optimal forging parameter sequence using a multi-objective optimization function in this invention.

[0044] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of multi-pass process sensing data acquisition and real-time input to the graph attention neural network in this invention;

[0045] Figure 6 This is a flowchart illustrating the logical process of calculating the grain boundary weakening index and dynamically warning of cracking risk during multi-pass forging in this invention. Detailed Implementation

[0046] Please refer to Figures 1 to 6 This invention provides a machine learning-based method for optimizing multi-pass forging parameters of difficult-to-deform alloys. It aims to address the microstructure instability and local grain boundary weakening problems caused by traditional machine learning models neglecting the cumulative effect of dynamic recrystallization across passes during multi-pass forging. This method integrates physical metallurgical mechanisms with data-driven modeling to construct a microstructure evolution trajectory with historical dependence between passes. Based on this trajectory, multi-objective parameter optimization is performed to ensure that the forging process effectively suppresses cracking risk while meeting macroscopic forming requirements.

[0047] The method includes the following steps:

[0048] S1, obtain the material property data, thermodynamic boundary condition data and geometric specification data of the initial billet;

[0049] S2, Based on the material property data and thermodynamic boundary condition data, a physical constraint sub-model is constructed, which includes a dislocation density evolution equation, a dynamic recrystallization volume fraction kinetic model, and a grain growth rate function;

[0050] S3 uses a high-precision infrared temperature measurement array and strain field optical measurement system to collect surface temperature distribution, equivalent strain field and strain rate field data in real time during each forging process, forming a multi-pass time-series process sensing dataset.

[0051] S4. The multi-pass time-series process sensing dataset is input into a pre-trained graph attention neural network. The network uses forging passes as nodes and thermal coupling relationships between passes as edges to propagate and aggregate the microstructure state of each pass across passes, and outputs the predicted values ​​of dislocation density, average grain size, and grain boundary energy density at the end of each pass.

[0052] S5, perform residual correction on the output of the physical constraint sub-model and the prediction result of the graph attention neural network to generate a micro-organization evolution trajectory that integrates physical priors and data observations;

[0053] S6. Based on the microstructure evolution trajectory, calculate the grain boundary weakening index corresponding to each pass. The grain boundary weakening index is defined as the ratio of the grain boundary energy density of the current pass to the critical grain boundary fracture threshold.

[0054] S7, construct a multi-objective optimization function, whose objective terms include total forging energy consumption, final grain uniformity index and maximum grain boundary weakening index, and constraint terms include equipment load limit, die life limit and final forging temperature window;

[0055] S8. An improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function to generate an optimal forging parameter sequence that satisfies the stability of the microstructure. The forging parameter sequence includes the initial forging temperature, final forging temperature, reduction, strain rate and pass interval time for each pass.

[0056] S9, the optimal forging parameter sequence is sent to the forging execution control system to drive the hydraulic servo mechanism to perform multiple forming operations in sequence.

[0057] In step S1, the material property data, thermodynamic boundary condition data, and geometric specification data of the target forging are obtained from the initial billet. The material property data includes the alloy composition by mass percentage, initial grain size, initial dislocation density, activation energy parameter, and thermal conductivity temperature function. Thermodynamic boundary condition data includes ambient temperature, mold preheating temperature, lubrication thermal resistance coefficient, and cooling medium heat transfer coefficient.

[0058] The target forging geometric specifications data include the final forging's three-dimensional profile, critical section thickness, draw ratio, and local reinforcing rib structure. These three types of data collectively constitute the initial input set for subsequent modeling and optimization. Material property data are obtained through laboratory metallographic analysis, X-ray diffraction, and thermophysical performance testing; thermodynamic boundary condition data are recorded in real-time by on-site sensors or set according to process specifications; the target forging geometric specifications data are exported from the computer-aided design model and converted into an analytical geometric feature vector. All data must be standardized before input to eliminate dimensional differences and unify to the same time reference.

[0059] In step S2, a physical constraint sub-model is constructed based on the material property data and thermodynamic boundary condition data. This sub-model consists of three core equations: a dislocation density evolution equation, a dynamic recrystallization volume fraction kinetic model, and a grain growth rate function. The dislocation density evolution equation adopts the Kocks-Mecking form, and its expression is as follows:

[0060]

[0061] in, Dislocation density, in units of ; The rate of change of dislocation density over time. The equivalent rate of change is expressed in units of . ; and is a material constant, determined by fitting the high-temperature compression experiment; The activation energy is the dynamic recovery energy, expressed in joules per mole. The gas constant is 8.314 joules per mole Kelvin; The absolute temperature is expressed in Kelvin. This equation describes the competition mechanism between dislocation multiplication and dynamic recovery during plastic deformation.

[0062] The dynamic recrystallization volume fraction kinetic model adopts the Avrami type equation, the expression of which is:

[0063]

[0064] in, This is the volume fraction of dynamic recrystallization, dimensionless. The heat preservation time for each pass is in seconds. The time required for recrystallization volume fraction to reach 50% is determined by the Zener-Hollomon parameter correlation. The Avrami index reflects recrystallization nucleation and growth mechanisms, with typical values ​​ranging from 1.5 to 3.0.

[0065] The grain growth rate function adopts the Hillert model, and its expression is:

[0066]

[0067] in, This refers to the current grain size, in micrometers. This refers to the grain size at the point where recrystallization is complete. This is the growth rate constant; This is the activation energy for grain boundary migration. This model is applicable to the static or quasi-static growth stage of grains after recrystallization.

[0068] The three equations described above constitute a closed physical evolution system, which can deduce the microstructure state step by step under given initial conditions and thermodynamic paths. This sub-model is embedded as prior knowledge in the subsequent data fusion process to ensure that the prediction results conform to the basic laws of metallurgy.

[0069] In step S3, a high-precision infrared temperature measurement array and a strain field optical measurement system are used to collect real-time data on the surface temperature distribution, equivalent strain field, and strain rate field during each forging process. The infrared temperature measurement array consists of no fewer than 64 independent temperature measurement units, with a spatial resolution of 5 mm and a sampling frequency of 100 Hz, covering the entire upper surface area of ​​the billet. Each temperature measurement unit outputs a continuous temperature time series, which is then interpolated in time and space to form a two-dimensional temperature field matrix. The strain field optical measurement system employs digital image correlation technology, equipped with a binocular high-speed camera, a frame rate of 500 frames per second, and a strain measurement accuracy of 1‰.

[0070] The system sprays a random speckle pattern onto the surface of the billet, calculates the overall displacement vector by tracking the speckle displacement, and then obtains the equivalent strain and strain rate tensor field through spatial gradient calculation. Data acquisition for each pass begins when the billet contacts the mold and ends when it is completely demolded, forming a complete pass-time data block. All data is stored according to pass number and is accompanied by timestamps, equipment status codes, and environmental disturbance markers, constituting a multi-pass time-series process sensing dataset.

[0071] In step S4, the multi-pass time-series process sensing dataset is input into a pre-trained graph attention neural network. This network uses forging passes as graph nodes, and the node feature vector consists of the mean temperature field, standard deviation of the strain field, peak strain rate, and pass duration for that pass, with a dimension of 128. The edges of the graph represent the thermo-mechanical coupling relationship between passes, and the edge weights are calculated jointly by the temperature gradient, cumulative strain increment, and time interval between adjacent passes, using the following formula:

[0072]

[0073] in, For the first The order and the number Edge weights between tracks; , The average temperature of the two passes; , The first Dao Ci, Di After each pass, the material undergoes cumulative equivalent change. This refers to the interval between track sessions; , , These are normalization coefficients, with values ​​of 0.4, 0.35, and 0.25 respectively. This weight reflects the intensity of the influence of historical races on the current race's micro-state.

[0074] The graph attention neural network comprises three graph convolutional layers, each performing a neighborhood aggregation operation on node features. Within each layer, attention coefficients are generated through a learnable query-key mechanism to weight and aggregate the influence of historical paths on the microstate of the current path. Specifically, for a node… Its neighboring nodes Attention coefficient The calculation is as follows:

[0075]

[0076] in, , For node features; The weight matrix is ​​a learnable weight matrix; This is the attention vector; This represents vector concatenation. The aggregated node features are processed by a nonlinear activation function and then passed to the next layer. The network output consists of the dislocation density prediction, average grain size prediction, and grain boundary energy density prediction at the end of each pass, corresponding to three independent fully connected decoders.

[0077] In step S5, the output of the physical constraint sub-model and the prediction of the graph attention neural network are subjected to residual correction. The correction uses weighted least squares, and the weight matrix is ​​determined by the uncertainty covariance of the physical model in the high-temperature, high-strain region. The corrected dislocation density... The calculation formula is:

[0078]

[0079] in, The dislocation density output by the physical model; The dislocation density predicted by the graph attention network; The diagonal weight matrix has elements ranging from 0.3 to 0.7, with lower values ​​in the high-temperature region to enhance data-driven correction and higher values ​​in the low-temperature region to preserve the dominance of the physical model. A similar correction mechanism is applied simultaneously to grain size and grain boundary energy density. The three sets of corrected microscopic parameters constitute a microstructure evolution trajectory that integrates physical priors and data observations, exhibiting cross-channel continuity and physical consistency.

[0080] In step S6, based on the microstructure evolution trajectory, the grain boundary weakening index corresponding to each pass is calculated. The grain boundary weakening index of the i-th pass... Defined as the current grain boundary energy density With critical grain boundary fracture threshold The ratio, that is:

[0081]

[0082] Grain boundary energy density From dislocation density With average grain size The joint calculation is expressed as follows:

[0083]

[0084] in These are material-related constants. The critical grain boundary fracture threshold is preset according to the alloy type: for nickel-based superalloys, It is 1.2 joules per square meter; for titanium-aluminum alloys, It is 0.8 joules per square meter; for austenitic stainless steel, It is 1.0 joules per square meter. When the grain boundary weakening index of the i-th pass... A value exceeding 1 indicates a risk of grain boundary weakening in the current pass, which needs to be suppressed during optimization.

[0085] In step S7, a multi-objective optimization function is constructed. The objective term of this function includes total forging energy consumption. Final grain uniformity index and maximum grain boundary weakening index The total forging energy consumption is obtained by summing the load-displacement integrals of each pass. This represents the standard deviation of the final grain size. For the average grain size, the multi-objective optimization function is defined as:

[0086]

[0087] in, , , The normalized weighting coefficients are set to 0.4, 0.3, and 0.3 respectively. Constraints include: the upper limit of equipment load is set to 90% of the hydraulic press's rated pressure; the die life limit is set to a cumulative reduction of no more than 5000 tons for a single die set; and the final forging temperature window is set to 30°C above and 20°C below the alloy recrystallization termination temperature. All constraints are embedded in the optimization framework in the form of inequalities.

[0088] In step S8, an improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function. The algorithm population size is 200, the crossover probability is 0.9, and the mutation probability is 0.1. Simulated binary crossover is used for the crossover operation, and Gaussian perturbation is used for the mutation operation, with a standard deviation of 5% of the parameter range. The algorithm introduces an elite retention strategy, preserving the Pareto optimal solution set in each generation, and employs a diversity maintenance mechanism based on crowding distance to ensure that the solution set is uniformly distributed in the target space. The optimization variables are the initial forging temperature, final forging temperature, reduction, strain rate, and pass interval time for each pass, forming a high-dimensional decision vector. After the algorithm iterates to convergence or reaches the maximum number of generations, it outputs a set of Pareto optimal forging parameters.

[0089] In step S9, the optimal forging parameter sequence is sent to the forging execution control system. The control system parses the parameter sequence and generates corresponding hydraulic servo commands, including the master cylinder pressure curve, slide speed profile, and cooling spray timing. The system monitors execution deviations in real time. If the actual temperature or strain deviates from the set value by more than the tolerance band, an online re-optimization mechanism is triggered, which calls the lightweight graph attention subnetwork to fine-tune local parameters, ensuring that the microstructure evolution trajectory always remains within the safe domain.

[0090] This embodiment fully realizes a closed-loop process from data acquisition, microstate prediction, physical-data fusion, risk assessment to multi-objective optimization and execution control. The entire methodology solves the problem of crack misjudgment caused by neglecting microstructural memory in traditional methods by explicitly modeling the cumulative effect of dynamic recrystallization between passes. The introduction of a graph attention neural network quantifies the impact of historical thermal paths on the current microstate, while the physical constraint sub-model ensures the reliability of predictions under extreme conditions. The grain boundary weakening index, as a calculable metallurgical risk indicator, directly guides the optimization direction, ensuring that the generated parameter sequence possesses both process feasibility and microstructural stability.

[0091] The system comprises a material and boundary condition input module, a physical constraint sub-model construction module, a multi-pass process sensing data acquisition module, a microstructure state cross-pass prediction module, a physical-data fusion correction module, a grain boundary weakening index calculation module, a multi-objective optimization function construction module, an optimal parameter sequence solution module, and a forging execution control module. These modules are interconnected via a unified data bus, supporting parallel execution of real-time data streams and batch processing tasks. The material and boundary condition input module receives structured input from the laboratory database and the process planning system; the physical constraint sub-model construction module calls the material parameter library to automatically generate differential equation solver instances; the multi-pass process sensing data acquisition module interfaces with the field sensor network to achieve millisecond-level data synchronization; the microstructure state cross-pass prediction module deploys a pre-trained graph attention neural network, supporting GPU-accelerated inference; the physical-data fusion correction module performs residual calculation and weighted fusion; the grain boundary weakening index calculation module outputs the risk level of each pass in real time; the multi-objective optimization function construction module dynamically assembles objectives and constraints; the optimal parameter sequence solution module encapsulates an improved non-dominated sorting genetic algorithm kernel; and the forging execution control module generates equipment control commands and provides feedback on the execution status. The system adopts a modular architecture, with standardized interfaces for each component, making it easy to integrate into existing smart manufacturing platforms.

[0092] The method and system described in this embodiment, through deep integration of physical mechanisms and machine learning, achieves accurate prediction and active control of the microstructure evolution during multi-pass forging of difficult-to-deform alloys, fundamentally improving the forming quality and service reliability of high-end forgings.

Claims

1. A method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning, characterized in that, include: Acquire material property data, thermodynamic boundary condition data, and geometric specification data of the target forging from the initial billet; Based on the material property data and thermodynamic boundary condition data, a physical constraint sub-model is constructed, which includes a dislocation density evolution equation, a dynamic recrystallization volume fraction kinetic model, and a grain growth rate function. By using a high-precision infrared temperature measurement array and strain field optical measurement system, the surface temperature distribution, equivalent strain field and strain rate field data of each forging process are collected in real time to form a multi-pass time-series process sensing dataset. The multi-pass time-series process sensing dataset is input into a pre-trained graph attention neural network. The network uses forging passes as nodes and thermal coupling relationships between passes as edges to propagate and aggregate the microstructure state of each pass across passes, and outputs the predicted values ​​of dislocation density, average grain size, and grain boundary energy density at the end of each pass. The output of the physical constraint sub-model and the prediction result of the graph attention neural network are subjected to residual correction to generate a micro-organism evolution trajectory that integrates physical priors and data observations. Based on the microstructure evolution trajectory, the grain boundary weakening index corresponding to each pass is calculated. The grain boundary weakening index is defined as the ratio of the grain boundary energy density of the current pass to the critical grain boundary fracture threshold. A multi-objective optimization function is constructed. The objective terms of the multi-objective optimization function include total forging energy consumption, final grain uniformity index and maximum grain boundary weakening index, and the constraint terms include equipment load limit, die life limit and final forging temperature window. An improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function to generate an optimal forging parameter sequence that satisfies microstructure stability. The optimal forging parameter sequence is sent to the forging execution control system, which drives the hydraulic servo mechanism to perform multiple forming operations in sequence.

2. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 1, characterized in that, The forging parameter sequence includes the initial forging temperature, final forging temperature, reduction, strain rate, and pass interval for each pass.

3. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 2, characterized in that, The material property data includes the alloy composition by mass percentage, initial grain size, initial dislocation density, activation energy parameter, and thermal conductivity temperature function; the thermodynamic boundary condition data includes ambient temperature, mold preheating temperature, thermal resistance coefficient of lubrication conditions, and heat transfer coefficient of cooling medium.

4. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 3, characterized in that, The high-precision infrared temperature measurement array consists of no fewer than 64 independent temperature measurement units, with a spatial resolution of 5 mm and a sampling frequency of 100 Hz; the strain field optical measurement system adopts digital image correlation technology, is equipped with a binocular high-speed camera, has a frame rate of 500 frames per second, and a strain measurement accuracy of 1‰.

5. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 4, characterized in that, The graph attention neural network contains three graph convolutional layers, each with 128-dimensional node features. The edge weights are calculated from the temperature gradient, cumulative strain increment, and time interval between adjacent traces. The attention coefficients are generated through a learnable query-key mechanism and are used to weight and aggregate the influence of historical traces on the microstate of the current trace.

6. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 5, characterized in that, The residual correction employs a weighted least squares method, with the weight matrix determined by the uncertainty covariance of the physical model in the high-temperature, high-strain region, and the corrected dislocation density. ,in Output for the physical model. For the graph attention network's predicted values, It is a diagonal weight matrix, and its elements take values ​​from 0.3 to 0.

7.

7. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 6, characterized in that, The critical grain boundary fracture threshold of the grain boundary weakening index is preset according to the alloy type. For nickel-based superalloys, the threshold is 1.2 joules per square meter; for titanium-aluminum alloys, the threshold is 0.8 joules per square meter; and for austenitic stainless steel, the threshold is 1.0 joules per square meter.

8. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 7, characterized in that, The multi-objective optimization function is defined as follows: ； in Total forging energy consumption, The final grain size standard deviation, The average grain size, Let be the grain boundary weakening index of the i-th pass. , , The normalized weighting coefficients are set to 0.4, 0.3, and 0.3 respectively; the upper limit of the equipment load is set to 90% of the rated pressure of the hydraulic press; the mold life limit is set to a cumulative reduction of no more than 5,000 tons for a single mold; and the final forging temperature window is set to 30°C above and 20°C below the alloy recrystallization termination temperature.

9. The method for optimizing multi-pass forging parameters of difficult-to-deform alloys based on machine learning according to claim 8, characterized in that, The improved non-dominated sorting genetic algorithm adopts an elite retention strategy, with a population size of 200, a crossover probability of 0.9, a mutation probability of 0.1, and a Gaussian perturbation for the mutation operation. The standard deviation is 5% of the parameter range, and a diversity maintenance mechanism based on crowding distance is introduced to ensure the uniformity of the Pareto front solution set.

Images