CN-121999915-A - Reverse design method, device, equipment and storage medium for high-hardness alloy composition

CN121999915ACN 121999915 ACN121999915 ACN 121999915ACN-121999915-A

Abstract

The invention provides a high-hardness alloy component reverse design method, a device, equipment and a storage medium, which are obtained by acquiring alloy component data and alloy microstructure characteristic data, inputting an alloy hardness prediction model and outputting a Vickers hardness prediction value, inputting the prediction value into the high-hardness alloy component reverse design model and outputting a component scheme, wherein the alloy hardness prediction model is obtained by training a hybrid neural network based on alloy components, alloy microstructure characteristics and a sample data set constructed by actually measured hardness, the high-hardness alloy component reverse design model is obtained by training based on a deep reinforcement learning frame, the deep reinforcement learning frame comprises at least one intelligent body, and the intelligent body learns in an alloy component design environment to generate the high-hardness alloy component scheme.

Inventors

LIU QING
ZHANG SHUNLI
SHANG WENJING

Assignees

晋中学院

Dates

Publication Date: 20260508
Application Date: 20251229

Claims (10)

1. A method of reverse engineering a high hardness alloy composition comprising: acquiring alloy component data and alloy microstructure characteristic data, inputting the alloy component data and the alloy microstructure characteristic data into an alloy hardness prediction model, and outputting a Vickers hardness prediction value; inputting the Vickers hardness predicted value into a reverse design model of the high-hardness alloy component, and outputting a scheme of the high-hardness alloy component; The alloy hardness prediction model is obtained by training a mixed neural network based on a sample data set constructed by alloy components, alloy microstructure characteristics and actually measured hardness, and the mixed neural network is configured to respectively process the alloy component data and the alloy microstructure characteristic data and perform characteristic fusion to predict a Vickers hardness value; The high-hardness alloy component reverse design model is obtained by training based on a deep reinforcement learning framework, the deep reinforcement learning framework comprises at least one intelligent body, the intelligent body learns in an alloy component design environment to generate a high-hardness alloy component scheme, wherein the state of the alloy component design environment is an alloy component vector, actions of the intelligent body are adjustment of element content in the alloy component vector, rewards are calculated through the actions to learn the high-hardness alloy component scheme, and the rewards are calculated based on a Vickers hardness predicted value represented by the current state.
2. The method of claim 1, wherein the hybrid neural network comprises a first branch, a second branch, a feature fusion layer, and an output layer; the first branch is a convolutional neural network and is used for processing the alloy microstructure characteristic data; The second branch is a deep neural network and is used for processing the alloy composition data; The feature fusion layer is used for fusing the output features of the first branch and the second branch; and the output layer is used for predicting the Vickers hardness predicted value according to the fusion characteristics output by the characteristic fusion layer.
3. The method of reverse engineering a high hardness alloy composition according to claim 2, wherein the feature fusion layer comprises: the characteristic splicing unit is used for splicing the output characteristics of the first branch and the second branch; And the residual adding unit is used for processing the spliced features through at least two parallel linear transformation paths and adding the feature vectors output by the paths.
4. The method of claim 1, wherein the method of obtaining the microstructure characterization data of the alloy comprises: extracting high-dimensional image features from the alloy microstructure image by using a pre-trained self-supervision visual transducer model; And performing principal component analysis and dimension reduction on the high-dimensional image characteristics to obtain the alloy microstructure characteristics.
5. The method of claim 4, wherein the alloy microstructure image comprises a raw alloy microstructure image and a composite alloy microstructure image, the composite alloy microstructure image acquisition method comprising: And acquiring an original alloy microstructure image, inputting the original alloy microstructure image into a generated countermeasure network model, and generating a synthetic alloy microstructure image corresponding to the condition by using the alloy composition data as the condition by the generated countermeasure network model.
6. The method of claim 1, wherein the reward is calculated based on a predicted vickers hardness value for the current state characterization, comprising: constructing a substitute prediction model; Judging whether the current calculation condition meets a preset model switching condition or not in the learning process of the intelligent agent; if not, calling the replacement prediction model to calculate rewards based on the Vickers hardness prediction value output by the current state, and if so, calling the alloy hardness prediction model to calculate rewards based on the Vickers hardness prediction value output by the current state; The preset model switching condition comprises at least one of first exploration of the alloy component vector, confidence of the alternative prediction model on a current prediction result being lower than a set threshold and accumulation of new component quantity predicted by the alternative prediction model reaching an update threshold.
7. The method of claim 1 or 6, wherein the bonus function used to calculate the bonus is configured to: outputting a negative reward when the predicted vickers hardness value is below a first threshold; outputting a linear positive prize proportional to the vickers hardness value when the vickers hardness predicted value is between the first and second thresholds; Outputting a positive prize that is super-linear with the vickers hardness value when the predicted vickers hardness value is higher than the second threshold; wherein the second threshold is higher than the first threshold.
8. A high hardness alloy composition reverse engineering apparatus, comprising: The hardness prediction module is used for acquiring alloy component data and alloy microstructure characteristic data, inputting the alloy component data and the alloy microstructure characteristic data into an alloy hardness prediction model, and outputting a Vickers hardness prediction value; The component optimization module is used for inputting the Vickers hardness predicted value into a high-hardness alloy component reverse design model and outputting a high-hardness alloy component scheme; The alloy hardness prediction model is obtained by training a mixed neural network based on a sample data set constructed by alloy components, alloy microstructure characteristics and actually measured hardness, and the mixed neural network is configured to respectively process the alloy component data and the alloy microstructure characteristic data and perform characteristic fusion to predict a Vickers hardness value; The high-hardness alloy component reverse design model is obtained by training based on a deep reinforcement learning framework, the deep reinforcement learning framework comprises at least one intelligent body, the intelligent body learns in an alloy component design environment to generate a high-hardness alloy component scheme, wherein the state of the alloy component design environment is an alloy component vector, actions of the intelligent body are adjustment of element content in the alloy component vector, rewards are calculated through the actions to learn the high-hardness alloy component scheme, and the rewards are calculated based on a Vickers hardness predicted value represented by the current state.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the high hardness alloy composition reverse engineering method of any one of claims 1 to 7 when the program is executed.
10. A non-transitory readable storage medium having stored thereon a computer program, which when executed by a processor, implements the high hardness alloy composition reverse engineering method according to any one of claims 1 to 7.

Description

Reverse design method, device, equipment and storage medium for high-hardness alloy composition Technical Field The invention relates to the technical field of alloy component design, in particular to a reverse design method, device and equipment for high-hardness alloy components and a storage medium. Background The high-hardness alloy is a core structural material in the key fields of high-temperature parts of aerospace engines, high-end cutting tools, heavy machinery wear-resistant parts and the like due to the excellent high strength, high wear resistance and high temperature resistance. The Vickers hardness is a key performance index for determining the deformation resistance, service life and extreme environmental adaptability of the material. Development of high hardness alloys relies on the traditional mode of "empirical drive-experimental trial and error". The research and development personnel need to preset the component proportion based on the past knowledge and experience, and verify and adjust the complex experimental procedures such as smelting, forging, heat treatment, microcosmic characterization, hardness test and the like through repeated development. The mode not only leads to long research and development period and high cost of manpower and material resources, but also is limited by subjective experience of experts and limited experimental scale, and is difficult to comprehensively and systematically explore high-dimensional and continuous component space, so that a potential optimal component scheme is easily omitted, and the design efficiency is low. To overcome the limitations of the conventional trial-and-error method, in the related art, a mapping model between alloy components and performances is built based on a data-driven machine learning method so as to accelerate performance prediction. However, most existing models attempt to build a direct black box mapping of "components" to "performance", and it is difficult to adequately capture complex nonlinear relationships and physical mechanisms of components that determine performance by affecting microstructure, resulting in limited prediction accuracy, poor interpretability of the model, and weak extrapolation capability in component space. These models, though, are capable of extracting high-dimensional features containing physical information from metallographic images. However, the technology of how to extract key microstructure features strongly correlated with performance from these high-dimensional features is still not mature. The traditional optimization algorithm (such as genetic algorithm and Bayesian optimization) is easy to sink into local optimum in a high-dimensional component space, the convergence speed is low, and after a performance prediction model is obtained, how to reversely find the optimum component according to target performance (such as high hardness) is still a challenging optimization problem. Disclosure of Invention The invention provides a reverse design method, device, equipment and storage medium for high-hardness alloy components, which are used for solving the defects that the existing high-hardness alloy component design method ignores the bridge effect of a microstructure, so that the prediction accuracy of a model is limited, the interpretation is poor, the exploration efficiency in a high-dimensional component space is low, and the reliability in actual design is insufficient. The invention provides a reverse design method of high-hardness alloy components, which comprises the following steps: acquiring alloy component data and alloy microstructure characteristic data, inputting the alloy component data and the alloy microstructure characteristic data into an alloy hardness prediction model, and outputting a Vickers hardness prediction value; inputting the Vickers hardness predicted value into a reverse design model of the high-hardness alloy component, and outputting a scheme of the high-hardness alloy component; The alloy hardness prediction model is obtained by training a mixed neural network based on a sample data set constructed by alloy components, alloy microstructure characteristics and actually measured hardness, and the mixed neural network is configured to respectively process the alloy component data and the alloy microstructure characteristic data and perform characteristic fusion to predict a Vickers hardness value; The high-hardness alloy component reverse design model is obtained by training based on a deep reinforcement learning framework, the deep reinforcement learning framework comprises at least one intelligent body, the intelligent body learns in an alloy component design environment to generate a high-hardness alloy component scheme, wherein the state of the alloy component design environment is an alloy component vector, actions of the intelligent body are adjustment of element content in the alloy component vector, rewards are calculated through the actions to learn