Search

CN-121999938-A - Physical knowledge machine learning high-entropy alloy phase prediction system and method based on semi-empirical parameters

CN121999938ACN 121999938 ACN121999938 ACN 121999938ACN-121999938-A

Abstract

The invention discloses a physical knowledge machine learning high-entropy alloy phase prediction system and method based on semi-empirical parameters, and belongs to the technical field of high-entropy alloy material design and machine learning intersection. The system takes semi-empirical parameter physical constraint as a core, integrates the physical interpretability of an empirical parameter method and the data driving advantage of machine learning through a three-level cooperative mechanism of 'feature multiplexing-independent prediction-Bayesian fusion', and solves the problems of narrow phase coverage, low multiphase prediction precision, black box defect and the like of the traditional method. The system can predict more than 10 high-entropy alloy phase types and a multiphase coexisting system, the single-phase prediction accuracy in 856-group multi-component high-entropy alloy test set reaches 100%, the multiphase coexisting system prediction accuracy reaches 77%, the phase formation physical mechanism explanation can be output, and a special accurate criterion for the B2 phase is provided, so that the high-entropy alloy design is promoted to be converted from a trial-and-error method to accurate prediction.

Inventors

  • CHEN XIZHANG
  • Jia Yanghao

Assignees

  • 温州大学

Dates

Publication Date
20260508
Application Date
20260128

Claims (10)

  1. 1. The physical knowledge machine learning high-entropy alloy phase prediction system based on the semi-experience parameters is characterized by comprising a semi-experience parameter calculation module, a physical knowledge feature processing module, an independent prediction module, a Bayesian fusion module and a result interpretation module which are sequentially and cooperatively connected; The semi-empirical parameter calculation module calculates 12 core semi-empirical parameters based on a P.Maetin semi-empirical model, wherein the core semi-empirical parameters comprise mixing enthalpy delta Hmix, omega parameters Omega, valence electron concentration VEC, atomic radius difference delta r and the like, and a physical basis is provided for a subsequent module; The physical knowledge feature processing module is based on 12 core semi-empirical parameters, expands to form 30 feature sets with definite physical significance, wherein 12 are original semi-empirical parameters, 19 are physical derivative features, and the physical derivative features comprise 、VEC×ΔR; The independent prediction module comprises a physical branch and a data branch, wherein the physical branch calls an alloycriterion function to output a phase result based on a semi-empirical criterion, the data branch trains XGBoost models based on 30 physical knowledge characteristics, outputs phase probability and a prediction result, and filters an abnormal sample through semi-empirical parameter rationality verification in the training process; The Bayesian fusion module quantifies the prediction uncertainty of the physical branch and the data branch, performs weighted fusion by combining the causal feature weight, and outputs the final phase prediction result; the result interpretation module outputs a physical knowledge interpretation report including key semi-empirical parameter contribution degrees, feature importance ordering and phase formation physical logic.
  2. 2. The system of claim 1, wherein the 30 features are categorized into four types by physical meaning, namely 10 physical basic parameters, 4 atomic size mismatch parameters, 6 thermodynamic parameters and 10 HEA specific phase stability parameters, all of which are calculated by an analytical formula, and have no statistical features of no physical meaning.
  3. 3. The system of claim 1, wherein the Bayesian fusion module comprises a causal junction construction module, a causal feature weight extraction module, an uncertainty quantization module, and a Bayesian fusion operator module; the causal junction construction mold module constructs a causal path of semi-empirical parameters, a prediction method and phase formation; the causal feature weight extraction sub-module outputs causal importance scores of semi-empirical parameters based on causal forest training; The uncertainty quantization sub-module calculates the criterion boundary distance of the physical branch and the prediction probability entropy of the data branch, and converts the criterion boundary distance and the prediction probability entropy into weight coefficients; And the Bayesian fusion calculation submodule combines causal weight and uncertainty, calculates posterior probability based on Bayesian theorem and outputs combined decision.
  4. 4. The system of claim 1, further comprising B2-phase specific prediction criteria, wherein the criteria are constructed based on 5 core features, including η (threshold 0.08-0.13), Ω (threshold 1-3.2), λ (threshold 0.25-0.55), D_k (threshold-5.7 to-1.2), e/a (threshold 1.8-2.2), and constructing a new semi-empirical parameter γG/(Δ Hmix Ω) (threshold 0.8-2.18).
  5. 5. A physical knowledge machine learning high-entropy alloy phase prediction method based on semi-empirical parameters is characterized by comprising the following steps: Step1, obtaining characteristic data of a high-entropy alloy phase through a high-entropy alloy phase pre-measurement training set; step2, calculating 12 core semi-empirical parameters based on basic physicochemical properties of the high-entropy alloy components through a semi-empirical parameter calculation module; Step3, performing physical derivative expansion on 12 core semi-experience parameters through a physical knowledge feature processing module to form 30 physical knowledge feature sets; Step4, respectively outputting phase prediction results through a physical branch and a data branch of the independent prediction module, wherein the physical branch is decided based on semi-empirical criteria, and the data branch outputs phase probability through a trained machine learning model; Step5, constructing a causal structure through a Bayesian fusion module, extracting causal feature weights, quantifying two branch prediction uncertainty, and carrying out weighted fusion to obtain a final phase prediction result; Step6, outputting a physical knowledge interpretation report, and defining key parameters and physical mechanisms of phase formation.
  6. 6. The method of claim 5, wherein the Step1 core semi-empirical parameters are calculated by using analysis methods such as a mixing rule, a standard deviation formula, a Miedema model, an ideal mixed entropy formula, etc., so as to ensure that the parameters are physically clear in meaning and can be calculated automatically.
  7. 7. The method of claim 5, wherein the model training process of the data branch in Step3 comprises semi-empirical parameter rationality checking, filtering samples beyond a physical threshold, and strengthening physical constraints of the model.
  8. 8. The method of claim 5, wherein the Bayesian fusion in Step4 satisfies three causal hypotheses, SUTVA hypothesis, no confounding hypothesis and enthusiasm hypothesis, ensuring physical consistency and causal reliability of the fusion result.
  9. 9. The method of claim 5, wherein the mathematical basis for the fusion process in Step4 is the Bayesian theorem: where P (Phase) is the prior probability of the semi-empirical parametric method, ) Likelihood probabilities for a machine learning model.
  10. 10. The method of claim 5, further comprising a B2 phase prediction step, wherein the B2 phase prediction step is used for realizing accurate identification of the B2 phase by combining a Bayesian fusion result based on the constructed B2 phase exclusive criterion, and the misjudgment rate is less than or equal to 8.59%.

Description

Physical knowledge machine learning high-entropy alloy phase prediction system and method based on semi-empirical parameters Technical Field The invention relates to the technical field of high-entropy alloy material design and machine learning intersection, in particular to a system and a method for predicting a high-entropy alloy phase by physical knowledge machine learning based on semi-empirical parameters. Background The high-entropy alloy is a novel metal material containing at least 5 main elements (the atomic ratio is 5% -35%), has important application prospects in the fields of aerospace, energy equipment and the like due to excellent high-temperature strength, corrosion resistance and other performances, and the phase structure directly determines the macroscopic performance of the material, so that accurate prediction of the phase structure is a key for shortening the research and development period and reducing the cost. The existing high-entropy alloy phase prediction method is mainly divided into two types, namely a traditional experience parameter method and a single machine learning method. The traditional empirical parameter method is based on a definite physical rule (such as Omega-Delta criterion and valence electron concentration criterion), but has poor adaptability to a complex system with the component number of more than 5, can only judge a few typical single phases (such as FCC/BCC), cannot identify multiphase coexistence or special intermetallic compound phases (such as Sigma and B2 phases), and has no code landing. Although the single machine learning method can process high-dimensional data, the method has the obvious defects that the phase prediction range is narrow, the existing research focuses on 2-3 typical phases, the accuracy rate is reduced by 20-30% when the types of the predicted phases exceed 3 types, similar phases or composite phases cannot be finely distinguished, the accurate requirement of material performance regulation and control is difficult to meet, the model is a black box, the physical mechanism formed by the phases cannot be explained, the causal relationship between the phase structure and the physical parameters is difficult to establish, and theoretical guidance cannot be provided for material design. Although physical informed machine learning is pointed out to improve prediction generalization and interpretability, the prior study has not combined semi-empirical parameters with the framework, and cannot realize wide-phase domain and detailed high-entropy alloy phase prediction. Therefore, it is needed to construct a fusion framework of "wide-phase domain coverage + careful distinction + physical interpretability", and to solve the deficiencies of the prior art, a high-entropy alloy method and software are proposed herein. Disclosure of Invention The method aims to solve three major core problems in the existing high-entropy alloy phase prediction method, namely 1) the phase coverage is narrow, more than 10 phase types and a multiphase coexisting system cannot be covered, 2) the multiphase prediction accuracy is low, the recognition capability of similar phases, composite phases and special intermetallic compound phases is insufficient, 3) the physical interpretability is poor, the defect of black box exists in traditional machine learning, and the causal relation between phase formation and physical parameters cannot be established. In order to solve the technical problems, the invention designs a high-entropy alloy phase prediction method by machine learning The model can identify the internal relation of the data set, a prediction model is established to rapidly predict the unseen sample, and the semi-empirical parameter results are fused through Bayes. In order to achieve the above purpose, the invention provides a physical knowledge machine learning high-entropy alloy phase prediction system based on semi-empirical parameters, which has the following core technical scheme: the system takes semi-empirical parameter physical constraint as a core and comprises 5 modules which work cooperatively: and the semi-empirical parameter calculation module is used for calculating 12 core semi-empirical parameters through an analytic formula based on a P.Maetin semi-empirical model, wherein the four dimensions of physical basis, atomic size mismatch, thermodynamics and HEA exclusive phase stability are covered, and a physical basis is provided for the system. And the physical knowledge characteristic processing module is used for expanding 30 characteristic sets with definite physical meanings based on 12 core semi-empirical parameters, wherein 18 physical derivative characteristics are obtained through original parameter combination, so that the introduction of statistical characteristics without physical meanings is avoided, and the interpretability of the model is ensured. The independent prediction module comprises a physical branch and a data branch, wher