CN-121999924-A - High-strength ductile cast iron design method integrating physical information neural network

CN121999924ACN 121999924 ACN121999924 ACN 121999924ACN-121999924-A

Abstract

The invention discloses a high-strength ductile cast iron design method integrating a physical information neural network, and belongs to the technical field of material design and intelligent manufacturing. Firstly, a nodular cast iron data set covering chemical components, microstructures and mechanical properties is systematically constructed, preprocessing and feature dimension reduction are carried out, a deep neural network regression model is further constructed, the physical rule of the nodular cast iron is used as a constraint condition, a physical information neural network is formed by embedding a composite loss function formed by data loss and physical loss weighting into a training process, the trained physical information neural network is used as an adaptability predictor of a genetic algorithm, multi-objective optimization and optimization are carried out on the chemical components, and an optimal parameter combination is reversely deduced. The invention realizes the deep fusion of data driving and physical mechanism, solves the problems of long period and poor extrapolation of a pure data model of the traditional trial-and-error method, and provides an innovative solution for the rapid and accurate design of high-performance spheroidal graphite cast iron.

Inventors

LI ZHENHUA
LIU YU
HE YUANHUAI
WEI HE
TU WENWEN
Jiang Dehuan

Assignees

昆明理工大学

Dates

Publication Date: 20260508
Application Date: 20260108

Claims (9)

1. The high-strength ductile cast iron design method integrating the physical information neural network is characterized by comprising the following steps of: s1, constructing a spheroidal graphite cast iron material performance data set; s2, performing data preprocessing operation based on the data set constructed in the S1; s3, executing characteristic engineering and data set division operation on the preprocessed data set; S4, constructing a deep neural network regression model; S5, based on the constructed deep neural network regression model, introducing a physical constraint mechanism and a composite loss function design; S6, training the network constructed in the S5; S7, model evaluation and verification; s8, performing multi-objective mechanical property optimization operation on the model after evaluation and verification, and completing method design.
2. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 1, which is characterized in that: the ductile iron material performance data set comprises chemical components, microstructure and mechanical properties; The chemical components comprise elements composing the spheroidal graphite cast iron and corresponding mass percentages thereof, wherein 17 elements comprise C, si, mn, S, P, cu, mo, ca, ba, sb, ni, sn, cr, al, ti, mg and RE; microstructure includes spheroidization rate, graphite size and pearlite content; the mechanical properties include tensile strength and elongation.
3. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 2, which is characterized in that: The data preprocessing operation includes: based on the ductile iron material performance data set, eliminating records with key feature missing or obvious abnormality, eliminating features with feature missing more than 90%, and filling missing values in the residual data in an arithmetic average mode.
4. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 3, wherein the step S3 comprises the following steps: S3.1, taking 13 chemical component parameters and 3 microstructure characteristics obtained after pretreatment as model input variables and taking mechanical properties as output variables; S3.2, performing feature dimension reduction processing on the input features to reduce model complexity and improve calculation efficiency; the feature dimension reduction processing method comprises the steps of performing correlation analysis on 16 input features and 2 output performances by using a Pearson correlation coefficient method, and screening out 13 features with the most obvious influence on the performances as final input; performing Z-score standardization treatment on the screened data; And S3.3, carrying out standardized processing on the data subjected to dimension reduction, and randomly dividing the data into a training set, a verification set and a test set according to a preset proportion.
5. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 1, which is characterized in that: The deep neural network regression model comprises an input layer of 13 neurons and four hidden layers, wherein the number of the neurons is 480, 240, 120 and 120 respectively, and leakyReLU activation functions and output layers of 2 neurons are adopted.
6. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 1, which is characterized in that: the physical constraint mechanism comprises 0.35 times microstructure constraint, 0.08 times boundary constraint, 0.01 times performance trade-off constraint and 0.005 times strength-plasticity contradiction constraint; the composite loss function is expressed as total loss = data loss + lambda physical loss, wherein the data loss is the mean square error between the predicted value and the true value; lambda is a super parameter, 0.1 is taken to balance the weight of the two losses.
7. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 1, which is characterized in that: The S6, the training mode of the network constructed in the S5 is that based on a training set, a gradient descent algorithm is adopted to iteratively train the neural network embedded with physical constraints, model parameters are optimized, and an Adam optimizer is adopted to train until model loss converges.
8. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 1, which is characterized in that: The evaluation and verification indexes comprise a decision coefficient and a mean square error, and the comparison objects comprise a random forest model, a support vector regression model and XGBoost.
9. The method for designing the high-strength ductile iron fused with the physical information neural network according to claim 1, which is characterized in that: The multi-objective mechanical property optimization method comprises the steps of taking a trained physical information neural network as an fitness function predictor, starting a genetic algorithm to perform multi-objective optimization, setting an optimization target to be that tensile strength Rm is more than or equal to 900 MPa and elongation A is more than or equal to 5%, setting the population size of the genetic algorithm to be 100, iterating 100 generations, setting a basic variation rate to be 0.15 by using a mixed crossing strategy, setting a tournament parameter to be 3, and performing global optimization in a chemical component space by selecting, crossing and variation operations; Meanwhile, in order to take economic benefit into consideration, only Cu and Mo are selected as main alloy elements, and the content of the rest alloy elements Ni, cr, ti, sb is limited to 0.

Description

High-strength ductile cast iron design method integrating physical information neural network Technical Field The invention relates to the technical field of nodular cast iron material design, in particular to a high-strength and toughness nodular cast iron design method integrating a physical information neural network. Background In recent years, with the development of the equipment manufacturing industry in China to high-end and intelligent directions, the performance requirements on key basic parts are increasingly improved. The performance optimization of spheroidal graphite cast iron as an important structural material is directly related to the reliability and service life of important equipment. Particularly in the fields of energy equipment, heavy machinery and the like, the trend of enlargement and complexity of cast iron components puts higher demands on the comprehensive performance of materials. Spheroidal graphite cast iron has become one of the preferred materials for manufacturing large complex structural members due to its good casting performance and mechanical properties. However, with the increase of the size of the components and the deterioration of the service environment, how to obtain the ductile iron material with excellent performance through composition and process design becomes a key technical bottleneck for restricting the development of the industry. At present, the material design of spheroidal graphite cast iron mainly faces the following technical problems: 1. The existing design method mostly depends on limited experimental data and empirical formulas, and deep rules contained in historical data cannot be fully utilized. Particularly under the condition of small samples, the traditional method is difficult to establish an accurate composition-organization-performance relation model, and the reliability of design results is insufficient. 2. The multi-objective optimization is difficult, and in practical engineering application, the nodular cast iron needs to meet the requirements of a plurality of performance indexes such as tensile strength, elongation and the like. The performance indexes often have mutually restricted relation, and the traditional method is difficult to realize the collaborative optimization of a plurality of performance indexes. 3. The physical mechanism is missing, and although the pure data-driven model can realize performance prediction to a certain extent, the model prediction result often lacks interpretability due to lack of guidance of the physical mechanism, and the prediction reliability in a data sparse area is poor. 4. The design efficiency is low, the traditional 'trial-and-error' design mode needs to be subjected to repeated experiments for many times, the research and development period is long, the cost is high, and the global optimal solution is difficult to obtain. Especially in the face of complex multicomponent systems, the design efficiency of conventional approaches is more difficult to meet the requirements of engineering applications. Disclosure of Invention In order to solve the technical problems, the invention provides a high-strength ductile cast iron design method integrating a physical information neural network. In order to realize the technology, the steps are as follows: s1, constructing a spheroidal graphite cast iron material performance data set; The ductile iron material performance data set comprises chemical components, microstructure and mechanical properties; The chemical components comprise elements composing the spheroidal graphite cast iron and corresponding mass percentages thereof, wherein 17 elements comprise C, si, mn, S, P, cu, mo, ca, ba, sb, ni, sn, cr, al, ti, mg and RE; microstructure includes spheroidization rate, graphite size and pearlite content; the mechanical properties include tensile strength and elongation. S2, performing data preprocessing operation based on the data set constructed in the S1; The data preprocessing operation comprises the following steps: based on the ductile iron material performance data set, eliminating records with key feature missing or obvious abnormality, eliminating features with feature missing more than 90%, and filling missing values in the residual data in an arithmetic average mode. S3, executing characteristic engineering and data set division operation on the preprocessed data set, wherein the steps comprise: S3.1, taking 13 chemical component parameters and 3 microstructure characteristics obtained after pretreatment as model input variables and taking mechanical properties as output variables; S3.2, performing feature dimension reduction processing on the input features to reduce model complexity and improve calculation efficiency; the feature dimension reduction processing method comprises the steps of performing correlation analysis on 16 input features and 2 output performances by using a Pearson correlation coefficient method, and screening out 13 feat