Optimization Design Of High Strength And Toughness FeCrAl Stainless Steel Based On Machine Learning Assistance

FeCrAl stainless steel can be applied to new nuclear reactors of various reactor types because of its good radiation resistance and high temperature thermal stability.In particular,FeCrAl ODS stainless steel prepared by doping Y₂O₃nano oxide particles into FeCrAl stainless steel has become an important candidate material for fuel rod cladding of the advanced Generation-IV nuclear reactors due to its high service temperature(>700?).At present,the traditional material design mainly depends on the empirical method or trial and error method,which often results in a waste of time and experimental cost.Therefore,a new alloy design idea is urgently needed to carry out to design alloy quickly and accurately.Machine learning(ML)can offer great assistance to alloy design,microstructure optimization,performance optimization,etc.,by discovering the correlation between user-defined characteristics of a material and its properties.As long as sufficient training data is collected and a reasonable prediction model is constructed,the best composition,microstructure,and performance of the material can be predicted theoretically,which greatly saves the manpower and physical resources,then accelerates materials innovation.Therefore,this paper used machine learning to quickly screen and design the matrix composition of FeCrAl stainless steel,systematically evaluates the accuracy of each model(BP neural network and RBF neural network)in designing stainless steel,optimizes the above machine learning model by genetic algorithm,and systematically studies the influence of each feature on machine learning performance.Thus,the number of features is minimized and the composition of the final FeCrAl stainless steel is determined by using the established neural network optimization model.In order to obtain high density,large volume and high number density nano oxide dispersed phase samples,the FeCrAl stainless steel was prepared by synchronous hot isostatic pressing.In order to quickly obtain samples with multiple deformations,wedge hot rolling was used for high flux preparation.Then,the effect of hot rolling deformation on the microstructure and mechanical properties of FeCrAl stainless steel with different components was systematically studied,and its role in the design of FeCrAl stainless steel was evaluated combined with the prediction results of machine learning.The main conclusions are as follows:(1)Different machine learning models have different accuracy in processing strength and elongation data sets.Among them,RBF neural network has good strength prediction accuracy,while the elongation prediction accuracy is relatively poor;The prediction accuracy of BP neural network strength model and elongation model is lower than that of RBF neural network.In addition,genetic algorithm can significantly optimize the above two machine learning models for the genetic algorithm of neural network,so as to obtain excellent prediction accuracy.Among them,the BP neural network optimized by genetic algorithm has higher prediction accuracy.Therefore,it is selected as the machine learning model of this experiment,and the alloy elements and test temperature are selected as the features.Finally,the composition of two FeCrAl stainless steels was determined,which were Fe-12Cr-4.5Al-2.0W-0.3Y₂O₃and Fe-12Cr-4.5Al-2.0W-0.3Y₂O₃-Ti;(2)After hot isostatic pressing,both samples show a ferrite matrix with bimodal structure and form coarse M₂₃C₆particles.There are also high-density nano Y₂O₃particles in the matrix.This Y₂O₃particle has cubic structure and obvious lattice mismatch with the ferrite matrix.Some Y₂O₃particles combine with Al to form coarse Y-Al-O composite oxide.It is worth noting that the addition of Ti preferentially forms tic with the carbon in the tissue,but does not form Y-Ti-O reported in other literatures,but its nano oxide particles still have high number density.(3)After hot rolling,the medium axial grains of the samples gradually change into fibrous grains,and the coarse carbides are refined and evenly distributed at the grain boundary.Many subgrains are formed in the grain due to continuous dynamic recrystallization.With the increase of hot rolling reduction,the proportion of sub grain size and deformed grain increases,resulting in the increase of dislocation density and elastic energy stored in ferrite matrix.At the same time,Y₂O₃particles and Y-Al-O particles have no significant coarsening and have good thermal stability.(4)After hot isostatic pressing,the tensile strength and elongation of the two samples measured at room temperature and 700?are close to the predicted values of machine learning,indicating that machine learning calculation can effectively guide the composition design of stainless steel.At the same time,the room temperature and high temperature mechanical properties(strength and plasticity)of the two samples after hot rolling are significantly improved.In conclusion,this study used machine learning combined with high-throughput alloy design to develop a new ODS FeCrAl alloy with excellent performance.The research method is also conducive to provide new solutions and methodological support for the research and development of other new structural alloys.