Optimization Design And Mechanical Properties Evaluation Of RAFM Steels And Low-activation High Entropy Alloys Based On Machine Learning

Reduced activation ferritic/martensitic(RAFM)steels are considered as the primary candidate structural materials for the fusion reactor blanket because of their good thermophysical properties,high resistance to neutron-irradiation and mass productivity.However,the application and research of RAFM steels are still confronted with two major challenges.Firstly,the obvious high-temperature softening effect of RAFM steels hinders their application at high temperatures(e.g.,above 550℃).In order to meet the requirements of higher service temperature of future commercial fusion reactors,it is necessary not only to further explore the potential of RAFM steels,but also to pay attention to the research and development of high entropy alloys(HEAs)with superior performance.Secondly,it is urgent to solve the key problem about how to accurately and efficiently evaluate the mechanical properties of RAFM steels under neutron irradiation.The traditional trial and error experiments,physical models,and theoretical calculation methods are expensive or inefficient,which are difficult to be employed to solve the above-mentioned key issues.The data-driven machine learning(ML)method provides a new powerful tool for the research and development of multicomponent materials.However,the current ML model is still difficult to realize the property-oriented design.In view of this,the research carried out a series of works using the ML method.The main research contents and results are as follows:1)The intelligent design model was developed to realize the property-oriented design of RAFM steels,which is used to design and prepare new RAFM steels with excellent high-temperature tensile properties.After systematically comparing 6 commonly used ML algorithms,the gradient boosting regression(GBR)algorithm was selected to construct the forward model.The corresponding coefficient of determination(R2)for predicted ultimate tensile strength(UTS)and total elongation(TE)are 0.98 and 0.89,respectively.Using artificial neural network regression(ANNR)algorithm,the reverse model was established to provide the possible combinations of compositions and heat treatments for given tensile properties.The intelligent design model,combining the forward model with the reverse model,was developed to design the compositions and heat treatment parameters for RAFM steels with the targeted tensile properties.The validity of the intelligent design model was verified by the experimental data of three RAFM steels reported in the relevant literatures.In order to further verify the actual performance of the design results of the intelligent design model and develop the high-temperature resistant RAFM steel,a new type of highstrength RAFM(HS-RAFM)steel was designed and prepared.In the test temperature range of 25-600℃,the UTS of the HS-RAFM steel is～200-400 MPa higher than the conventional RAFM steels while maintaining comparable TE.Therefore,this strategy is suitable for the property-oriented design of RAFM steels and can also be considered as a very promising approach to develop high-performance structural materials.2)By combining ML with calculation of phase diagram(CALPHAD)methods to build an integrated design system,the design of RAFM steels strengthened by MX precipitation was realized,which provides a new way to realize the structure and performance oriented design.Based on the datasets of the equilibrium volume fraction of different phases calculated by CALPHAD,the reliable microstructural prediction models were constructed using ML method.The microstructural prediction models using the gradient boosting classifier(GBC)algorithm achieve the accuracy of 98.8%and 95.9%to identify the presence of δ ferrite and coarsening phases in RAFM steels,respectively.The random forest regression(RFR)and support vector regression(SVR)algorithms were selected to construct the prediction models for the volume fraction of MX and M23C6 precipitations,and the corresponding R2 were 0.93 and 0.96,respectively.Then,four microstructural prediction models are coupled together to build an integrated design system in combination with the aforementioned intelligent design model.A new MX-precipitation strengthened RAFM(MX-RAFM)steel was designed according to the structure and performance requirements.Finally,through experimental verification,it is found that microstructure and tensile properties of MX-RAFM steel from experiments are in good agreement with the predicted results,which confirms the effectiveness of the integrated design system.3)The prediction models for neutron irradiation hardening and embrittlement based on ML were developed,taking the study of the ductile-brittle transition temperature(DBTT)of neutron-irradiated RAFM steels as an example.Based on the collected dataset of RAFM steels under neutron irradiation,the key features affecting DBTT of neutron-irradiated RAFM steels were identified by correlation screening and recursive elimination methods.Based on the selected key features,the prediction model for DBTT of neutron-irradiated RAFM steels was constructed by RFR algorithm,which shows a good prediction ability with R2 about 0.92.The prediction model for DBTT of un-irradiated RAFM steels was constructed as well.The prediction model for ductilebrittle transition temperature shift(ΔDBTT)of neutron-irradiated RAFM steels was constructed by combining the prediction models for DBTT before and after irradiation.By comparing the predicted ΔDBTT by the model with the related experimental data,R2 is 0.91 and the root mean square error(RMSE)is 14.5,indicating that this prediction model for ΔDBTT has high predictability and reliability.4)Based on ML method,a design strategy for simultaneously optimizing the structure and performance of HEAs was proposed,and the low-activation HEA Fe30Cr35V15W15Mn5 was designed and fabricated.Five different machine learning algorithms were used to screen the key features through correlation screening,recursive elimination,and exhaustive screening.The GBC and random forest classifier(RFC)algorithms and their corresponding key variables are used to build two classification models with accuracy of more than 90%to identify the phases of HE As.Using RFR algorithm and its corresponding key variables,a regression prediction model for the hardness of HEAs is constructed,and its R2 reaches 0.94.Using these models,a lowactivation HEA Fe30Cr35V15W15Mn5 was selected from about 285000 candidate materials and fabricated successfully in experiments.It has BCC phase structure,and the experimental value of hardness is 699.3±27.6 HV.This is consistent with the prediction results,indicating that our design strategy is suitable for the design of lowactivation HEAs.In summary,ML,CALPHAD,and experimental verification methods were combined to realize the optimization design of high-temperature tensile properties and prediction of irradiation properties of RAFM steels,as well as the design of lowactivation HEAs in this paper.It provides new ideas and methods for efficient research and development of candidate structural materials for fusion reactors in the future.