| New material design is the base of economy development,while the traditional material research and development usually need 20 to 30 years.Materials Genome Project aims to combine high-throughput calculations,high-throughput experiments and material big data closely,and achieve to cut down the time and expense by half.With the continuous development of artificial intelligence and data science,data-driven material design is considered as the fourth paradigm of materials research and development,and Materials Information Science has become a hotspot in the interdisciplinary research of computer and material science currently.In this paper,we used Sure Independent Screening and Sparsifying Operator(SISSO),multi-modal fusion machine learning etc.to forward predict the fatigue strength and charpy impact of steel,and used particle swarm optimization and multi-objective particle swarm optimization etc.to reverse design high-strength and high-toughness steels.The main tasks and innovations of this paper include:1.Material properties prediction based on SISSO algorithm:The relationship between material composition,process,structure and material attributes is very complex.While the traditional machine learning models are mostly like a "black box",which not able to tell the relationship between prediction targets and features with an analytical expression,and lack interpretability for prediction results.In this paper,we adopted the Sure Independent Screening and Sparsifying Operator(SISSO)algorithm,improved the combination explosion problem caused by feature combination of the algorithm and expanded the searched space of features.Compared with the traditional machine learning algorithms like Support Vector Regression(SVR),Decision Tree(DT),Gradient Boosting Decision Tree(GBDT),etc.,SISSO model can acquire an interpretable and concise expression while predicting the material properties,with the accuracy of 0.96 which is comparable closed to the "black box" models.Furthermore,for the physical interpretability of model,the expression formula of the best SISSO descriptors is consistent with the previous laboratory test rules,(including the positive or negative correlation between fatigue strength and quenching temperature,tempering temperature,diffusion temperature,carbon content,chromium content,nickel content etc.)which shows the reliable physical meaning and interpretability of SISSO.In terms of model stability,10-flod cross validation on fatigue strength dataset was performed in this paper.In the optimal descriptors of 10 flods,it was found six times same as the descriptors in the whole-data formula expression trained by SISSO 2D model,which indicating good stability of SISSO.2.Material properties prediction based on Multi-modal fusion:Dataset with multi-modal has strong expression ability.Combined with the rich semantic information contained in the periodic table of the elements,the material composition data was mapped into images through visual means and the atomic-scale micro-features of specific material properties(such as relative atomic mass,valence electron,electronic arrangement,etc.)were supplemented through modal enhancement during the process of image mapping.Afterwards,the multi-modal fusion and model training with the pre-processed data and images were performed to improve the accuracy of material property prediction.Experimental results show that based on the multi-modal fusion model,the prediction accuracy rate of fatigue strength on the steel fatigue strength dataset is maintained at about 0.96,while the prediction accuracy rate of charpy impact value has improved compared with single-modal models(such as Neural Network,Support Vector Regression,Decision Trees and Gradient Boosting Decision Tree),which has increased to 0.97.3.Material Reverse Design and Multi-objective optimization:After establishing the prediction model of material properties,the value of given input can be forward predicted.While in the development of new materials which has given the single or multiple target properties,it is more important to accurately recommend the next "experimental point",that is,the Material Reverse Design.Experimental results show as below:(1)SISSO model is used to reduce the feature dimension,the feature expression which obtained by model training is used to guide the material attributes reverse design.For the fatigue strength dataset of steel materials,the search efficiency of SVR,DT and GBDT based on particle swarm optimization(which searched on the six-dimensional subspace selected by SISSO)has been greatly improved compared with the grid search on the three main feature subspace,from more than 10 minutes to 2 minutes.In addition,the efficiency of particle swarm search based on SISSO model is even higher than that based on other machine learning models such as SVR,DT and GBDT,and the search can be completed in 0.01 second.(2)The feature expression formula obtained from the training of SISSO was analyzed to establish the analytical model for reverse design,and the relationship model of the main related characteristics to fatigue strength of the steel materials was displayed visually.Due to the extrapolation ability of the formula,the reverse design result based on SISSO model has made a breakthrough in the original metal fatigue strength dataset.The designed value of metal fatigue strength is higher than that in the original dataset and other machine learning models.Moreover,with this analytical formula,the relationship model between metal fatigue strength and main characteristics(C,Cr and Ni)was visually displayed.With the values of other features fixed,the fatigue strength is increased as the content of C,Cr or Ni increased,which is consistent with the rule of material science in low-carbon steel.(3)Aiming at multiple target attributes(fatigue strength and charpy impact value),the multi-objective particle swarm algorithm was used to perform multi-objective optimization.Through multi-objective particle swarm optimization,the process parameters(including heat treatment temperature,time and cooling rate)and composition content are adjusted within the scale of the experiment standard,the multiple target attributes as fatigue strength and charpy impact value are optimized.The quality of steel has improved with higher fatigue strength and higher charpy impact value,which realized the reverse design of high fatigue strength(which is 1190 Mpa)and higher chary impact(which is 221 J/cm2)steel.With three times improvement of searching efficiency,the designed result of multi-objective particle swarm algorithm has reached the approximate values of several single-objective optimizations that optimized separately,realizing the improvement of steel quality.The design schema also has significance for guiding the design of alloys with high strength and high impact toughness.Finally,the data-driven material performance forward prediction methods based on SISSO were integrated into the Materials Genome Engineering Web Application Platform,which using NodeJs.Flask and other technologies to build the front and back ends of the service.Users submit computing tasks with the Web front-end interaction,including:training data,computing resource requirements,and parameters of different machine learning models,etc.,and use XML format language to interact with rear-end services.Rear-end services perform the functions of computing tasks,real-time task status feedback,computing results storage and query,etc.. |
PDF Full Text Request