Machine Learning-based Alloy Design And Property Optimization Of Novel Fe-based Metallic Glasses

Fe-based metallic glasses(MGs),which exhibit outstanding soft magnetic properties,high strength and hardness,high elastic modulus and wear resistance,have great potential for application in high-tech fields such as sensors,transformers,wireless charging,5G devices and coatings.However,the development and application of Fe-based MGs are still facing two major challenges.Firstly,the interplay between the factors related to soft magnetic properties and thermal stability of these multicomponent Fe-based MGs are complex.Overcoming the trade-off between the saturation flux density(Bs)and the onset crystallization temperature(Tx)has been a long-term topic in the field.Secondly,the embrittlement induced by annealing is one of the bottlenecks for the widespread application of Febased MGs in the electronic field.Traditional "try and error" experimental methods are expensive and inefficient,which are hard to be employed to developed new Febased MGs with further improved properties.In this thesis,therefore,we aim to solve the above-mentioned key issues of Fe-based MGs.Machine learning methods were applied to predict and guide the development of Fe-based MGs with prominent overall performance.On top of that,effects of composition and other characteristic parameters on the soft magnetic properties,thermal stability,crystallization behavior and annealing-induced embrittlement of Fe-based MGs are then systematically analyzed.The main content and conclusions of the thesis are as follows:(1)We built an eXtreme Gradient Boosting(XGBoost)machine-learning(ML)model for developing advanced Fe-based MGs with a decent combination of Bs and thermal stability.Also,we carefully reveled the key parameters that play a dominant role in determining the Bs and Tx.Based on the dataset of Bs and Tx for 252 Febased MGs,the XGBoost ML model was trained and optimized to achieve high accuracy.The corresponding correlation coefficient(R2)for prediction Bs and Tx is 0.934 and 0.947,respectively.The feature importance ranking was obtained by the ML model,and it was found that the valence electron concentration and atomic size difference of the MGs play a dominant role in BS and Tx.The quantitative relationships between the key features and properties(BS and Tx)were established to guide the alloy design and property optimization.A series of Fe-based MGs with high Bs and Tx were successfully designed by the ML model,among which the Fe82.55B13.79Si0.9Zr2 76 MG has Bs of 1.61 T and Tx of 738 K while the Fe81.55B14.79Si0.9Zr2.76 MG has Bs of 1.59 T and Tx of 757 K.These results confirmed the reliability of the XGBoost ML model.(2)Furthermore,the prediction of glass transition temperature(Tg)and Tx of MGs was further realized by constructing a ML model involving seven algorithms.The reported Tg and Tx of MGs were used to build a characteristic temperature dataset with the sample number of 841.Algorithms including Lasso,elastic network,support vector regression,kernel ridge regression,random forest,gradient boosting and multilayer perceptron were tested to simultaneously implement the ML model by the Pipeline technique.It was found that the random forests model exhibited the best performance for both Tg and Tx.The feature dimensionality reduction was successfully achieved by the Pearson correlation coefficient analysis.For prediction of Tg with the random forests model,the correlation coefficient(R2)reaches 0.887 with an average absolute error of 19.0 K,while the R2 reaches 0.902 with an average absolute error of 18.4 K for Tx.Compared with previous studies,based on the extended Tg and Tx database,our current ML model shows higher accuracy and good generalization ability.(3)A series of new Fe-based MGs in which the y(FeNi)phase is the primarily competing phase were designed by combining phase diagram calculation and ML,and the alloying effects on the precipitation behavior of y(FeNi)phase,soft magnetic properties and annealing-induced embrittlement of these alloys were investigated systematically,which provides a feasible method to solve the embrittlement of Fe-based MGs after annealing.Based on the phase diagram calculations and experimental data,the ML framework was established to predict the new Fe-based MGs with the y(FeNi)as the primary crystallized phase.On this basis,the crystallization behavior,soft magnetic properties and annealing-induced embrittlement of the newly developed Fe50Ni28-xCo4MoxB18(x=1,2,3)alloys were studied.It was found that the increase of Mo content is beneficial to suppress the precipitation of ?(FeNi)and thus increase the onset crystallization temperature,leading to the enhanced thermal stability of the MGs.The Bs of these specially designed alloys reached the maximum value when being annealed at the temperature of Tx-25?.In addition,the annealing-induced embrittlement of the Fe50Ni25Co4Mo3B18 MG was significantly mitigated,and the ribbon still exhibited good toughness when even been annealed at Tx-25?.In summary,this thesis aims to address the key issues in the field of Fe-based MGs via the combination of ML,phase diagram calculation and experimental methods.Prediction of thermal stability,Bs and Tx by the ML strategy was attempted,and annealing-induced embrittlement of Fe-based MGs was eliminated in several newly designed compositions.The findings presented in this thesis provide a new avenue for the alloy design and property optimization of high-performance Febased MGs.