基于融合策略的块体金属玻璃形成能力预测

Glass forming ability prediction of bulk metallic glasses based on fused strategy

提出一种融合策略来提高随机森林(RF)、k近邻(KNN)、梯度提升决策树(GBDT)和极端梯度提升(XGBoost)模型预测块体金属玻璃(BMGs)形成能力(GFA)的准确性。该策略使用训练好的卷积神经网络(CNN)模型提取来特征向量,且合金的成分信息是唯一的输入变量,不需要通过实验获得其他物理和化学特性。此外,通过网格搜索方法和k折交叉验证对RF、KNN、GBDT和XGBoost模型的超参数进行了优化。结果表明,本文作者提出的4种融合模型CNN-RF、CNN-KNN、CNN-GBDT和CNN-XGBoost比上述4种机器学习模型(即RF、KNN、GBDT和XGBoost模型)预测精度更高,这表明训练好的CNN模型比人工特征提取更为高效。此外,与已报道的机器学习模型和GFA判据相比,本文作者提出的融合模型可以更精准地预测金属玻璃的形成能力。

In order to improve the prediction accuracy of random forest(RF),k-nearest neighbor(KNN),gradient boosted decision trees(GBDT)and extreme gradient boosting(XGBoost)models,a fused strategy was proposed for predicting the glass forming ability(GFA)of bulk metallic glasses(BMGs).Feature vectors were extracted using a trained convolutional neural network(CNN),and alloy composition information was the only variable input without requiring various physical and chemical properties acquired from experiments.Besides,the hyperparameters of RF,KNN,GBDT and XGBoost models were optimized by grid search method and k-fold cross validation.The obtained results show that the accuracy of CNN-RF,CNN-KNN,CNN-GBDT and CNN-XGBoost fused models proposed in this work in predicting GFA is higher than that of the four machine learning models mentioned above(i.e.,RF,KNN,GBDT and XGBoost models),implying that the trained CNN could extract feature more effectively than manual feature construction.Furthermore,compared with previously reported machine learning models and GFA criteria,the proposed fused models could predict the GFA of BMG more accurately.