金属氢化物材料具有储氢/热密度高,工作温度区间广,无污染无腐蚀性的优点,被认为是理想的储氢/热材料。金属氢化物储氢/热材料可以通过掺杂不同元素形成多元合金,以开发具有不同目标性能的材料。这种方法依赖实验合成,十分耗费时间和经济成本。数据驱动的机器学习性能预测模型可以解决这一问题,通过测试对比最小二乘回归、最小绝对收缩和选择操作符回归、岭回归、弹性网络回归、支持向量回归和随机森林回归多种回归算法,成功建立了金属氢化物微观材料性质与宏观形成能之间的关系。测试结果显示随机森林回归具有最好的预测性能,在训练集和测试集上相对误差均较小,仅为3.078和8.2011,且决定系数较高,具有良好的回归能力和泛化能力。SHAP分析中表明组成金属氢化物的基态原子体积的均值和最值具有高达5.56和1.26的SHAP值,这2个因素很大程度上决定了金属氢化物材料的形成能大小。对Mg基,Ca基,AB、AB2及AB5型金属氢化物材料预测结果分析显示预测相对误差均在9%以下,证明了模型准确性及普适性,可用于未知数据集的形成能预测。

Metal hydride thermal/hydrogen energy storage material is considered ideal candidate due to high energy density,wide workingtemperature range and lack of corrosive pollutants. Multi-component metal hydride alloys can be formed by doping with differentelements to obtain various target properties. However, conventional material development method relies on experimental synthesis,having the disadvantages of time-consuming and costly. Data-driven machine learning prediction model is capable of addressing thisproblem. By comparing varieties of regression algorithms such as least squares regression,least absolute shrinkage and selection operatorregression,ridge regression,elastic net regression,supporting vector regression,and random forest regression,the relationship betweenthe microscopic properties of metal hydride materials and their macroscopic formation energy are established. Results show that randomforest regression have the best prediction performance,with lowest relative errors on both the training and test sets of 3.078 and 8.2011,high R-squared values,and great generalization and regression abilities. SHAP analysis reveals extreme and mean value of ground stateatom of metal hydride exhibit the greatest SHAP value of 5.56 and 1.26,suggesting their significant influence on the formation energy.Analysis for the prediction value of Mg-base,Ca-base,AB type,AB2 type,and AB5 type metal hydrides shows the highest relative errorbelow 9%,further proving the accuracy and universality of the model for all types of metal hydride. This model can be used to predict theformation enthalpy of unknown datasets.