Collaborative Innovation Center of Advanced Microstructures, Laboratory of Solid State Microstructures, School of Physics, Nanjing University, Nanjing 210093, China
Fund Project:Project supported by the National Natural Science Foundation of China (Grant No. 118742007)
Received Date:18 June 2019
Accepted Date:20 August 2019
Available Online:01 November 2019
Published Online:05 November 2019
Abstract:Ferroelectrics undergoes a reversible structural phase from the ferroelectric phase to the paraelectric phase when its temperature exceeds the critical temperature namely Curie temperature Tc. As ferro-paraelectric phase transition is always accompanied by heat-flow, dielectric and pyroelectric anomaly, the value of Tc is extremely important for ferroelectrics. In this paper, the Curie temperature of lead-based perovskite ferroelectric solid solution is studied by machine learning methods including kernel ridge regression (KRR), support vector regression (SVR) and extremely randomized trees regression (ETR). We collect the Tc values of 205 different lead-based perovskites from published experimental papers, both simple perovskites with only one type of B site ion and complex perovskites with up to 5 kinds of ions in B position such as PMN-PFN-PZT are gathered. The diversity of our dataset is guaranteed for the good generalization of our model in perovskite solid solution of different complexity. The features are constructed from the physical and chemical properties of the B site elements in corresponding materials. The weighted-average and variance of the elemental properties are calculated and fed to machine learning models. We use the 5 runs of ten fold cross-validation method to evaluate the machine learning models. The hyperparameters are also chosen carefully with the cross-validation to avoid over fitting. The radial basis function kernel is used in both KRR and SVR. The insensitive error in the SVR is set to be 4 which is comparable to the random error in experiment. From our cross-validation, we find that the mean average errors (MAEs) between the predicted and experimental values of the machine learning methods are 14.4 K, 14.7 K, and 16.1 K, respectively. And the root-mean-square errors (RMSEs) are 22.5 K, 23.4 K, 23.8 K, respectively. After the optimization and the evaluation, our three machine learning models are stacked together by averaging the output of each regression model and thus building an ensemble model. The MAE of the ensemble model is 13.9 K. The RMSE of the ensemble model is 21.4 K. The predicted values keep a correlation coefficient of 0.97 with the experimental values. From the variance reduction in ETR, we derive the importance of our features when determining the Curie temperatures. The five most important factors in our ETR model are " weighted-average thermal conductivity”, " weighted-average conductivity”, " variance of specific heat capacity”, " weighted-average element number”, and " weighted-average relative atomic displacement”. We predict the Curie temperatures higher than those of 200000 types of lead-based perovskites after being trained. Now, we provide two ferroelectric materials that may have high Curie temperatures: 0.02PbMn1/2Nb1/2O3-0.98PbTiO3 (0.02PMN-0.98PT) and 0.02PbGa1/2Nb1/2-0.02PbMn1/2Nb1/2O3-0.96PbTiO3 (0.02PGN-0.02PMN-0.96PT). The predicted Curie temperatures of them are 481 ℃ and 466 ℃, respectively. Keywords:machine learning/ ferroelectric/ Curie temperature/ perovskite
模型的超参数直接影响了模型的拟合能力和拟合策略, 选取一组合适的超参数对模型的构建非常重要. 在定义了交叉验证的策略之后, 在一定的超参数区间内通过随机搜索[36]和网格搜索的方式选取若干组超参数并使用该超参数构建模型. 通过之前定义的交叉验证策略对模型效果进行检验, 最终选取使得交叉验证结果最理想的一组超参数作为模型最终的超参数. 例如, 当优化SVR中的超参数ε时, 在0—10的取值范围内以0.5为步长对ε进行网格搜索. 如图1所示, 对比了ε取不同值时模型的测试误差, 同时将该参数模型应用于全部数据集时的支持向量数目也一并列出, 以作参考. 当ε取值小于4时模型的误差较低; 当ε取值进一步增大时模型的误差出现快速上升. 另一方面随着ε取值的增大, 模型中的支持向量的数目也在减少. 在泛化能力相似的情况下, 倾向于选择较简单的模型以避免过拟合, 因此ε的值最终取为4. 此时, 模型经过完整地学习训练集中所有205项数据后, 其模型中的支持向量数目为147项, 也就是说模型仅使用其中147项材料的数据对未知材料进行预测. 图 1 支持向量回归中的超参数ε的优化及支持向量数目分析 Figure1. Optimization of hyperparameters in support vector regression and the analysis of the number of support vectors
表1本文三种机器学习方法所采用的超参数 Table1.Hyperparameters of the three machine learning methods in this study.
22.4.模型的集成 -->
2.4.模型的集成
上述三种模型, 从不同的角度出发对钙钛矿铁电材料的居里温度进行了学习, 均取得了良好的成效. 为了进一步提高预测的准确度, 通过集成的方法, 将上述三种回归模型融合为一个统一的模型. 集成模型同时训练上述三种机器学习模型, 并使用三种模型分别对居里温度进行预测. 将三种机器学习模型的预测均值作为最终的输出. 为了检验集成的效果, 也对集成模型使用交叉验证的方法进行了评估. 另外, 也尝试了赋予三个模型其他合理的权重, 如图2所示, 当将三种模型融合在一起时, 往往可以得到比最优单模型更好的效果. 测得的最优融合比例为0.6∶0.2∶0.2 (KRR∶SVR∶ETR), 其MAE为13.7 K. 等权重模型融合测得的MAE为13.9 K (图2中的红色点). 出于保持模型简单性的考虑, 直接采用了等权重的方法对模型进行融合. 图 2 不同模型权重的融合实验结果 Figure2. Performance of ensembled model with different base model weight.