-
针对快速获取钙钛矿材料带隙值的问题, 建立特征融合神经网络模型(CGCrabNet), 利用迁移学习策略对钙钛矿材料的带隙进行预测. CGCrabNet从材料的化学方程式和晶体结构两方面提取特征, 并拟合特征和带隙之间的映射, 是一个端到端的神经网络模型. 在开放量子材料数据库中数据(OQMD数据集)预训练的基础上, 通过仅175条钙钛矿材料数据对CGCrabNet参数进行微调, 以提高模型的稳健性. 数值实验结果表明, CGCrabNet在OQMD数据集上对带隙的预测误差比基于注意力的成分限制网络(CrabNet)降低0.014 eV; 本文建立的模型对钙钛矿材料预测的平均绝对误差为0.374 eV, 分别比随机森林回归、支持向量机回归和梯度提升回归的预测误差降低了0.304 eV、0.441 eV和0.194 eV; 另外, 模型预测的SrHfO 3和RbPaO 3等钙钛矿材料的带隙与第一性原理计算的带隙相差小于0.05 eV, 这说明CGCrabNet可以快速准确地预测钙钛矿材料的性质, 加速新材料的研发过程.
The band gap is a key physical quantity in material design. First-principles calculations based on density functional theory can approximately predict the band gap, which often requires significant computational resources and time. Deep learning models have the advantages of good fitting capability and automatic feature extraction from the data, and are gradually used to predict the band gap. In this paper, aiming at the problem of quickly obtaining the band gap value of perovskite material, a feature fusion neural network model, named CGCrabNet, is established, and the transfer learning strategy is used to predict the band gap of perovskite material. The CGCrabNet extracts features from both chemical equation and crystal structure of materials, and fits the mapping between feature and band gap. It is an end-to-end neural network model. Based on the pre-training data obtained from the Open Quantum Materials Database (OQMD dataset), the CGCrabNet parameters can be fine-tuned by using only 175 perovskite material data to improve the robustness of the model. The numerical and experimental results show that the prediction error of the CGCrabNet model for band gap prediciton based on the OQMD dataset is 0.014 eV, which is lower than that obtained from the prediction based on compositionally restricted attention-based network (CrabNet). The mean absolute error of the model developed in this paper for predicting perovskite materials is 0.374 eV, which is 0.304 eV, 0.441 eV and 0.194 eV lower than that obtained from random forest regression, support vector machine regression and gradient boosting regression, respectively. The mean absolute error of the test set of CGCrabNet trained only by using perovskite data is 0.536 eV, and the mean absolute error of the pre-trained CGCrabNet decreases by 0.162 eV, which indicates that the transfer learning strategy plays a significant role in improving the prediction accuracy of small data sets (perovskite material data sets). The difference between the predicted band gap of some perovskite materials such as SrHfO 3and RbPaO 3by the model and the band gap calculated by first-principles is less than 0.05 eV, which indicates that the CGCrabNet can quickly and accurately predict the properties of new materials and accelerate the development process of new materials. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] -
超参数名称 含义 值 $ {d}_{{\rm{m}}} $ 元素特征构造得到的向量维度 512 $ {N}_{{{f}}} $ 化学式中最大元素种类 7 N 注意力机制层数 3 n 注意力机制头数 4 I 参与训练的元素种类和 89 T 图卷积层数 3 $ {V}_{{\rm{c}}{\rm{g}}} $ 节点嵌入后元素向量维度 16 $ {w}_{1} $∶$ {w}_{2} $ 权重比参数 7:3 Epochs 最大迭代次数 300 batch_size 批处理大小 256 
下载:
导出CSV
元素嵌入方法 Train MAE Val MAE Test MAE One-Hot 0.185 0.423 0.433 Magpie 0.428 0.546 0.566 Mat2vec 0.203 0.408 0.420 下载:
导出CSV
Train MAE Val MAE Test MAE CGCNN 0.502 0.605 0.601 Roost 0.178 0.447 0.455 CrabNet 0.226 0.422 0.427 HotCrab 0.177 0.422 0.440 CGCrabNet 0.187 0.408 0.413 下载:
导出CSV
机器学习方法 超参数名称 取值 RF 子学习器数量 90 SVR 核函数 多项式核 多项式核次数 3 正则化强度 2 伽马参数 2 零系数 1.5 GBR 子学习器数量 500 学习率 0.2 最大深度 4 损失函数 绝对误差函数 下载:
导出CSV
化学式 带隙计算值 CGCrabNet RF SVR GBR NbTlO3 0.112 0.658 1.458 1.296 1.614 ZnAgF3 1.585 1.776 1.836 2.194 1.840 AcAlO3 4.102 3.212 2.881 3.197 2.963 BeSiO3 0.269 1.116 2.813 2.963 3.777 TmCrO3 1.929 1.682 1.612 1.987 1.668 SmCoO3 0.804 0.644 0.821 1.043 0.724 CdGeO3 0.102 0.586 0.911 1.675 0.196 CsCaCl3 5.333 4.891 4.918 5.116 5.157 HfPbO3 2.415 2.724 1.733 2.346 1.967 SiPbO3 1.185 1.327 1.407 1.543 1.079 SrHfO3 3.723 3.683 2.821 3.370 3.253 PrAlO3 2.879 3.139 2.665 2.091 2.984 BSbO3 1.405 1.123 0.653 -0.025 0.579 CsEuCl3 0.637 0.388 1.500 4.477 0.949 LiPaO3 3.195 3.100 2.443 -0.306 2.553 PmErO3 1.696 1.309 1.550 1.682 1.252 TlNiF3 3.435 2.806 2.063 3.049 3.255 MgGeO3 3.677 1.256 0.979 1.623 1.073 NaVO3 0.217 0.785 0.911 0.180 0.989 RbVO3 0.250 0.616 1.736 0.290 1.534 KZnF3 3.695 3.785 2.853 3.203 3.295 NdInO3 1.647 1.587 1.653 0.889 1.590 RbCaF3 6.397 6.974 6.482 6.372 6.028 RbPaO3 3.001 2.952 2.864 -0.234 2.937 PmInO3 1.618 1.480 1.896 1.222 1.754 KMnF3 2.656 2.991 2.647 2.428 2.730 NbAgO3 1.334 1.419 1.369 1.227 1.265 CsCdF3 3.286 3.078 2.990 2.724 2.879 KCdF3 3.101 3.125 2.789 2.365 2.990 CsYbF3 7.060 6.641 6.325 6.523 6.736 NaTaO3 2.260 1.714 1.680 2.093 1.715 CsCaF3 6.900 6.874 6.291 6.416 6.379 RbSrCl3 4.626 4.470 4.966 4.647 4.795 AcGaO3 2.896 3.199 2.740 2.869 2.981 BaCeO3 2.299 1.789 3.918 2.696 3.655 注: CGCrabNet, RF, SVR和GBR分别代表特征融合神经网络、随机森林回归、支持向量回归和梯度提升回归模型计算得到的带隙值. 下载:
导出CSV
-
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26]
下载:
计量
- 文章访问数:2262
- PDF下载量:90
- 被引次数:0