Npj Comput. Mater.: “中心-环境”深度迁移学习—快速预测钙钛矿氧化物

海归学者发起的公益学术平台

分享信息，整合资源

交流学术，偶尔风月

Formation energies predicted by the ML with CE features (DNN-CE) and DFT using various datasets.

来自上海大学材料基因组工程研究院的刘轶教授和冯凌燕教授团队，展示了一种结合自主设计的“中心-环境”（CE）特征模型和深度迁移学习方法，来预测钙钛矿氧化物稳定性的高效高通量筛选策略。他们用5329个尖晶石氧化物结构的形成能开发了基于结构信息“中心-环境”（CE）特征的深度神经网络（DNN）源域模型，然后通过学习855个钙钛矿氧化物结构的小数据集对模型参数进行微调，获得了在钙钛矿氧化物目标域中具有良好可迁移性的迁移学习模型。通过迁移学习模型所预测的钙钛矿结构的形成能平均绝对误差（MAE）仅为0.106 eV/atom，优于只使用钙钛矿数据训练模型0.132 eV/atom的MAE。在迁移学习方法的帮助下，他们发现，在大型数据集上训练的 DNN-CE 模型可以迁移应用于不同的晶体结构。结合包含结构信息的“中心-环境”特征和迁移学习方法，可以有效解决小数据集上训练的机器学习（ML）模型精度差和大数据集上训练的ML模型可迁移性差的问题。

General schematic diagram of DNN-CE models and the workflow of transfer learning method in this work.

为构建目标域虚拟样本空间，使用元素周期表中的73种元素代替钙钛矿中的阳离子，生成5329个立方钙钛矿结构。基于迁移学习模型，进一步预测了包含73种元素的5329个潜在钙钛矿结构的形成能。

Crystal structures and constituent elements of spinel oxides and perovskite oxides studied in this work.

结合机器学习预测的形成能和包括容忍因子（0.7 < t ≤ 1.1）和八面体因子（0.45 < μ < 0.7）的结构因子判据，作者预测了1314种潜在的热力学稳定的钙钛矿氧化物。在这些预测的潜在钙钛矿氧化物中，已有144种被实验合成，其他计算工作预测了其中的10种，Materials Project数据库中收录了301种，而另外859种氧化物尚未在文献中报道过。

Statistical distribution of the formation energy of perovskite structures predicted by machine learning and the screening process for stable perovskite structures.

作者的研究结合基于结构信息的“中心-环境”机器学习特征和迁移学习方法，利用现有大数据以较低成本预测新晶体结构，为昂贵的高通量计算筛选材料设计提供了有效的加速策略。预测的新型钙钛矿氧化物为新型钙钛矿实验合成，探索可再生能源和电子材料应用提供了丰富的材料候选平台。

Center-environment deep transfer machine learning across crystal structures: from spinel oxides to perovskite oxides

Yihang Li, Ruijie Zhu, Yuanqing Wang, Lingyan Feng & Yi Liu

In data-driven materials design where the target materials have limited data, the transfer machine learning from large known source materials, becomes a demanding strategy especially across different crystal structures. In this work, we proposed a deep transfer learning approach to predict thermodynamically stable perovskite oxides based on a large computational dataset of spinel oxides. The deep neural network (DNN) source domain model with “Center-Environment” (CE) features was first developed using the formation energy of 5329 spinel oxide structures and then was fine-tuned by learning a small dataset of 855 perovskite oxide structures, leading to a transfer learning model with good transferability in the target domain of perovskite oxides. Based on the transferred model, we further predicted the formation energy of potential 5329 perovskite structures with combination of 73 elements. Combining the criteria of formation energy and structure factors including tolerance factor (0.7 < t ≤ 1.1) and octahedron factor (0.45 < μ < 0.7), we predicted 1314 thermodynamically stable perovskite oxides, among which 144 oxides were reported to be synthesized experimentally, 10 oxides were predicted computationally by other literatures, 301 oxides were recorded in the Materials Project database, and 859 oxides have been first reported. Combing with the structure-informed features the transfer machine learning approach in this work takes the advantage of existing data to predict new structures at a lower cost, providing an effective acceleration strategy for the expensive high-throughput computational screening in materials design. The predicted stable novel perovskite oxides serve as a rich platform for exploring potential renewable energy and electronic materials applications.