利用晶体图神经网络和对抗样本方法探索CsSnBr_(3-x)I_(x)中高能量转换效率构型并提高模型的可解释性  

Enhancing interpretability in the exploration of high-energy conversion efficiency in CsSnBr_(3-x)l_(x) configurations using crystal graph convolutional neural networks and adversarial example methods

作　　者：王韬 赖晓珑 卫亚东 郭鸿 金浩 Tao Wang;Xiaolong Lai;Yadong Wei;Hong Guo;Hao Jin(College of Physics and Optoelectronic Engineering,Shenzhen University,Shenzhen 518060,China;Centre for the Physics of Materials and Department of Physics,McGill University,Montreal H3A 2T8,Canada)

机构地区：[1]College of Physics and Optoelectronic Engineering,Shenzhen University,Shenzhen 518060,China [2]Centre for the Physics of Materials and Department of Physics,McGill University,Montreal H3A 2T8,Canada

出　　处：《Science China Materials》2024年第4期1183-1191,共9页中国科学（材料科学）（英文版）

基　　金：supported by Guangdong Basic and Applied Basic Research Foundation (2022A1515012006);Shenzhen Natural Science Fund (the Stable Support Plan Program 20231121110218001);the National Natural Science Foundation of China (12175150);China Postdoctoral Science Foundation (2023M742403)。

摘　　要：晶体图卷积神经网络(CGCNN)不需要手动创建描述符,因而改变了材料研究方式.然而,它们在可解释性以及对结构畸变的敏感性方面的不足,阻碍了该模型在材料掺杂工程中的应用.因此,我们提出了一种具有创新性的对抗样本方法,用于引导特征构建并增强CGCNN的可解释性.在这项研究中,重点是寻找具有高能量转换效率的CsSnBr_(3-x)I_(x)构型.首先,训练了一个CGCNN分类器作为基准.随后,对输入数据进行扰动以生成低性能分类器,并基于错误预测结果来识别对抗样本.将对抗样本与正常样本进行比较,发现对抗样本中存在着明显的结构畸变,这为创建描述符提供了思路.基于此,建立了一个可解释模型,该模型超越了使用原子位置扰动的CGCNN模型以及使用通用特征的梯度提升分类器模型.值得注意的是,之前被忽视的描述符,即“不等价原子数”,在提供关键见解方面发挥了重要作用.进一步研究发现,具有明显畸变的结构可以表现出增强的功率密度,从而提高能量转换效率.本工作不仅阐明了原子替代对能源转换效率的影响,还为构建可解释的机器学习模型提供了可行方案.Crystal graph convolutional neural networks(CGCNNs)have revolutionized materials research by eliminating the need for manual feature engineering.However,their lack of interpretability and sensitivity to structural distortions hinders their application in substitution engineering.Therefore,we propose an innovative adversarial example method for guiding feature construction and enhancing the interpretability of CGCNNs.In this study,our focus lies on identifying CsSnBr_(3-x)I_(x) configurations with high-energy conversion efficiency.Initially,we train a CGCNN classifier as a benchmark.Subsequently,we perturb input data to generate a low-performance classifier and identify adversarial examples based on incorrect predictions.Upon comparing these examples with normal examples,we observe substantial structural distortions in adversarial cases,serving as inspiration for the creation of disorder-related features.Consequently,an interpretable model is developed,which surpasses CGCNNs with atomic position perturbations and a gradient-boosting classifier using general features.Notably,the_previously overlooked feature "number of unequal atoms" plays an important role in offering crucial insights.Further analysis reveals that configurations with pronounced disorder can exhibit increased power density,thereby enhancing the energy conversion efficiency.Our work not only elucidates the impact of atom substitution on energy conversion efficiency but also provides a roadmap for constructing interpretable machine learning models.

关 键 词：卷积神经网络 可解释性 对抗样本 能量转换效率 位置扰动 输入数据 分类器 结构畸变 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TB34[自动化与计算机技术—控制科学与工程]