摘要
发现和优化纳米复合材料可以通过融合从头算方法和机器学习(ML)来实现,目前的瓶颈在于缺乏原子尺度上描述单体间相互作用的ML模型.我们开发了名为双区域网络(DRN)的深度学习算法,通过训练学习原子之间的多尺度交互作用,预测了由337种不同分子组成的分子对在随机构型中的电子耦合强度,平均绝对误差低至2.8 meV.基于两种截止半径的层次化材料表象被证明对DRN模型的高迁移性和鲁棒性至关重要,这种材料表征方式不仅捕获了共轭片段的局部特征,而且在模型训练之前编码了重要的分子片段间相互作用.本研究建立的ML模型为描述单体间相互作用提供了普适的建模框架,为复杂纳米复合材料的逆向设计打下了基础.
The discovery and optimization of functional nanocomposites can be potentially accomplished by joint abinitio and machine learning(ML)exploitation,which is currently hindered by the absence of an ML model to appropriately describe intermonomer interactions in atomic scale.We developed a deep learning model named double-region network(DRN)to fill this gap via simultaneously learning multi-scale interactions.An ultra-low mean absolute error of 2.8 meV is achieved to predict electronic couplings of 337 distinct molecule types in random configurations,with a tiny training set of 21 configurations per molecule type.The hierarchical material representation based on atomic chemical environments with small and large cutoff radii is demonstrated to be crucial for the high transferability and robustness of the DRN model.Such representation not only captures the local features of conjugated fragments,but also encodes the important intermolecular fragment interactions prior to model training.The ML model established in this study offers a general framework for describing intermonomer interactions and opens an opportunity for the inverse design of complex nanocomposites.
作者
李翀
梁超
伊丽米然木·肉扎洪
王彪
李华山
Chong Li;Chao Liang;Yilimiranmu Rouzhahong;Biao Wang;Huashan Li(School of Physics,Sun Yat-Sen University,Guangzhou 510275,China)
基金
supported by the National Natural Science Foundation of China(NSFC,52072417 and 11832019)
the NSFC Original Exploration Project(12150001)
the Natural Science Foundation of Guangdong Province(2018B030306036)
Guangdong Science&Technology Project(2019QN01C113)
the Project of Nuclear Power Technology Innovation Center of Science Technology and Industry for National Defense(HDLCXZX-2021-HD-035)
Guangdong International Science and Technology Cooperation Program(2020A0505020005)。
