npj Computational Materials

Graph deep learning accelerated efficient crystal structure search and feature extraction

Chuan-Nan Li, Han-Pu Liang, Xie Zhang, Zijing Lin & Su-Huai Wei

npj Computational Materials 9: 176 (2023)

doi.org/10.1038/s41524-023-01122-4

Published online: 30 September 2023
Abstract| Full Text | PDF OPEN

Abstract: Structural search and feature extraction are a central subject in modern materials design, the efficiency of which is currently limited, but can be potentially boosted by machine learning (ML). Here, we develop an ML-based prediction-analysis framework, which includes a symmetry-based combinatorial crystal optimization program (SCCOP) and a feature additive attribution model, to significantly reduce computational costs and to extract property-related structural features. Our method is highly accurate and predictive, and extracts structural features from desired structures to guide materials design. We first test SCCOP on 35 typical compounds to demonstrate its generality. As a case study, we apply our approach to a two-dimensional B-C-N system, which identifies 28 previously undiscovered stable structures out of 82 compositions; our analysis further establishes the structural features that contribute most to energy and bandgap. Compared to conventional approaches, SCCOP is about 10 times faster while maintaining a comparable accuracy. Our framework is generally applicable to all types of systems for precise and efficient structural search, providing insights into the relationship between ML-extracted structural features and physical properties.

摘要: 现代材料设计中，结构搜索和特征提取是一个核心课题，但目前采用的方法效率还有待提高。近年来兴起的机器学习（ML）为解决这个问题提供了机遇。本文中我们开发了一种基于ML的预测分析框架，包括基于对称性的组合晶体优化程序（SCCOP）和具有叠加性的特征归因模型，显著地降低了计算成本，并可提取与性质相关的结构特征。我们的方法具有高度准确性和预测性，并可从预测的结构中提取结构特征以指导材料设计。我们首先在35个典型化合物上测试了SCCOP普适性。在此基础上，我们进一步将新方法应用到B-C-N二维体系，从82种成分中寻找出28个以前未被发现的动力学稳定结构；通过特征归因模型，我们进一步确定了对能量和带隙贡献最大的结构特征。与传统方法相比，SCCOP的结构预测速度提高了一个数量级，同时保持了与第一性原理计算相当的准确性。我们的新框架适用于所有化合物组分，可用于精确而高效的结构搜索，为ML提取的结构特征与物理性质建立了关系，提供了新见解。

Editorial Summary

Peering into Crystals: How Can Graph Neural Networks Unveil the Mysteries of Crystals?

Crystal structure prediction aims to identify the thermodynamically stable crystal structures with desired properties under specific temperature and pressure for given chemical compositions. Traditional methods often require vast computational resources. Therefore, developing more efficient and more accurate new methods has been a long-standing challenge in condensed matter physics, materials science, and chemistry. Led by Prof. Zijing Lin at the University of Science and Technology of China and Prof. Su-Huai Wei at the Beijing Computational Science Research Center, Chuan-Nan Li and Han-Pu Liang et al. have developed a crystal structure search program (SCCOP) based on Graph Neural Networks (GNN) and a feature extraction model. This innovative scheme allows for rapid generation of low-energy structures for given atomic compositions and extraction of structural features related to the target material properties. What sets this model apart is its integrated machine-learning approach, including structure generation with symmetry and distance constraints, structure search with graph representation of crystal, fine-tuning with transfered learning, GNN-based structure optimization, and the construction of a feature attribution model. This considerably shortens the discovery time and design cycle of new functional materials. Compared to traditional methods relying on density functional theory and expert knowledge analysis, this framework offers higher efficiency and better interpretability while maximizing the benefits of machine learning. Using the B-C-N two-dimensional system as a case study, SCCOP successfully predicted many previously unreported crystal structures as well as established relationships between structure stabilization, band gaps, and coordination numbers. It also identified key factors in energy reduction and bandgap formation. The research team discovered five stable B-C-N wide bandgap materials with excellent mechanical properties and low thermal conductivity. The significance of this study lies in offering a novel tool and methodology for the materials science and engineering community. It promises to accelerate the discovery and design of new materials, potentially propelling further advancements in materials science.

窥晶探秘：图神经网络揭示的晶体之谜？

晶体结构预测的目标是在给定组分的情况下，预测在一定温度和压力条件下具有最佳热稳定性的晶体结构。传统的预测方法通常依赖于大量的计算资源，如何更进一步开发高效、准确的新型方法是凝聚态物理、材料科学和化学等领域长期以来的一项重要挑战。为解决这一问题，中国科大的林子敬教授和北京计算科学研究中心的魏苏淮教授带领团队成员李川南和梁汉普等，开发了一种基于图神经网络（GNN）并具有特征提取的晶体结构搜索程序（SCCOP），实现了在给定组分下快速生成所需要的结构，并提取与材料性质相关的结构特征。该模型的创新之处在于它采用了一体化的机器学习方法，包括利用对称性和距离约束生成结构、晶体图表示的结构搜索、迁移学习微调、基于GNN的结构优化以及特征归因模型的构建等，可大大缩短了新型功能材料的发现和设计周期。与传统的基于密度泛函理论和专家知识分析的方法相比，这一框架在充分利用机器学习的同时，提供了更高的效率和可解释性。本研究以B-C-N二维体系为例，用SCCOP成功预测出了大量未报导过的晶体结构，并建立起结构稳定性、带隙与配位数之间的关系，识别出降低能量与形成带隙的关键因素。研究团队发现的五种稳定的B-C-N宽带隙材料具有出色的力学性能和低热导率，为今后的应用提供了理论支撑。这项研究的重要性在于它为材料科学和工程领域提供了一种新的工具和方法，可以加速新材料的发现和设计，有望推动材料科学的进一步发展。