Buehler Markus J
Laboratory for Atomistic and Molecular Mechanics (LAMM), Massachusetts Institute of Technology 77 Massachusetts Ave. Cambridge MA 02139 USA
Center for Computational Science and Engineering, Schwarzman College of Computing, Massachusetts Institute of Technology 77 Massachusetts Ave. Cambridge MA 02139 USA.
Mater Adv. 2022 Jun 24;3(15):6280-6290. doi: 10.1039/d2ma00223j. eCollection 2022 Aug 1.
Deep learning holds great promise for applications in materials science, including the discovery of physical laws and materials design. However, the availability of proper data remains a challenge - often, data lacks labels, or does not contain direct pairing between input and output property of interest. Here we report an approach based on an adversarial neural network model - composed of four individual deep neural nets - to yield atomistic-level prediction of stress fields directly from an input atomic microstructure, illustrated here for defected graphene sheets under tension. The primary question we address is whether it is possible to predict stress fields without any microstructure-to-stress fields pairings, nor the existence of any input-output pairs whatsoever, in the dataset. Using a cycle-consistent adversarial neural net with either U-Net, ResNet and a hybrid U-Net-ResNet architecture, applied to a system of graphene lattices with defects we devise an algorithmic framework that enables us to successfully train and validate a model that reliably predicts atomistic-level field data of unknown microstructures, generalizing to reproduce well-known nano- and micromechanical features such as stress concentrations, size effects, and crack shielding. In a series of validation analyses, we show that the model closely reproduces reactive molecular dynamics simulations but at significant computational efficiency, and without knowledge of any physical laws that govern this complex fracture problem. The model opens an avenue for upscaling where the mechanistic insights, and predictions from the model, can be used to construct analyses of very large systems, based off relatively small and sparse datasets. Since the model is trained to achieve cycle consistency, a trained model features both forward (microstructure to stress) and inverse (stress to microstructure) generators; offering potential applications in materials design to achieve a certain stress field. Another application is the prediction of stress fields based off experimentally acquired structural data, where the knowledge of solely positions of atoms is sufficient to predict physical quantities for augmentation or analysis processes.
深度学习在材料科学应用方面前景广阔，包括物理定律的发现和材料设计。然而，合适数据的获取仍然是一个挑战——通常，数据缺乏标签，或者不包含感兴趣的输入与输出属性之间的直接配对。在此，我们报告一种基于对抗神经网络模型的方法——该模型由四个独立的深度神经网络组成——用于直接从输入的原子微观结构生成应力场的原子级预测，本文以拉伸状态下有缺陷的石墨烯片为例进行说明。我们解决的主要问题是，在数据集中没有任何微观结构与应力场的配对，甚至不存在任何输入 - 输出对的情况下，是否有可能预测应力场。通过将具有U-Net、ResNet和U-Net-ResNet混合架构的循环一致对抗神经网络应用于有缺陷的石墨烯晶格系统，我们设计了一个算法框架，使我们能够成功训练和验证一个模型，该模型能够可靠地预测未知微观结构的原子级场数据，并能推广以重现诸如应力集中、尺寸效应和裂纹屏蔽等著名的纳米和微观力学特征。在一系列验证分析中，我们表明该模型能紧密重现反应分子动力学模拟结果，但计算效率显著提高，且无需了解任何支配这个复杂断裂问题的物理定律。该模型为扩大规模开辟了一条途径，其中模型的机理见解和预测可用于基于相对较小且稀疏的数据集构建对非常大的系统的分析。由于该模型经过训练以实现循环一致性，一个经过训练的模型具有正向（微观结构到应力）和反向（应力到微观结构）生成器；这在材料设计中提供了潜在应用，以实现特定的应力场。另一个应用是基于实验获取的结构数据预测应力场，其中仅原子位置的知识就足以预测用于增强或分析过程的物理量。
