We suggest two metrics for assessing the quality of atomistic configurations of disordered materials, both of which are based on quantifying the orientational distribution of neighbours around each atom in the configuration. The first metric is that of geometric invariance: i.e., the extent to which the neighbour arrangements are as similar as possible for different atoms, allowing for variations in frame of reference. The second metric concerns the degree of local symmetry. We propose that for a set of configurations with equivalent pair correlations, ranking highly those configurations with low geometric invariance but with high local symmetry selects for structural simplicity in a way that does not rely on formal group theoretical language (and hence long-range periodic order). We show that these metrics rank a range of SiO2 and a-Si configurations in an intuitive manner, and are also significantly more sensitive to unphysical features of those configurations in a way that metrics based on pair correlations are not. We also report that implementation of the metrics within a reverse Monte Carlo algorithm gives rise to an energy landscape that is too coarse (at least in this initial implementation) for amorphous structure "solution".
On Strong-Scaling and Open-Source Tools for High-Throughput Quantification of Material Point Cloud Data: Composition Gradients, Microstructural Object Reconstruction, and Spatial Correlations
Physics-Informed Machine Learning and Uncertainty Quantification for Mechanics of Heterogeneous Materials
本文提出了一种基于物理信息神经网络(PINNs)的非均质固体弹性形变求解模型及相应的不确定性量化(UQ)。对于本研究,PINNs框架-模数由Nvidia开发的,其中我们实现了一个模块的力学非均质固体。我们使用PINN近似动量平衡,假设各向同性的线弹性本构行为对损失函数。与控制方程一起,相关的初始/边界条件也柔和地参与损失函数。分析了其中异质性表现为基质中的空隙(低弹性模量区域)和纤维(高弹性模量区域)的固体,并对从商业有限元(FE)分析包获得的解决方案验证了结果。本研究还表明,PINNs可以捕捉到精确的材料界面处的应力跳变。此外,本研究探讨了通过在几何形状和材料属性的变化与替代功能在PINN相关的优势。所提出的UQ研究表明,PINNs解决方案的平均值和标准差与Monte Carlo FE结果吻合良好。PINN预测的单个代表性空隙和单纤维复合材料的有效杨氏模量与FE预测的杨氏模量相比非常好,这建立了PINN制剂作为一种有效的均匀化工具。