This website requires JavaScript.

Deep Learning for Multiscale Damage Analysis via Physics-Informed Recurrent Neural Network

Shiguang Deng
Dec 2022
摘要
Direct numerical simulation of hierarchical materials viahomogenization-based concurrent multiscale models poses critical challenges for3D large scale engineering applications, as the computation of highly nonlinearand path-dependent material constitutive responses at the lower scale causesprohibitively high computational costs. In this work, we propose aphysics-informed data-driven deep learning model as an efficient surrogate toemulate the effective responses of heterogeneous microstructures underirreversible elasto-plastic hardening and softening deformation. Ourcontribution contains several major innovations. First, we propose a noveltraining scheme to generate arbitrary loading sequences in the sampling spaceconfined by deformation constraints where the simulation cost of homogenizingmicrostructural responses per sequence is dramatically reduced via mechanisticreduced-order models. Second, we develop a new sequential learner thatincorporates thermodynamics consistent physics constraints by customizingtraining loss function and data flow architecture. We additionally demonstratethe integration of trained surrogate within the framework of classic multiscalefinite element solver. Our numerical experiments indicate that our model showsa significant accuracy improvement over pure data-driven emulator and adramatic efficiency boost than reduced models. We believe our data-driven modelprovides a computationally efficient and mechanics consistent alternative forclassic constitutive laws beneficial for potential high-throughput simulationsthat needs material homogenization of irreversible behaviors.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答