CRII:OAC: Machine Learning- Enhanced Multiscale Simulation of Fiber Composites

许多工程材料，如纤维复合材料，具有跨越多个长度尺度的分层结构。这些材料的分析和设计依赖于多尺度模拟，如果结构较大且材料变形依赖于其加载历史，其计算成本将显著增加。这些高成本阻碍了不确定性传播和设计优化等计算密集型研究。为了应对这一挑战，该项目采用了高性能计算和机器学习方面的最新进展，在不影响准确性的情况下，以数量级加速多尺度模拟。所开发的方法和工具适用于许多材料系统，纤维复合材料的测试平台有益于广泛的学术和工业努力，因为这些材料在汽车和航空航天工业中大量使用。这项工作开发了网络基础设施基础，将加速多尺度模拟，同时(1)最大限度地减少尺度间通信中产生的信息损失，(2)考虑各种不确定性来源，如微观结构性质和形态的空间变化。该项目建立了机械机器学习（ML）模型，模拟复杂和历史依赖的微观结构变形，这些变形体现了广泛的纳米尺度和中尺度效应，以确保可转移性。ML模型与消息传递接口（MPI）设计集成，该设计利用多尺度模拟的分层特性，在计算集群的计算节点内部和跨计算节点实现两级并行性。在对微结构因制造不确定性而发生空间变化的轻质纤维复合材料进行多尺度模拟时，采用信息传递来管理尺度间和尺度内的数据传输。对复合材料的模拟旨在增加对其性能如何受到微观结构间和微观结构内部不确定性、组成特性、变形历史和微观结构形态的影响的理解。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Many engineered materials such as fiber composites have a hierarchical structure that spans multiple length scales. The analysis and design of these materials rely on multiscale simulations whose computational costs significantly increase if the structure is large and if the material deformation depends on its loading history. These high costs prohibit computationally intensive studies such as uncertainty propagation and design optimization. To tackle this challenge, the project employs recent advances in high performance computing and machine learning to accelerate multiscale simulations by orders of magnitude without compromising accuracy. The developed methods and tools are applicable to many materials systems and the testbed on fiber composites benefits a wide range of academic and industrial efforts since these materials are heavily used in, for example, the automobile and aerospace industries.This work develops cyberinfrastructure foundations that will enable acceleration of multiscale simulations while (1) minimizing the information loss incurred in inter-scale communication, and (2) considering various uncertainty sources such as spatial variation of microstructural properties and morphologies. The project builds mechanistic machine learning (ML) models that emulate complex and history-dependent microstructural deformations that embody a broad range of nanoscale and mesoscale effects to ensure transferability. The ML models are integrated with a message passing interface (MPI) design that leverages the hierarchical nature of the multiscale simulation to achieve two-level parallelism, both within and across the computational nodes of a compute cluster. The message passing is employed to manage inter-scale and intra-scale data transfer during multiscale simulation of a light-weight fiber composite whose microstructures spatially vary due to manufacturing uncertainties. The simulations on composite materials aim to increase understanding of how their properties are affected by inter- and intra- microstructural uncertainties, constituent properties, deformation history, and microstructure morphology.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.