Learning-based multiscale method for solid materials
Abstract
The 2019 Army Modernization Strategy describes `Materials by Design as an Army priority research area. The empirical development and optimization of new materials is an extremely expensive and time-consuming iterative exercise because the macroscopic behavior of materials is the result of interconnected mechanisms that operate over a wide range of length and time scales. Experimental studies are expensive and limited in their ability to observe the entire spectrum of mechanisms. Therefore, multiscale modeling has emerged as an essential tool. However, the practical implementation of the multiscale modeling is challenging due to the need to balance accuracy and computational complexity. We propose an alternate approach that exploits machine learning combined with model reduction in multiscale modeling of materials. The basic idea is to create a surrogate that approximates the solution operator of the fine scale model and use the surrogate in coarse calculations. This approach provides the fidelity of concurrent multiscale modeling at a few times the computational cost of an empirical model. We have demonstrated this approach in a two-scale setting. We propose to build on this to address a series of increasingly difficult problems leading up to a high fidelity, computationally ecient, and rigorous multiscale methodology with broad applications to various problems of material strength and failure. We propose to study a number of fundamental issues that arise in this methodology. These include (i) exploiting emerging architectures to learn history dependence of material behavior and to develop new insights into the underlying mechanics; (ii) finding ecient representation of microstructure (polycrystalline texture, defect structures etc.) and its evolution under extreme loading conditions; (iii) Understanding how uncertainties propagate through the multiscale hierarchy, quantifying system-level uncertainties due to uncertainties in the individual levels of the hierarchy, and enabling a materials by design approach by computing system level sensitivity to individual mechanisms; and (iv) integrating experimental data when available with data obtained by direct numerical simulation to train the ML surrogates in such a manner that priorities experimental data. We also propose to demonstrate the methodology by applying it to specific phenomena of interest to the Army and DoD. We study canonical problems and example material systems that highlight fundamental aspects of the underlying phenomena. Speci cally, we propose to focus on (a) strength of metals including consequences of texture evolution and localization during processing and dynamic loading; and (b) failure of fiber-reinforced multi-ply laminated composites under ballistic loading accounting for fiber cracking, fiber debonding, fiber pull-out and matrix cracking at the microscopic level that manifest as di use damage at the macroscale, and inter-ply delamination at the mesoscale. Finally, we propose to collaborate with researchers at ARL and other Army and DoD laboratories to use the issues they face to motivate fundamental research, and then to find pathways to transition our results into these organizations. The proposed research will be carried out at the California Institute of Technology by the PI and two graduate students. Since we address core issues of interest to the Army, Department of Defense and industry, the potential impact of the proposed work is both deep and broad.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 20, 2022
- Source ID
- W911NF2210269
Entities
People
- Kaushik Bhattacharya
Organizations
- Army Contracting Command
- California Institute of Technology
- United States Army