A Machine Learning Framework for Bridging the Mechanical Responses of a Material at Multiple Structure Length Scales

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

The safety of structures bearing load depends upon the ability of the engineers to predict the behavior under operating conditions during the design phase. This means that the engineers must know the behavior of materials that are used in these structures accurately. Damage and failure in materials start at the atomic scale and propagate through the structural length scales to manifest itself. Therefore, it is important to be able to predict material response through multiple scales based on experimental observations and computational modeling. Much effort has been put towards achieving this capability but the immense amount of information that is obtained from advanced experimental techniques and physics-based numerical simulations has proven intractable. This award aims to transform the current practices in the field of mechanics of materials by introducing machine learning to optimize the extraction and fusion of information and knowledge from the disparate and expansive experimental and numerical datasets. The goal is to dramatically improve the efficiency and output compared to the current protocols of material behavior prediction, which will also accelerate the materials discovery process. It is expected that theses outcomes will provide significant competitive advantages to the US industry in a broad range of advanced materials technology areas, including those related to healthcare, energy, and national security. The award will also be a vehicle to train graduate and undergraduate students at the intersection of materials science, mechanics of materials, and data and information sciences. The research outcomes and developed tools will be transferred into commercial practice through multiple industrial collaborations. It is proposed to accomplish the research objective described above by developing and deploying a novel Bayesian machine learning framework that is centered on systematically uncovering the physics controlling the multiscale materials phenomena of interest. The overall strategy involves establishing suitable high-fidelity reduced-order (i.e., surrogate) models to capture the localization tensors for elastic and plastic deformations in multiphase polycrystalline microstructures. In turn, these models will be used to formulate a computationally efficient strategy for Bayesian sequential design of experiments that identifies the most optimal experiments offering the highest potential for information (or knowledge) gain. As a result, several high-throughput experimental assays will be designed and evaluated to critically examine their value for reliably calibrating the unknown material parameters in sophisticated plasticity theories. Based on the results of these investigations, novel high-throughput protocols will be designed and implemented to demonstrate the significant cost and time savings achieved in the multiscale characterization of the mechanical behavior of heterogeneous structural materials. Specifically, the new protocols will be validated using samples of polycrystalline dual-phase steels. 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.