CAREER: Computational Failure Mechanics Across Multiple Scales with Deep Reinforcement Learning

点击翻译按钮获取中文摘要

This Faculty Early Career Development (CAREER) Program grant will support fundamental research in using human-interpretable knowledge generated by artificial/augmented intelligence (AI) to discover hidden physics mechanisms that lead to the failure of materials. While there has been considerable progress made in basic understanding of various failure mechanisms, e.g. brittle fracture, strain localization, and ductile flow, the recent advancements in experimental techniques, such as digital image correlation and micro-computed tomography imaging, have led to an influx of data on failure that is not always easy to incorporate into models manually. The current effort leverages the AI's ability to repeatedly generate and test hypotheses such that it can self-identify and discover mistakes in previous modeling efforts and identify new physics too subtle or difficult to identify and interpret manually. These computer-generated discoveries will be a powerful tool in accelerating the progress of science. By understanding how failures of materials and structures occur, the research will advance national prosperity and welfare by helping engineers making more efficient, robust and precise designs-by-analyses for infrastructure, structural components and devices. As part of the grant, the PI will also facilitate mentoring relationships with selected underrepresented high school students and high school teachers in the Harlem district of New York City to enable future generations of engineers and scientists in leveraging the opportunities afforded by AI. A new meta-modeling paradigm is planned to adaptively generate models to capture the effects of evolving microstructures due to micro-cracks, plastic slip, and wear at sub-scales, and then to recursively upscale responses to the scale of interest. By conceptualizing a constitutive law (theoretical or data-driven) as a directed multi-graph, i.e. a flow network of information, we idealize the process of writing a constitutive law as a combination of actions operated on the directed multi-graph. With a reward function defined as a function of accuracy, robustness, speed, and consistency, we then invent a game whose goal is to maximize the rewards with finite actions (e.g. shape of yield surface, hardening rules, damage, nonlocality, refinement criterion) against constraints (e.g. material-frame indifference, thermodynamic laws). This technique is then applied in hierarchical and concurrent multiscale coupling models. In the hierarchical models, hybridized machine learning models can be used as a surrogate to bridge scales. In the concurrent model, a multi-phase field model is used to enable several types of models (theoretical, data-driven, hybridized, homogenization-from sub-scale-simulations) to be deployed at different domains of interests (e.g. crack tip, shear band) to yield the most accurate simulations.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.