Session: 13-19-01: Scientific Machine Learning (SciML) for Characterization, Modeling, and Design of Structures and Materials I

Paper Number: 166261

Robust Scientific Machine Learning for Experimental Mechanics 

Materials have hierarchical and heterogeneous structures that fundamentally drive their deformation and failure mechanisms. The relationship between a material’s structure and its behavior -- such as the impact of the microstructure of a polycrystalline metal on twinning, dislocation slip, grain boundary sliding, and the evolution of multi-crack systems – is highly intricate. These deformation mechanisms involve both complex stochastic and deterministic mechanisms whose interactions are an area of active research and debate. Characterizing, understanding, and accurately modeling these interactions is key to advance materials science. However, the multi-scale nature of these processes and capturing the various mathematical underpinnings and complexity of the phenomena across scales, both length and time, remains an open challenge.

As experimental techniques have evolved and the volume of data generated in materials science has grown rapidly – ranging from imaging data to information encoded in language and tables to mathematical formulations – many of the scientific investigations facing the mechanics community have become rooted in the challenge of finding structure in enormous volumes of multi-modal data that contain complex and stochastic interactions. However, much of the machine learning (ML) architecture adopted by the mechanics community was initially developed for applications outside of science and engineering. These traditional machine learning approaches usually function as black boxes, meaning that they do not inherently enforce physical laws and constraints, and therefore run the risk of producing results that lack physical validity, making them less reliable for scientific discovery and engineering applications.  

To address this challenge, there is a growing need for scientific machine learning, or ‘SciML’, architectures that are capable of reasoning over multimodal and multi-scale information, incorporating relevant physical constraints for comprehensive scientific reasoning for key materials challenges. This talk will discuss the need for architectures that incorporate image analysis, textual analysis, and incorporation of mathematical constraints – that is, architectures that have solid grounding in physical principles and be explainable, with three fundamental requirements: (i) clear objectives, (ii) quantifiable evaluation, and (iii) well-defined extensibility. Specifically, we will discuss the need for SciML architectures that seamlessly integrate image analysis, textual data interpretation, and mathematical constraints to create models that are not only accurate but also physically explainable. These requirements will be discussed in the context of various mechanics problems currently under investigation.

The integration of machine learning with physics-based reasoning is a frontier in the rapidly evolving landscape of the experimental mechanics of materials. It is key to create frameworks that effectively integrate data of multiple modalities with physical validity to advance predictive an interpretable models that are scientifically robust across multiple length and time scales.

Presenting Author: Samantha Daly UCSB

Presenting Author Biography: Samantha (Sam) Daly is a Professor in the Department of Mechanical Engineering at UCSB. She received her Ph.D. from Caltech in 2007 and subsequently joined the University of Michigan, where she was on the faculty until 2016 prior to her move to UCSB. Her research interests lie at the intersection of the experimental mechanics of materials and data science. Currently, the Daly group is engaged in the development of new methods for multi-scale material characterization and application of machine learning to understand the deformation and failure of advanced structural materials. Prof. Daly is a Fellow of The American Society of Mechanical Engineers (ASME), and currently serves on the Executive Committee of the ASME Applied Mechanics Division and as Vice Chair of the U.S. National Committee of Theoretical and Applied Mechanics.

Authors:

Samantha Daly UCSB