Sandia Labs FY22 Laboratory Directed Research & Development Annual Report

USING MACHINE LEARNING TO CREATE RAPID STRONGLINK MECHANISMS CAD-TO-SIMULATION-READY MODELS. Computer aided design (CAD) to simulation

also dramatically impacted using a new multi-agent reinforcement learning algorithm that learns a complex procedure for dimensionally reducing a 3D CAD assembly to a simulation-ready set of

workflows for nuclear deterrence (ND) have shown dramatic performance improvements with ML. This work targets some of the most inefficient, tedious,

CAD model (left) with multiple fasteners (right), rapidly reduced to simulation-ready state using new ML methods.

and error-prone bottlenecks using new ML-based methods. Common mechanisms such as fasteners and springs can now be quickly identified and reduced to simulation-ready proxies, significantly reducing the need for most human interactive geometry and meshing operations. Also developed is a new dynamic, in-situ classification tool that can be used as the basis for a common sharable knowledge base, capitalizing on the expertise of experienced engineers within a community of analysts. The time analysts need to prepare shell-based Finite Element Analysis models to simulate critical ND transportation systems was

interconnected sheet bodies. This project has led to two Technical Advances and a patent submission. Presentations were made at various conferences and symposia, including the Siemens Geometry and Meshing Lecture Series and the 16 th and 17 th U.S. Congress on Computational Mechanics. Additionally, it won the Best Technical Presentation award at the 2022 Society for Industrial and Applied Mathematics International Meshing Roundtable (IMR), and a published, peer-reviewed paper will be presented at the 2023 IMR in Amsterdam, Netherlands. (PI: Steven James Owen)

37

LABORATORY DIRECTED RESEARCH & DEVELOPMENT