Closing the Reality Gap: Evolving Deep Material Networks into Robust, Multi-Physics Materials Digital Twins
Please login to view abstract download link
Realizing the promise of materials digital twins requires surrogate models that balance extreme computational efficiency with high-fidelity physics. While data-driven approaches often sacrifice interpretability for speed, the Deep Material Network (DMN) offers a physics-informed architecture rooted in micromechanical homogenization theory. This presentation details recent advancements transforming the DMN from a mechanical homogenizer into a comprehensive multi-physics digital twin engine. In this presentation, we will explore the versatility of the DMN’s fundamental building blocks, illustrating how they can be designed to simulate physics beyond mechanics, including thermo-elasto-viscoplasticity and thermal conductivity in complex architectures. We will also discuss advanced training strategies that take full advantage of training data to enhance interpretation and local field fidelity. Finally we will conclude this presentation by demonstrating how the DMN evolves into a probabilistic engine capable of forward uncertainty quantification and inverse calibration, effectively serving as a distribution of surrogate microstructures. By acting as an uncertainty-aware framework that can disentangle experimental noise and self-calibrate as new data arrives, the DMN provides the decision-support layer necessary for real-time process control and the realization of autonomous materials digital twins. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. SAND2026-15795A
