Accelerating phase field simulations to understand liquid-metal dealloying

Liquid metal dealloying. The initial species fields (cA and cB at t=0) are contaminated with low-amplitude random white noise (left fields). Given the chaotic nature of the dealloying process, the initial noise perturbation eventually leads to widely different solid phase fields at late time (e.g., right fields, at t=6⁢μs, after 6 million time steps).
Liquid metal dealloying. The initial species fields (cA and cB at t=0) are contaminated with low-amplitude random white noise (left fields). Given the chaotic nature of the dealloying process, the initial noise perturbation eventually leads to widely different solid phase fields at late time (e.g., right fields, at t=6⁢μs, after 6 million time steps).

When a corrosive liquid stays in contact with metal alloys for a long time, it can cause a process called dealloying. One type of this process is called liquid-metal dealloying (LMD). Scientists have created models to help understand how this happens and why the metal changes shape in complex ways. However, the equations used in these models are complicated and hard to solve using computers.

In a partnership with the Office of Nuclear Energy, scientists at Sandia have come up with a new approach called U-Shaped Adaptive Fourier Neural Operators (U-AFNO). This is a machine learning model that uses new techniques to learn from data. U-AFNO uses a special type of network called U-Nets to find and recreate important details in the physical fields. It also uses a vision transformer (ViT) to process information in a different way.

Sandia’s research shows that the U-AFNO model works better than other mixed methods when it comes to making predictions in fully auto-regressive settings, which means it can make better guesses based on previous information.

Sandia experts linked to work

  • Christopher Bonneville
  • Habib N. Najm
  • Cosmin Safta

Sponsored by

Department of Energy Office of Science logo

Publications

Bonneville, C., Bieberdorf, N., Hegde, A., Asta, M., Najm, H. N., Capolungo, L., & Safta, C. (2024) “Accelerating phase field simulations through a hybrid adaptive Fourier neural operator with U-Net backbone.” Nature/npj Computational Materials, the journal article has been accepted. https://arxiv.org/html/2406.17119v2


March 3, 2025