Surrogate models for microstructure simulation

line graph — Comparing high-fidelity and surrogate models: Dislocation dynamics in 2D showing good agreement between the different approaches.

CMI researchers from Lawrence Livermore National Laboratory conducted the research for this highlight.

Innovation
Developed neural-network (NN) based surrogate models for microstructure simulation with up to 10,000X speedup

Achievement
Graphical NN models retain accuracy with much smaller computational cost in simulating dislocation motion as well as nucleation from melt

Significance and impact

High-fidelity microstructure simulations such as molecular dynamics (MD) are very expensive.
Faster machine-learning surrogate models allow access to much larger length- and time-scales while maintaining accuracy.
One of the most promising approaches for modeling force fields.

Hub Level Goal Addressed

Increasing the speed of materials discovery and deployment
Accelerated microstructure simulation for alloys and magnets

image of line graph showing red and blue — Comparing high-fidelity and surrogate models: Precipitation volume in a binary mixture through nucleation and growth (MD vs ML)