Accelerated prediction of microstructure evolution

side-by-side image of a phase field simulation that took 7 hours and a machine learning prediction that took 15 seconds
Side-by-side images show (left) a phase field simulation that took seven hours and (right) a machine learning prediction that took 15 seconds

CMI scientists at Lawrence Livermore National Laboratory conducted this research.

Achievement:
A deep learning toolkit for predicting 3D microstructure evolution has been developed and implemented for the first time

Significance and Impact:

  • A machine learning toolkit that is orders of magnitude more efficient at predicting microstructure evolution
  • Toolkit delivers accurate predictions of alloy solidification in 3D with phase field models or direct atomistic simulations are extremely expensive

Details and Next Steps:

  • Direct 3D simulations of alloy solidification are very expensive
  • Machine learning with convolutional and recurrent neural networks were trained to learn the 3D time evolution
  • Periodic 3D convolution and point group symmetry implemented