Fig. 4: Recursive training performance via a quilting strategy. | npj Computational Materials

Fig. 4: Recursive training performance via a quilting strategy.

From: Deep material network via a quilting strategy: visualization for explainability and recursive training for improved accuracy

Fig. 4

a Example of DMN recursive training showing how to construct and initialize a N = 4 DMN from a N = 2 DMN in 2D-microstructure training. After training the shallower network, the optimized network is used for quilting an array network. The array network is initialized with a quarter of each weight of the shallow optimized DMN unit cells and the known normal vectors. b Recursively training of deeper and deeper networks (N = 1, 3, 5, and 7) for 4000 epochs by quilting patches of optimized DMN unit cells to initialize deeper networks at 4001, 8001, and 12,001 epochs. Orange arrows indicate the recursive training steps, and blue arrows the quilting steps. c Evolution of the training error achieved via recursive training compared to the error obtained from random initialization of the DMN parameters. Insets show the online predictions using Norton viscoplastic constitutive relationship for networks with different depths.

Back to article page