Fig. 10: CNN model parameter direct inspection example.

The authors design a simple 3D CNN to predict effective elastic properties of high contrast composites from synthetic microstructures. a 3D microstructure data, colored by material phase. b CNN filter weights, colored according to their sign (near zero values are rounded to zero). Positive weights in the CNN filters (red) suggest a preference for a structural pattern and negative weights (blue) penalty deviations. A comparison between input microstructures and CNN filter weights shows that CNN filters can learn simplified characteristics of the input microstructure. Figure reprinted from ref. 86 with permission. Copyright 2018 Elsevier.