Fig. 3. Architecture and performance of the machine-learning (ML) model.
a Architecture of the residual network (ResNet). b Architecture of the residual block. BN stands for batch normalization and ReLU the rectified linear unit. c Effects of four boundary conditions (BCs) on the ResNet performance. d Validation loss versus epochs showing the training progress of plain convolutional neural network (CNN) and ResNet for different numbers of convolutional layers. e Final validation loss for the cases in (d). f Effects of training groups on the ResNet performance with 51 convolutional layers. g, h Density scatter plots of the true versus ML-predicted values of the coordinate (g) (x, y) and (h) (z). The color indicates the relative point density. i Comparison of the ground-truth (gray) and ML-predicted (colored) shapes for datapoints randomly picked from the validation set. The color contour visualizes the prediction error Δri of each deformed plate.