Figure 1.

Impact of residual learning for the design problem. Design problem involves predicting formation enthalpy from vector-based materials representation composed of 126 structure-derived and 145 composition-derived physical attributes on the OQMD-SC. They are trained using 9:1 random train:test splits (test set is same as validation set). Plain Network do not have shortcut connections; stacked residual network (SRNet) places shortcut connection after stacks of multiple layers; individual residual network (IRNet) leverage individual residual learning around each layer. The three subplots shows the validation error curves during training for each network; x-axis represents the training iteration (x1000) and y-axis represents the MAE. The models are implemented using TensorFlow and trained using Adam optimizer with a mini batch size of 32 and a learning rate of 1e-4 and a patience of 400 epochs (training stops if the validation error does not improve for last 400 epochs).