Figure 2.

Performance of deep learning models of different depths in model architecture. The models are trained and tested on the lowest DFT-computed formation enthalpy of 256, 622 compounds. Here, we present the impact of depth of architecture for one sample split from our ten-fold cross validation. (a) Shows the mean absolute error (MAE) on the test dataset of 25, 662 compounds with unique compositions at different epochs for one split from the cross validation. The DNN models keep learning new features from the training dataset with the increase in the number of layers up to 17 layers, after which they begin to slowly overfit to the training data. (b) Shows the MAE for different depths of deep learning model architectures and also illustrates mean absolute error of the best performing conventional ML model trained using physical attributes computed on the same training and test sets. The deep learning model starts outperforming the best performing conventional ML model with an architecture depth of 10 layers, achieving the best performance at 17 layers, we refer to the best performing DNN model as ElemNet. The detailed architecture for ElemNet is available in the Method section.