Figure 31.

Example of early stopping. For this example, we trained a NN (three hidden layers with ReLU activation and 250, 100, and 10 neurons, respectively, followed by linear activation in the output layer) using the Adam optimizer,³⁴⁵ to predict the CO₂ uptake for structures in the database from Boyd et al.¹³ using RACs and pore geometry descriptors as features. We can observe that after approximately 43 epochs (dotted vertical line) the training error still decreases, whereas the validation error starts to increase again.