batch

This keyword sets the size of each batch used during the optimization procedure. The syntax is:

batch <batch_size>

Here, <batch_size> sets the batch size \(N_\mathrm{bat}\), which must satisfy \(N_\mathrm{bat}\geq 1\) and defaults to \(N_\mathrm{bat}=1000\).

In principle one can train against the entire training set during every iteration of the optimization procedure (equivalent to \(N_\mathrm{bat}\) being identical to the number of structures in the training set). It is, however, often beneficial for computational speed and potentially necessary for memory reasons to consider only a subset of the training data at any given iteration. \(N_\mathrm{bat}\) sets the size of this subset.

Usually, training sets with more diverse structures require using larger batch sizes to achieve maximal accuracy. In many cases, a batch size between \(N_\mathrm{bat}=100\) and \(N_\mathrm{bat}=1000\) should be sufficient to achieve good results. If you have a powerful GPU (such as a Tesla A100), you can use a large batch size (such as \(N_\mathrm{bat}=1000\)) or the full-batch (\(N_\mathrm{bat}\geq\) number of configurations in the training set). If you use a small batch size for a powerful GPU, you will simply waste the GPU resources. If you have a weaker GPU, you can use a smaller batch size.

If you observe oscillations in the RMSE values for energies, forces, and virials even for generations \(\gtrsim 10^5\), it is a sign that the batch size is too small.