force_delta

This keyword can be used to bias the loss function to put more emphasis on obtaining accurate predictions for smaller forces. The syntax is:

force_delta <delta>

where <delta> sets the parameter $\delta$, which must satisfy $\delta \ge 0$ eV/A and defaults to $\delta =0$ eV/A (i.e., no bias).

When $\delta =0$ eV/A, the loss term associated with the forces is proportional to the RMSE of the forces:

$$\sqrt{\frac{1}{3N}\sum _{i=1}^N{({\mathit{F}}_i^{\mathrm{NEP}}-{\mathit{F}}_i^{\mathrm{tar}})}^2}$$

When $\delta >0$ eV/Å, this expression is modified to read:

$$\sqrt{\frac{1}{3N}\sum _{i=1}^N{({\mathit{F}}_i^{\mathrm{NEP}}-{\mathit{F}}_i^{\mathrm{tar}})}^2\frac{1}{1+\Vert {\mathit{F}}_i^{\mathrm{tar}}\Vert /\delta }}$$

In this case, a smaller $\delta$ implies a larger weight on smaller forces.