xnn.ffnn.models.reaxff.guarded_exp

xnn.ffnn.models.reaxff.guarded_exp(x)

exp with the argument capped at 40.

Every exponential this guards appears inside a saturating ratio (the ReaxFF correction and coordination functions), where capping the argument changes the result by less than 1e-17 – but keeps single and double backward passes finite in float32 when training drives the model through strongly over/under-coordinated configurations.

Parameters:

x (Tensor) – Exponent.

Returns:

exp(min(x, 40)).

Return type:

Tensor