failures, solutions, tips about autoforce

[amirhajibabaei]

Noise
The noise parameter (model.noise or calc.model.noise) has the default value of 0.01.
Reducing the noise often increases the fitness of the model on energies but also worsens the forces.
Also, too small noise could lead to an instability (large jump after update, the cause of which is not clear yet).
But in one case including the hydrogens and OH bond formation (with very fast and small energy oscillations) the same instability was fixed by reducing the noise to 0.0001.
Therefore it depends on the physical system.
In general, any case including hydrogen needs special treatment.

[amirhajibabaei]

Fixes for above issues:

"rattle" for random displacements -> forces nonzero.

Noise scale is now optimized.

The regression algorithm operates in 3 stages: optimize with forces, minize energy residuals using per-atom-type, regress using forces and energy residuals.

The jumps were related to descriptor.
When new atom entered the LCE a new feature vector is created.
For enforcing norm > 0, the zeros element of this vector was defined as finite number (>0) which caused the discontinuity.
This issue is fixed.