A new minimization procedure can be implemented to make the code less dependent on the input step given by the user.
It would be great to have a modified line minimization to optimize the step.
The ideal behavior would be that: The kong-Liu ratio never decreases in one step more than 10% of its previous value (avoid too few steps that do not exploit the importance sampling), but it chooses within this constrain the biggest step with the scalar product between gradients in two successive iterations positive.
The only nontrivial aspect is the choice of how to deal with the two different minimizations (structure and dynamical matrix)
A new minimization procedure can be implemented to make the code less dependent on the input step given by the user. It would be great to have a modified line minimization to optimize the step. The ideal behavior would be that: The kong-Liu ratio never decreases in one step more than 10% of its previous value (avoid too few steps that do not exploit the importance sampling), but it chooses within this constrain the biggest step with the scalar product between gradients in two successive iterations positive.
The only nontrivial aspect is the choice of how to deal with the two different minimizations (structure and dynamical matrix)