Feature Request: Automatic differentiation support for spin wave models

[elliottperryman]

Bonjour!

I have what I believe to be a burdensome but novel feature request. I propose that the observable dispersion curves computed by Takin be automatically differentiated with respect to the parameters of the spin wave model (exchange coupling, anisotropy, etc). I don't believe the instrument resolution would need to be considered, although both the energy and intensity of the dispersion curve would need to be differentiable. With these two, I believe the gradient of the log likelihood of the data can be computed with respect to the model parameters.

Why this feature would be useful:

• Current model-data fits are done using global fit minimization. A gradient w.r.t. the model parameters would allow much better fits to the data. As I understand it, current methods involve fixing some parameters, tuning others, and global minimization searches. A Newton's method search would be much faster I believe.
• Fit sensitivity can be then estimated using the Hessian of the data with respect to the model parameters.
• My own interests. This would make using gradient based MCMC samples of model parameters possible, as well as making autonomous experimentation methods possible.

The risks and costs of this feature:

• This could be completely infeasible and too much work!

Novelty: To my knowledge, no other tool has this capability. This I believe would be a novel and useful advancement.

I know this is a lot of work. I think this is a big risk, hopefully big reward project, so perhaps some other inputs would be useful. Maybe I am the only one that wants this.

Thanks for reading,
Elliott

[tweber-ill]

Hi Elliott, thanks, this actually sounds doable, I'll add it to the TODO list.