analytical hessian

[lucasem1]

Dear Developers

is there a Nequip branch that can calculate the hessian analytically?

Best regards.
Luca Sementa

[Linux-cpp-lisp]

Hi Luca,

This can be achieved by adapting a copy of PartialForceOutput (https://github.com/mir-group/nequip/blob/main/nequip/nn/_grad_output.py#L160) to use torch.autograd.functional.hessian instead of torch.autograd.functional.jacobian. Please note that due to limitations of the underlying PyTorch implementations, this can be quite slow depending on your system and model. Please also note that for the same reason TorchScript deployment when torch.autograd.functional.* is used is not supported.

If you write a HessianOutput class based on PartialForceOutput, please feel free to submit a PR against the develop branch after reading CONTRIBUTING.md; I'd be very happy to merge such an implementation into nequip.

Thanks!