GradDFT is a JAX-based library enabling the differentiable design and experimentation of exchange-correlation functionals using machine learning techniques.
Grad DFT is not intended as a general purpose high performance DFT code. Its domain of applicability is for training neural functionals. Accordingly, if we wish to perform production level simulations using functionals learnt in Grad DFT, we need to be able to use these functionals in high performance codes.
While learning is performed using Gaussian basis sets (and there will be some biases imparted by this), in principle, functionals are of the density so are basis set agnostic. This means that we can interface learnt models to any DFT code, really.
There are two things to do/decide upon here.
(1) Build a library interface which can be compiled and linked in with standard DFT codes in a similar spirit to Libxc. This could perhaps be an addition to Libxc itself or a a completely standalone library. Likely, we would end up with a different repo for this code than this one.
(2) What is the model export format we should use which is most seamlessly to integrate with the interface?
Grad DFT is not intended as a general purpose high performance DFT code. Its domain of applicability is for training neural functionals. Accordingly, if we wish to perform production level simulations using functionals learnt in Grad DFT, we need to be able to use these functionals in high performance codes.
While learning is performed using Gaussian basis sets (and there will be some biases imparted by this), in principle, functionals are of the density so are basis set agnostic. This means that we can interface learnt models to any DFT code, really.
There are two things to do/decide upon here.
(1) Build a library interface which can be compiled and linked in with standard DFT codes in a similar spirit to Libxc. This could perhaps be an addition to Libxc itself or a a completely standalone library. Likely, we would end up with a different repo for this code than this one.
(2) What is the model export format we should use which is most seamlessly to integrate with the interface?