Fall 2020 Seminar Series: Gabor Csanyi
Host: Tim Mueller
I will make the somewhat bold claim that over the past 10 years, a new computational task has been defined and solved: this is the analytic fitting of the Born-Oppenheimer potential energy surface as a function of nuclear coordinates under the assumption of medium-range interactions, out to 5-10 Å. The resulting potentials are reactive, many-body, reach accuracies of a few meV/atom, with costs that are on the order of 1-10 ms/atom. This leaves the following challenges for ML potentials: treatment of long range interactions in a nontrivial way, e.g. consistency of treatment of open and periodic boundary conditions, environment dependent multipolar description, excited states (adiabatic surfaces), magnetism. We also still need a “shakedown” of the details among various approaches (neural networks, kernels, polynomials), and more standard protocols of putting together the training data. Tradeoffs between system- (or even project-) specific datasets and potentials vs. more general potentials will be ongoing. Further afield, another interesting question is what part of this technology can be reused to fit analytic surrogate models of *electronic* functions and functionals, such as reduced Hamiltonians, Green’s Functions, density matrices etc, not to mention many body wave functions. There have been forays in this direction already.
Zoom Seminar Info:
Meeting ID: 982 0915 3548