I am toying with the idea of using current libraries available with SCF (Self-Consistent Field) subspace iteration codes to train an ANN (Artificial Neural Network) to solve for SCF eigenvalue problems. Are there any obvious deficiencies with this idea?
Firstly, I would like to make sure that it is possible to train an ANN to approximate an algorithm, in this case, a numerical algorithm. I would hope that the ANN would provide some speedups over the original algorithm.
Since I don't have the requisite knowledge to form the machine learning problem in a suitable and original mathematical framework, I will rely on other people's numerical codes to train an ANN to do what I'm looking to accomplish. This would be the starting point for further exploration.