Proposed next steps | Several possible directions: Start optimisation / solve least squares problem direct. Are we stuck in a local minima? Don't necessarily need to benchmark produced charges, but would be good to see how different they are.
Ethers seem to be the most problematic. Need to analyse this further. Perform analysis of the gradients during fitting - are ethers trying to constrain the parameters differently than the rest of the data? Split the analysis into polar and non-polar. Electric field expected to yield better results for polar molecules over non-polar. Need to determine if the test set is balanced too much towards non-polar. Potentially look more into fitting to ESP rather than electric field?
Maybe we need to expand the sets to include some nitrogen data to see the real benefits of the refits?
| @Simon Boothroyd | |