I concur that using a neural net as a sub-block of the simulator is probably a bad idea, because NN are approximations and summing them over simulation-time is only going to result in worse approximations.
I'm quite bullish on the other hand for using NN in chemistry. If you use NN at a higher level (for example predicting which reactions will happen and finding pathways), it makes a lot more sense, it's akin having an experienced chemist. A tool like Chematica (database of chemical reactions to find pathways which got bought by Merk) can almost certainly benefit from neural networks where you can try to find similarly-shaped molecules that will have a similar behavior but be cheaper to produce.
I'm also bullish, on the use of all the NN machinery, for the use of simulators. NN frameworks are one way to be able to use efficiently large amount of compute. In fact there are kind of dual problems : In machine learning you try to minimize some energy function while in Hamiltonian Physics simulation you try to keep the energy function constant.
Making approximations in these equations (for example in the way to represent a field, summing in a different order, neglecting zeros), can often result in faster computation for the same accuracy. In fact, if you write the simulator as a Monte-Carlo sum, you can train a NN controller using the trajectory approximation, to do what you want to do, folding the exact simulation computation time inside the training loop of the NN.
I'm quite bullish on the other hand for using NN in chemistry. If you use NN at a higher level (for example predicting which reactions will happen and finding pathways), it makes a lot more sense, it's akin having an experienced chemist. A tool like Chematica (database of chemical reactions to find pathways which got bought by Merk) can almost certainly benefit from neural networks where you can try to find similarly-shaped molecules that will have a similar behavior but be cheaper to produce.
I'm also bullish, on the use of all the NN machinery, for the use of simulators. NN frameworks are one way to be able to use efficiently large amount of compute. In fact there are kind of dual problems : In machine learning you try to minimize some energy function while in Hamiltonian Physics simulation you try to keep the energy function constant.
Making approximations in these equations (for example in the way to represent a field, summing in a different order, neglecting zeros), can often result in faster computation for the same accuracy. In fact, if you write the simulator as a Monte-Carlo sum, you can train a NN controller using the trajectory approximation, to do what you want to do, folding the exact simulation computation time inside the training loop of the NN.