Once you can simulate any physical system cheaply and effectively. You can rapidly start to explore a wider range of the design space then you can with experiment alone.

We know all of the fundamental science we need to perform these simulations. They are just too expensive.

As Dirac said: ‘the underlying physical laws necessary for the mathematical theory of ... the whole of chemistry are thus completely known, and the difficulty is only that the exact application of these laws leads to equations much too complicated to be soluble.’

I believe ML will allow us to solve these equations with enough accuracy. I think it is a tool for connecting scales, i.e., finding good reduced dimensional descriptions of high dimensional systems.

This is the age old problem that every physical modeler grapples with. It is a problem for climate scientists and quantum chemists. And it’s not a coincidence that climate modeling is being changed by ML techniques just as much as molecular simulation.

@TimothyDuignan Hard disagree. Predict next high temperature superconductor. Ultimately, what matters is experimental discovery that allows to verify theoretical models and build new ones. If we accelerate theory, experiment will be even more of a bottleneck then it is now

@Sergei_Imaging Yeah superconductors is definitely one I’m least confident about. Experiment will always still be king in the end. I’m just hoping we can move away from needing so much trial and error and guess work.

@TimothyDuignan yeah I disagree. I predict that atomistic/molecular simulations will be as important for continuum simulations of macroscopic matter w/ e.g. finite elements in 10 years as they are now: not at all.