I like this post and think this is an important question to discuss. We recently highlighted some work on our blog (not our work!) using per-atom internal representations from NNPs for different downstream tasks—hopefully this does, as you say, provide us a way to bridge between the data-rich world of in silico data and the data-poor world of experimental measurements.
Thanks for sharing, Corin! That’s a very interesting paper (among a string of interesting works from the A. Bronstein lab) - I hope we’ll see more and more of the utility of MLIP representations across scales, indeed. As you guys work on this stuff day-to-day: do you have strong intuitions about what MLIP representations are capturing? I guess they definitely capture local structure, but do they capture some notion of possible dynamics/fluctuations, or do they capture something more like experimental uncertainty in measuring coordinates?
I don't know! That's a great question. I suspect local dynamics would be learnable but I don't know if it's intrinsically in the embeddings. I imagine that the local covalent and non-covalent environment is the main component…
I like this post and think this is an important question to discuss. We recently highlighted some work on our blog (not our work!) using per-atom internal representations from NNPs for different downstream tasks—hopefully this does, as you say, provide us a way to bridge between the data-rich world of in silico data and the data-poor world of experimental measurements.
Blog: https://rowansci.com/blog/local-atomistic-environment
Thanks for sharing, Corin! That’s a very interesting paper (among a string of interesting works from the A. Bronstein lab) - I hope we’ll see more and more of the utility of MLIP representations across scales, indeed. As you guys work on this stuff day-to-day: do you have strong intuitions about what MLIP representations are capturing? I guess they definitely capture local structure, but do they capture some notion of possible dynamics/fluctuations, or do they capture something more like experimental uncertainty in measuring coordinates?
I don't know! That's a great question. I suspect local dynamics would be learnable but I don't know if it's intrinsically in the embeddings. I imagine that the local covalent and non-covalent environment is the main component…