4 ms·

Trying to predict where electrons go in a molecule, but using a classical model. This can be done with supervised machine learning - you can use quantum mechani

Trying to predict where electrons go in a molecule, but using a classical model. This can be done with supervised machine learning - you can use quantum mechanics to get lots of labeled data - but it's a tough problem, because chemical physicists have very high standards for accuracy.

The_rationalist 6y ago
Could hidden variables theories help simulations in some ways? E.g performance
lambdatronics 6y ago
What's the difference between your approach and density-functional theory?
- comicjk 6y ago
  Density functional theory is still quantum mechanics, but operating on the expected electron density itself, rather than the many-body problem of all the electrons. It's pretty good, but not fast - around a CPU-minute for a medium-sized molecule. I'm working on approximating the electron density using just the nuclei positions and some neural networks. The throughput is tens of thousands of times higher than for DFT. But since the model itself contains little or no physics, the training data has to be very clean and complete.
  - FridgeSeal 6y ago
    Would you be able to leverage any of the work that's being done around so called "neural ODE's"? The Julia community seems to be doing some really cool at the intersection of the 2 fields and it seems like it could be useful if you've already got some kind of pre-existing model/structure to hang the ML part off: * https://julialang.org/blog/2019/01/fluxdiffeq/ https://julialang.org/blog/2019/01/fluxdiffeq/ * https://mitmath.github.io/18337/lecture15/diffeq_machine_learning https://mitmath.github.io/18337/lecture15/diffeq_machine_lea...
c1ccccc1 6y ago
Sounds very cool! I'd be interested in looking at the code if you're up for sharing it. Are you computing the labels using DFT, or with a more exact method?
- comicjk 6y ago
  I work at Schrodinger Inc, collaborating with the people who make TorchANI (https://aiqm.github.io/torchani/ https://aiqm.github.io/torchani/). This currently predicts only molecular energies in the public codebase, but I expect they will add electron density using the same framework. Currently I am using hybrid DFT (wB97X-D) with plans to move up to wB97M-V and possibly Quantum Monte Carlo. The TorchANI group likes wB97X and up-training with DLPNO-CCSD(T).

The_rationalist 6y ago

Could hidden variables theories help simulations in some ways? E.g performance

lambdatronics 6y ago

What's the difference between your approach and density-functional theory?

comicjk 6y ago
Density functional theory is still quantum mechanics, but operating on the expected electron density itself, rather than the many-body problem of all the electrons. It's pretty good, but not fast - around a CPU-minute for a medium-sized molecule. I'm working on approximating the electron density using just the nuclei positions and some neural networks. The throughput is tens of thousands of times higher than for DFT. But since the model itself contains little or no physics, the training data has to be very clean and complete.
- FridgeSeal 6y ago
  Would you be able to leverage any of the work that's being done around so called "neural ODE's"? The Julia community seems to be doing some really cool at the intersection of the 2 fields and it seems like it could be useful if you've already got some kind of pre-existing model/structure to hang the ML part off: * https://julialang.org/blog/2019/01/fluxdiffeq/ https://julialang.org/blog/2019/01/fluxdiffeq/ * https://mitmath.github.io/18337/lecture15/diffeq_machine_learning https://mitmath.github.io/18337/lecture15/diffeq_machine_lea...

c1ccccc1 6y ago

Sounds very cool! I'd be interested in looking at the code if you're up for sharing it. Are you computing the labels using DFT, or with a more exact method?

comicjk 6y ago
I work at Schrodinger Inc, collaborating with the people who make TorchANI (https://aiqm.github.io/torchani/ https://aiqm.github.io/torchani/). This currently predicts only molecular energies in the public codebase, but I expect they will add electron density using the same framework. Currently I am using hybrid DFT (wB97X-D) with plans to move up to wB97M-V and possibly Quantum Monte Carlo. The TorchANI group likes wB97X and up-training with DLPNO-CCSD(T).