FermiNet: State of the art approx of molecular orbitals
1–10 of 19 posts
Re: FermiNet: State of the art approx of molecular orbitals
#2I’m really excited to read the paper for this! I’ve been pondering for a while now how well DNN’s would model complex QC wave functions. Much high level quantum chemistry research involves human theoreticians finding “quirks” and other features which allow computing specific properties like say estimating quantum tunneling in photosynthesis. These involve high level symmetries or various green’s functions which help correlate to different domains. But most simulation only use DFT (roughly discretized orbitals) or other “naive” optimization methods over a pure numerical solution of the ache of infer equations.
Still running true ab initio QM simulations of a few atoms can take months on a single computer (never had a chance to run simulations on a cluster or GPU). DNN’s however have the ability to find higher dimensional patterns difficult for humans to find but which could significantly speed up QM simulations.
Currently doing QM simulations of chemical reactions for any number of reactions is in feasible but if work like FermiNet could make it feasible for small teams to simulate more complex chemical reactions it could open up an entire field of chemical/industrial processes to startups. As in you could reasonably simulate chemical processes sufficiently to optimize current process or find entirely novel reactions. This would significantly reduce the capital expenses most research in these areas require.
In short if I were a VC I would be _very_ keen I’m watching this field. There tremendous value hidden behind this general problem.
Re: FermiNet: State of the art approx of molecular orbitals
#3I’m really excited to read the paper for this! I’ve been pondering for a while now how well DNN’s would model complex QC wave functions. Much high level quantum chemistry research involves human theoreticians finding “quirks” and other features which allow computing specific properties like say estimating quantum tunneling in photosynthesis. These involve high level symmetries or various green’s functions which help…
> In short if I were a VC I would be _very_ keen I’m watching this field. There tremendous value hidden behind this general problem.
As someone who used to be in this very space and even tried to get a startup off the ground based on it, I can tell you with absolute certainty that this will lead absolutely nowhere.
The short of it is that literally no business will accept data that is generated this way until someone shows that every neural net model trained in this way produces solutions that are mathematically equivalent to a validated method.
At best it might be used as a filter step in some pipeline, but that's not going to have much of an effect, and certainly not something on which to bet the success of a startup.
Re: FermiNet: State of the art approx of molecular orbitals
#4I’m really excited to read the paper for this! I’ve been pondering for a while now how well DNN’s would model complex QC wave functions. Much high level quantum chemistry research involves human theoreticians finding “quirks” and other features which allow computing specific properties like say estimating quantum tunneling in photosynthesis. These involve high level symmetries or various green’s functions which help…
To me there are two ways to look at this. Either the laws of physics as we have them are too general when compared to what can be encountered in reality (in other words, actual reality is simple enough that a Turing machine approximated by logic circuits can manage well enough at finding a description) or, models found by fitting data are too specialized, ignoring subtleties not captured by loss functions acting on the data they were trained on.
Any physical model gained by fitting data which purports to be faster than those based on a computable approximation of the laws of physics is so constrained. There is room to maneuver however. If the model being replaced has a limited and known domain of applicability due to approximations made for tractability, a fitted model with suffciently large capacity and expressiveness will for sure improve things.
It's just that it's unlikely to be generally applicable without violating what we know about physics, which is why I am skeptical of the latter part of your post.
Re: FermiNet: State of the art approx of molecular orbitals
#5I’m really excited to read the paper for this! I’ve been pondering for a while now how well DNN’s would model complex QC wave functions. Much high level quantum chemistry research involves human theoreticians finding “quirks” and other features which allow computing specific properties like say estimating quantum tunneling in photosynthesis. These involve high level symmetries or various green’s functions which help…
> In short if I were a VC I would be _very_ keen I’m watching this field. There tremendous value hidden behind this general problem. As someone who used to be in this very space and even tried to get a startup off the ground based on it, I can tell you with absolute certainty that this will lead absolutely nowhere. The short of it is that literally no business will accept data that is generated this way until someone…
all I'd ask for is a few blind predictions on some reasonably interesting molecules, where no existing method can make an accurate prediction without a very expensive calculation (QM on supercomputers is common, you could easily generate the energies for a bunch of modest molecules from a diverse set).
I say that as somebody who has evaluated VC pitches for O(n) approximations of QM as a startup idea.
Re: FermiNet: State of the art approx of molecular orbitals
#6I’m really excited to read the paper for this! I’ve been pondering for a while now how well DNN’s would model complex QC wave functions. Much high level quantum chemistry research involves human theoreticians finding “quirks” and other features which allow computing specific properties like say estimating quantum tunneling in photosynthesis. These involve high level symmetries or various green’s functions which help…
To me there are two ways to look at this. Either the laws of physics as we have them are too general when compared to what can be encountered in reality (in other words, actual reality is simple enough that a Turing machine approximated by logic circuits can manage well enough at finding a description) or, models found by fitting data are too specialized, ignoring subtleties not captured by loss functions acting on t…
>Any physical model gained by fitting data which purports to be faster than those based on a computable approximation of the laws of physics is so constrained.
Constrain the model so that there aren't any superfluids, semiconductors, plasmas, metals or Bose-Einstein condensates and you can still simulate any medicine I know of.
Re: FermiNet: State of the art approx of molecular orbitals
#7Earlier quoted context omitted.
To me there are two ways to look at this. Either the laws of physics as we have them are too general when compared to what can be encountered in reality (in other words, actual reality is simple enough that a Turing machine approximated by logic circuits can manage well enough at finding a description) or, models found by fitting data are too specialized, ignoring subtleties not captured by loss functions acting on t…
> Any physical model gained by fitting data which purports to be faster than those based on a computable approximation of the laws of physics is so constrained. Constrain the model so that there aren't any superfluids, semiconductors, plasmas, metals or Bose-Einstein condensates and you can still simulate any medicine I know of.
Lithium?
Re: FermiNet: State of the art approx of molecular orbitals
#8Earlier quoted context omitted.
> Any physical model gained by fitting data which purports to be faster than those based on a computable approximation of the laws of physics is so constrained. Constrain the model so that there aren't any superfluids, semiconductors, plasmas, metals or Bose-Einstein condensates and you can still simulate any medicine I know of.
Lithium?
Yeah, take metals out of the equation and it probably a lot more likely.
You are still going to need things like: Sodium, Potassium, Magnesium, Iron, etc.
Re: FermiNet: State of the art approx of molecular orbitals
#9Earlier quoted context omitted.
> In short if I were a VC I would be _very_ keen I’m watching this field. There tremendous value hidden behind this general problem. As someone who used to be in this very space and even tried to get a startup off the ground based on it, I can tell you with absolute certainty that this will lead absolutely nowhere. The short of it is that literally no business will accept data that is generated this way until someone…
all I'd ask for is a few blind predictions on some reasonably interesting molecules, where no existing method can make an accurate prediction without a very expensive calculation (QM on supercomputers is common, you could easily generate the energies for a bunch of modest molecules from a diverse set). I say that as somebody who has evaluated VC pitches for O(n) approximations of QM as a startup idea.
This sounds like a good way to approach it. If there’s a system that is currently impossible to make decent predictions on then any partially accurate result could be beneficial. I’d be suspicious of someone claiming O(n) on QM however. They’d have to be very simple estimates, not sure how useful they’d be.
To your parents comment, it’d depend heavily on the molecules and systems under study and startup goals.
> The short of it is that literally no business will accept data that is generated this way until someone shows that every neural net model trained in this way produces solutions that are mathematically equivalent to a validated method.
This would seem to be wrong approach this early on...
> At best it might be used as a filter step in some pipeline, but that's not going to have much of an effect, and certainly not something on which to bet the success of a startup.
I don’t think a startup based on providing ‘QM simulation as a service’ would work very well and be fraught with issues. However, many industries could make significant usage of a pipeline filtering possible solutions which could be validated experimentally or with more traditional methods.
IMHO, to work with this a startup would need to be a vertically integrated company solving a specific class of problems (say batteries). Even then the current State of the Art still seems a few years off before I’d want to do a startup in the area, though it’s much closer with the recent results.
Re: FermiNet: State of the art approx of molecular orbitals
#10Earlier quoted context omitted.
> Any physical model gained by fitting data which purports to be faster than those based on a computable approximation of the laws of physics is so constrained. Constrain the model so that there aren't any superfluids, semiconductors, plasmas, metals or Bose-Einstein condensates and you can still simulate any medicine I know of.
Lithium?
I meant metals in the solid-state physics sense, with the oceans of electrons and band gaps and stuff. You can keep your monatomic ions (which are crucial for biology in many ways).