Turns out the dimensionality/accuracy tradeoff depends on the property and the number of atoms - but only very little on the actual molecule (see the low variance in this plot).
~4N coordinates describe a molecule. Can we make it work with fewer?
Symmetries allow to reduce this - but some properties are harder than others.
@Alibanjafar uses quantum alchemy to quantify the property-dependent limit.
➡️ https://t.co/nOxO4wdcmN
Do you want to become a PhD student in computational chemistry? Are you curious how to consider many systems at once rather than one-by-one? Have you heard about machine learning or perturbation theory? Do you know any programming language? Apply here: https://t.co/FOJ31R0TC6
Picture: this is what you want in a perturbative treatment: in this case, the more expensive a term the less relevant on average! A happy little coincidence :)
New work on increasing photoelectron circular dichroism strength with Quantum Alchemy.
Perturbations work despite strong coupling of degrees of freedom and identifies key driving forces.
https://t.co/umgu6HMsEo work with Anton Artemyev, Boris Lagutin, Philipp Demekhin and me.
Finding conservation laws and analytical expressions for them with Kernel Ridge Regression - but with unknown labels ;)
Wonderful work from Meskerem Mebratie in collaboration with Rüdiger Nather, Werner Seiler and me, all @uni_kassel
We are #hiring a #postdoc to work on differentiable #compchem with Quantum Alchemy: https://t.co/caTDK856fU
Topics revolve around using response functions and (alchemical) derivatives to model and predict electronic structure properties.
@EugeneVinitsky@Andrew_S_Rosen A much less elegant workaround than Andrew's solution in the meantime might be a small script running directly in the cluster which periodically checks, pulls and submits. We use a similar setup to automatically distribute jobs across multiple clusters.
Related previous work: https://t.co/iZFph57yQ3 where predicting UV-VIS spectra faced the same issue: the search space is so densely populated, that many molecules yield almost identical spectra. Predicting multiple properties is probably more efficient.
Using 13C and 1H NMR shifts together for structure identification requires two orders of magnitude fewer training data than sticking to 13C alone, since the permissible error of the machine learning prediction increases.
Exciting work with @Dom1Lemm and @ProfvLilienfeld.
Pumped about our first #ML paper in @digital_rsc on label noise in #ChemicalSpace: "Impact of noise on inverse design: The case of NMR spectra matching" https://t.co/EVz1N4TLiU
With @ferchault and @Dom1Lemm Elucidation success depends on noise level and training set size!
1/n Excited to share our preprint (@ferchault , Konstantin @KKarandashev): "Understanding Representations by Exploring Galaxies in Chemical Space"! 🌌 A Monte Carlo approach to probe chemical feature spaces without exhaustive enumeration or ML training.🚀
https://t.co/PKgaHIn8WC
We're hiring again! Funded PhD position at @uni_kassel 🇩🇪🇪🇺 on derivatives in chemical space and language models. The group uses differentiable chemistry and quantum alchemy for design of materials and molecules. Come and join us! #compchem#chemtwitter
https://t.co/Bmdve7izT6
@KelvinIdan @JoshRackers@hbdft2008@ProfvLilienfeld Maybe robots for chemistry / chenputation is going to give us much more experimental data at lower cost. But probably not at lower cost than some (DF-)DFT..
It was good fun to write that review with @hbdft2008 and @ProfvLilienfeld!
➡️ML is data-limited, and DFT likely is the way out
➡️Data is too scarce, e.g. for excited states, non-equilibrium geometries, charged systems and more elements
There's Plenty of Room at the Bottom😉
Nice review by Bing Huang, @ferchault, and @ProfvLilienfeld on just how important DFT is to the future of AI for molecules.
Still so many open questions (especially in biophysics) but quantum chemistry is definitely a fundamental part of the equation!
https://t.co/xgzU53YvVn
this paper's nuts. for sentence classification on out-of-domain datasets, all neural (Transformer or not) approaches lose to good old kNN on representations generated by.... gzip https://t.co/6eZiXlJxOX
@JoshRackers Also: If matter is indeed nearsighted (as Kohn suggested), then the dimensionality of CCS would be bound: one can only fit so many atoms in the neighborhood of another one. Then squeezing many discrete points into a finite space makes that space almost continuous.
@JoshRackers Yes, indeed! Curious thing: we can interpolate between elements (QM is ok with fractional nuclear charges) even though we can only observe integer realizations. That change is often analytical (https://t.co/esaAJAWaxA).