Bálint Máté

new preprint on solvation free energies: tl;dr: We define an interpolating density by its sampling process, and learn the corresponding equilibrium potential with score matching. arxiv.org/abs/2410.15815 with @francoisfleuret and @tristanbereau (1/n)

Now accepted at #NeurIPS2025! We basically speed up diffusion models for sampling molecular conformations by 30x :) More exciting stuff coming soon!

🚀 After two years of intense research, we’re thrilled to introduce Skala — a scalable DL density functional that hits chemical accuracy on atomization energies and matches hybrid-level performance on main group chemistry — all at the cost of a semi-local functional. ⚛️🔥🧪⚗️🧬

🚀 After two+ years of intense research, we’re thrilled to introduce Skala — a scalable deep learning density functional that hits chemical accuracy on atomization energies and matches hybrid-level accuracy on main group chemistry — all at the cost of semi-local DFT. ⚛️🔥🧪🧬

@francoisfleuret math.umd.edu/~tadmor/ki_net…

Serious question: Von Neumann is always described as an absolute genius. What did he do that would put him anywhere close to Turing, Godel, or anyone on the Solvay conference pic?

@eeevgen @francoisfleuret @tristanbereau In this particular experiment, we ran MCMC for 40k steps and dropped the first 10k as burn-in. I see your concern with direct MCMC, and I guess it depends on the target. We could get away with it for this simple target (only LJ interactions, particles are not too densely packed).

@balintmt @francoisfleuret @tristanbereau Sorry, I still have a bit not understanding what is typical budget for MCMC from target to get good results?

New preprint on performing thermodynamic integration along the trajectories of a denoising diffusion model: arxiv.org/abs/2406.02313. With @francoisfleuret and @tristanbereau (1/n)

@eeevgen @francoisfleuret @tristanbereau Hi, thanks for the questions! 1. Yes, we do need samples at t=1. 2. Do you mean something like annealed MCMC? If so, we don't do that, just plain MCMC at the target potential.

Hi, very nice sequence of papers! I have couple of questions just to be sure I understand correctly. Let’s consider simple case, where one bridge side (prior) is simple distribution at t=0 and target at t=1. 1. Do I understand correctly, that to learn “neural” path you method still need to have samples at t=1? 2. If yes, is this true that to get them with MCMC you will still use some path, like usual geometric say? What is then typical budget of training samples to train neural path?

Finally, we also look at what happens if we predict the hydration free energy of methane using the potential that was trained on water (and vice versa). (10/10)

The approach is tested on the estimation of hydration free energies of rigid water and methane (LJ + Coulomb interactions). We find good agreement with experimental reference values. (9/n)

new preprint on solvation free energies: tl;dr: We define an interpolating density by its sampling process, and learn the corresponding equilibrium potential with score matching. arxiv.org/abs/2410.15815 with @francoisfleuret and @tristanbereau (1/n)

That’s a wrap! 🤩🙏🏻 Thank you everyone for making this possible and so enjoyable! Stay tuned for a second edition in a couple of years! 😉 #ML4PhysChem #workshop #generativeAI #AI4Science

To build the next generation of intelligent agents, developing efficient world models is essential. We introduce Δ-IRIS, an agent that learns behaviors by imagining millions of trajectories in its world model. Paper: arxiv.org/abs/2406.19320 Code: github.com/vmicheli/delta… 🧵👇

To validate all this, we compare the estimates of the average density and the excess chemical potential to grand canonical MC simulations. (6/n)

Given the estimates of the canonical free energies at fixed particle count we end up with a grand canonical sampler at any choice of the chemical potential. (5/n)