ML+KMC and neural nets tackle surfaces

Surface simulations often fail for a simple reason: the interesting events are rare, and brute-force calculations spend most of their time waiting. Kinetic Monte Carlo is the shortcut that jumps from one event to the next instead of tracking every vibration. (pmc.ncbi.nlm.nih.gov) That method is widely used for diffusion, crystal growth, and heterogeneous catalysis on surfaces, where atoms or molecules hop between sites and react at very different rates. The hard part is assigning reliable rates to every possible hop and reaction, especially once nearby adsorbates start changing the barriers. (pmc.ncbi.nlm.nih.gov) A May 2024 paper in *The Journal of Chemical Physics* pushed that bottleneck by pairing kinetic Monte Carlo with neural-network potentials that evaluate adsorption and activation energies on the fly. The authors, Tomoko Yokaichiya, Tatsushi Ikeda, Koki Muraoka, and Akira Nakayama, tested the scheme on hydrogen on palladium(111) and platinum(111), plus carbon-monoxide oxidation on platinum(111). (pubs.aip.org) Their setup stores configurations it has already seen, so later steps can reuse those energies instead of recalculating them. The paper says that makes it practical to include lateral interactions — the way neighboring adsorbates reshape the energy landscape of a surface reaction. (pubs.aip.org) A separate 2026 paper in *Scientific Reports* tackles a related surface problem from the structure side rather than the kinetics side. Jakub Antoš and Pavel Kocán present PairPotMCinator, a Monte Carlo code that uses pair-potential forces to predict how organic molecules arrange themselves on solid surfaces. (nature.com) The authors frame it as a preprocessor for heavier quantum calculations: generate close-to-relaxed candidate structures first, then hand those candidates to density functional theory. They test the method on ordered pigment layers on highly oriented pyrolytic graphite and on the silicon(111)-indium 7×3 surface. (nature.com) The same “learn or simplify first, solve the expensive physics second” logic is now showing up away from surfaces too. A new April 2026 arXiv preprint by Kenta Yoshimura replaces hand-designed trial densities in nuclear density functional theory with multilayer perceptrons that are optimized directly by energy minimization. (arxiv.org) That paper reports binding energies for calcium-40, zirconium-90, and lead-208 within 0.5% of existing extended Thomas-Fermi calculations, and it reproduces nuclear “pasta” shapes including spheres, rods, and slabs. Yoshimura also writes that single-precision arithmetic performed comparably to double precision, a detail aimed at graphics-processing-unit hardware. (arxiv.org) These projects are not the same method, and they do not solve the same physics. But they converge on the same computational trade: use machine learning or reduced interactions to search huge configuration spaces fast enough that high-fidelity calculations can be reserved for the most promising states. (pubs.aip.org) (nature.com) (arxiv.org) For readers seeing these papers grouped together online, the common thread is not one new platform or one shared codebase. It is a broader shift in computational physics and chemistry toward tools that skip dead time, narrow the search, and make intractable surface and many-body problems small enough to run. (pmc.ncbi.nlm.nih.gov) (arxiv.org)