Researchers at JAIST and ByteDance Seed have paired neural-network quantum Monte Carlo with a new local pseudopotential method, cutting computational cost while keeping errors within accepted bounds.

Much of modern chemistry and materials science runs on simulation. Before anyone mixes chemicals or grows a crystal, researchers increasingly use supercomputers to model how electrons behave in a material — because electrons, ultimately, decide whether a compound will conduct electricity, catalyze a reaction, or hold together under stress. The catch is that the most accurate simulations are also the most expensive, and there is constant demand for methods that are faster without sacrificing accuracy.²

A team from the Japan Advanced Institute of Science and Technology (JAIST), working with researchers at ByteDance Seed in China, reports a step toward that goal. In a study published July 10 in Nature Computational Science, they describe a technique that extends the reach of neural network-based quantum Monte Carlo, one of the most accurate simulation methods available, by combining it with a newly developed approach to handling the electrons that matter least.^1,2

Accurate, but expensive

Quantum Monte Carlo methods tackle the quantum behavior of electrons using statistical sampling. In recent years, researchers have begun using neural networks to represent the extraordinarily complex wave functions that describe many interacting electrons, and these neural network quantum Monte Carlo methods have attracted attention for their high accuracy. Their weakness is cost: the computational demands have been so steep that the methods have largely been confined to small molecular systems.²

Part of the problem is that a full simulation tracks every electron in every atom, including the tightly bound inner-shell electrons that play little role in chemistry. In heavier elements, those core electrons vastly outnumber the outer valence electrons that actually form bonds — yet they consume computing resources all the same.

Replacing the core

Chemists have long dealt with this by using pseudopotentials, mathematical stand-ins that replace an atom’s core electrons with an effective potential so that only the chemically active electrons need to be simulated explicitly. The difficulty has been making pseudopotentials work well inside quantum Monte Carlo, where certain types of pseudopotentials introduce complications.

The JAIST-ByteDance team developed what it calls a Bayesian localization of pseudo-Hamiltonian approach and combined it with neural network techniques. The result, described in the paper as empowering neural network-based quantum Monte Carlo with local pseudopotentials, achieved both accurate predictions and reduced computational cost.^1,2 The team included Associate Professor Tom Ichibha and doctoral student Ryunosuke Fujimaru, one of the study’s co-first authors, from JAIST, along with collaborators at ByteDance Seed.²

Tests reported by the team plotted prediction error against the distance between atoms in materials, comparing the new method with a range of existing approaches. The new method’s results stayed within the accepted range of error while enabling much faster simulations.²

Physics and chemistry, with an AI assist

For Ichibha, the result reflects a broader shift in how computational science is done. “By integrating AI techniques into research fields that have traditionally been advanced through physics and chemistry, significant progress has been achieved,” he said. “This method is expected to contribute to the discovery of novel materials and the understanding of biological phenomena, and the present results will greatly advance research in these areas.”²

The practical promise is scale. By trimming the cost of each simulation, the approach opens the door to high-precision analysis of large materials systems and complex chemical reactions that were previously out of reach because of limited computing resources.² That matters for problems in which small energy differences decide outcomes — which catalyst works, which battery material is stable, how a protein’s active site behaves.

What comes next

The researchers plan to extend the method to a wider range of elements, which would broaden its usefulness in discovering new materials, designing high-performance catalysts, and understanding how biomolecules function. They also anticipate applications in solid-state physics and in calculations of excited states, the energized configurations of electrons that govern how materials absorb and emit light.²

Those are areas where existing methods often struggle and where better accuracy at manageable cost could help resolve long-standing open questions in quantum and materials science.² The collaboration itself — a Japanese national graduate university working alongside the research arm of a major technology company — is also a sign of how much industrial AI expertise is now flowing into fundamental science.

Neural network approaches to quantum chemistry have generated real excitement over the past several years, but excitement has often run ahead of practicality. Work like this, which focuses less on headline accuracy and more on bringing those methods within reach of larger, messier, more realistic systems, may be what ultimately turns them into everyday tools for chemists and materials scientists.

---

Notes

1. Fu, W., Fujimaru, R., Li, R., Liu, Y., Wen, X., Li, X., Hongo, K., Wang, L., Ichibha, T., Maezono, R., Chen, J., & Ren, W. (2026). Empowering neural network-based quantum Monte Carlo with local pseudopotentials. Nature Computational Science. https://doi.org/10.1038/s43588-026-01008-7
2. Japan Advanced Institute of Science and Technology. (2026, July 22). A breakthrough in neural network quantum computation [Press release]. EurekAlert! https://www.eurekalert.org/news-releases/1136876