Researchers have long used computer simulations that follow the behavior of electrons to understand materials and biological processes. One promising approach, known as neural network quantum Monte Carlo, combines artificial neural networks with a statistical sampling method that approximates solutions to quantum problems. This combination delivers high accuracy. However, the calculations demand so much computer power that they have been practical only for very small molecules.
A new computational technique overcomes this barrier. Scientists at the Japan Advanced Institute of Science and Technology, working with colleagues from ByteDance Seed in China and other institutions, joined neural network methods with an approach they call Bayesian localization of a pseudo Hamiltonian. In simple terms, a Hamiltonian is the mathematical description of a system’s total energy in quantum mechanics; the “pseudo” version is a simplified stand-in, and Bayesian localization is a probability-based way of concentrating effort on the most important parts of that description. The result keeps the accuracy of the older method while sharply lowering the computing cost.
How the advance expands the reach of simulations
The new technique makes it possible to study much larger materials and more complex chemical reaction systems that were previously out of reach because of limited computing resources. Future work that extends the method to a wider range of chemical elements should support the search for new materials, the design of better catalysts, and clearer understanding of how biomolecules function. The same tools are also expected to help with problems in solid-state physics and with calculations that involve excited electronic states, areas where many questions remain open.
By bringing artificial intelligence into fields that have traditionally relied on physics and chemistry, the approach advances next-generation simulation technology. It offers a practical route to high-precision analysis of systems that matter for both materials discovery and biological insight.
The researchers have described the methods and results iof this study in a paper published in Nature Computational Science.