When do machine-learning potentials need a configuration-dependent kernel? Learning long-range interactions with Fourier neural operators
This paper demonstrates that while configuration-independent kernels suffice for homogeneous systems, Fourier neural operators (FNOs) are necessary to accurately capture long-range interactions in machine-learning potentials when the screening medium is inhomogeneous, as they uniquely learn the configuration-dependent coupling required to model remote modulation effects.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the microscopic world of atoms, forces do not always stop at a neighbor's doorstep. While some interactions are short and sweet, others stretch across vast distances, reaching through entire materials to influence how atoms behave far away. This is particularly true in liquids like water or in solid crystals made of charged ions, where the electric pull of one particle can ripple through the medium to affect another particle on the other side of the container. For decades, scientists have relied on computer models to simulate these materials, but standard models often struggle with these long-range connections. They are like people who only listen to those standing right next to them, missing the whispers coming from across the room. To fix this, researchers have tried adding a "long-range" module to their models, usually by assuming the electric forces follow a simple, unchanging rule, much like gravity always pulling down with the same strength regardless of the weather. But in the real world, the environment changes; materials can screen or block these forces differently depending on what is around them, making the simple rule insufficient.
A team of researchers at UCLouvain in Belgium has developed a new way to teach computers how to handle these complex, shifting long-range forces. Instead of hard-coding a single rule for how atoms talk to each other over a distance, they built a system that learns the rules itself, adapting to the specific environment it finds itself in. They tested this system, which they call a Fourier neural operator, on a variety of challenges, from synthetic salt-like fluids to real-world materials like liquid water and a gold cluster sitting on a magnesium oxide surface. Their work reveals that when the environment is uniform, a simple, fixed rule works well enough. However, when the environment is messy or uneven—such as when different parts of a material block electric forces to different degrees—only a flexible, learning system can get the physics right.
The researchers began by creating a synthetic world of charged particles to test their ideas. In this controlled setting, they could tune how strongly the particles screened each other, simulating everything from a nearly empty space where forces travel freely to a thick, crowded medium where forces die out quickly. They compared their new learning system against older methods that used a fixed, pre-written rule for these long-range interactions. In the uniform cases, where the screening was the same everywhere, the new system performed just as well as the old fixed-rule methods, accurately predicting the energy and forces between atoms. This confirmed that the new approach could learn the standard physics without needing to be told the answer in advance.
The real test came when the researchers introduced inhomogeneous screening, a situation where the ability of the material to block or transmit forces changes from one spot to another. Imagine a crowd of people where some areas are packed tight, blocking sound, while other areas are empty, letting sound travel far. In this uneven environment, the way two distant atoms interact depends not just on the atoms themselves, but on the specific arrangement of the material between them. The researchers found that the older methods, which relied on a fixed rule, failed to capture this nuance. They could not adjust their "listening" based on the changing crowd between the atoms. In contrast, the new learning system, which uses a nonlinear approach to adapt its rules, successfully tracked these changes. It reduced the error in predicting atomic forces by nearly half compared to the fixed-rule models. The system learned that the connection between two distant atoms is not a static line but a dynamic relationship that shifts as the medium around them rearranges.
To prove this was not just a result of the synthetic data, the team applied their method to three real-world materials: liquid water, a solid electrolyte used in batteries called lithium lanthanum zirconate, and a tiny cluster of two gold atoms sitting on a crystal surface. In the case of liquid water and the battery material, the new system matched the accuracy of the best existing methods, showing it is a reliable tool for standard bulk materials. However, the gold cluster experiment provided the most striking evidence of its power. In this setup, the gold atoms sat on a surface that had aluminum impurities buried deep underneath, more than ten angstroms away. These distant impurities changed the electronic environment enough to flip the gold cluster's preference: it wanted to stand up straight in a clean environment but lay flat when the aluminum was present. Standard models that only look at immediate neighbors missed this flip entirely. The fixed-rule long-range models got it partially right but still made significant errors. The new learning system, however, captured this subtle reversal with high precision, predicting the change in preference with an error of less than five thousandths of an electron volt per cell.
The study concludes that the need for a flexible, learning-based approach arises specifically when the material's ability to screen forces varies across space. If the environment is uniform, a simple, fixed rule is sufficient. But once the screening medium becomes uneven, the connection between distant atoms becomes dependent on the specific configuration of the material between them. The researchers showed that their new system can learn this complex, configuration-dependent behavior using only the energy and force data from simulations, without needing to be explicitly programmed with the physics of screening. This offers a powerful new route for simulating materials where charge transfer, interfaces, or defects create complex, long-range interactions that traditional models struggle to describe. The work suggests that for the most challenging materials science problems, the future lies not in writing better rules, but in teaching computers how to discover the rules themselves.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.