← Latest papers
🔬 materials science

Capturing long-range interaction with reciprocal space neural network

This paper proposes a reciprocal space neural network method that transforms real-space structural information to enable machine learning interatomic potentials to accurately capture long-range interactions, thereby improving the prediction of global properties and facilitating large-scale dynamic simulations of complex systems like ionic materials and defective crystals.

Original authors: Hongyu Yu, Liangliang Hong, Shiyou Chen, Xingao Gong, Hongjun Xiang

First appeared 2022-11-30 · Updated 2026-11-17
📖 4 min read☕ Coffee break read

Original authors: Hongyu Yu, Liangliang Hong, Shiyou Chen, Xingao Gong, Hongjun Xiang

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

To understand the materials that make up our world, from the silicon in a computer chip to the salt in our shaker, scientists often turn to a powerful tool called molecular dynamics. This is a way of simulating how atoms move and interact over time, but to run these simulations, researchers need a reliable map that tells them how much energy is stored in any given arrangement of atoms. For decades, the gold standard for creating these maps has been a method based on the laws of quantum mechanics, which is incredibly accurate but so computationally expensive that it can only handle tiny groups of atoms for very short periods. To study larger systems or longer timescales, scientists have turned to machine learning, training computers to learn the relationship between atomic positions and energy. However, most of these modern learning models operate on a simple assumption: that an atom only cares about its immediate neighbors. They look at the local neighborhood and ignore what is happening far away. This works well for many materials, but it fails for others, particularly those made of charged particles like salts or certain crystals, where the pull or push of an atom can reach across the entire structure, influencing the material's behavior in ways that local snapshots cannot capture.

A team of researchers at Fudan University and the Shanghai Qi Zhi Institute has developed a new approach to fix this blind spot. They realized that to understand how distant atoms influence one another, they needed to look at the material not just as a collection of individual points in space, but as a whole pattern of waves. In physics, there is a mathematical way to translate the arrangement of atoms in a crystal into a different kind of space, often called reciprocal space, which is naturally suited for describing patterns that repeat over long distances. The researchers built a new type of computer model that takes the standard picture of a crystal, translates it into this wave-based language, and uses a neural network to learn the energy contributions from these long-range patterns. They then combined this global view with the traditional local view, creating a hybrid model that sees both the immediate neighbors and the distant influences simultaneously.

The team tested this new method on several different materials to see if it could truly improve predictions. First, they looked at a dataset of hybrid organic-inorganic perovskites, a type of crystal used in solar cells, to predict their band gaps, which is a measure of how well they conduct electricity. When they compared their new model against existing methods, the hybrid approach that included the long-range wave information was the most accurate, significantly reducing the error in its predictions. In contrast, they found that simply adding a standard long-range calculation without the new neural network component did not help, and using only the long-range information without the local details was not enough on its own. The success came from the specific way they combined the two perspectives.

Next, the researchers tackled a more difficult challenge: simulating crystals with defects, or missing pieces, where long-range forces are especially critical. They created a model of a salt crystal, sodium chloride, and introduced random jitters to the atoms to simulate real-world imperfections. They trained their model to predict the total energy of these structures, a task that requires accounting for the electrical attraction between every single atom in the system. The results showed that their new method, which learns the long-range forces directly from the data, was far more accurate than previous models, especially when the crystal was large and contained defects. They also tested this on a complex semiconductor material called gallium nitride, which is used in high-power electronics. In this case, the new model not only predicted the energy more accurately but also provided a much more stable picture of how the forces between atoms behaved, reducing the errors in force predictions by a significant margin compared to standard models.

The key to this improvement was not just adding more data, but changing how the computer "sees" the crystal. By translating the atomic positions into a format that naturally captures the repeating, long-distance patterns of the material, the model could learn the subtle, far-reaching interactions that traditional models miss. The researchers also discovered that the way they trained the model mattered greatly; they found that teaching the computer to understand the local environment first, and then adding the long-range information in a second step, produced much better results than trying to learn everything at once. This new approach offers a promising path forward for designing better materials, particularly for systems where the behavior of the whole depends on the invisible connections between atoms that are far apart.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →