← Latest papers
🔬 materials science

A Universal Diffusion Model for the Quantum Position Operator

This paper introduces a universal, E(3)-equivariant graph neural network-based diffusion model that accurately predicts quantum position operator moments across diverse materials, enabling scalable calculations of electronic polarization, localization, and key response functions like infrared spectra and piezoelectric tensors for complex systems beyond the reach of routine first-principles methods.

Original author: Ransell D'Souza, Austin Zadoks, Riccardo Dal Molin, Nataliia Manko, Junfeng Qiao, Nicola Marzari, Antimo Marrazzo

Published 2026-11-19
📖 5 min read🧠 Deep dive

Original author: Ransell D'Souza, Austin Zadoks, Riccardo Dal Molin, Nataliia Manko, Junfeng Qiao, Nicola Marzari, Antimo Marrazzo

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

Inside every insulating material, from the glass of a window to the ceramic in a capacitor, electrons are not free to roam. They are bound to atoms or shared in bonds, creating a hidden landscape of electric charge that determines how the material reacts to the world. When an electric field pushes on these materials, or when they are squeezed and stretched, the electrons shift slightly, creating a response that powers everything from sensors to energy harvesters. To understand and predict this behavior, scientists need to track the precise average positions of these electrons, known as Wannier charge centers. These points act like the effective centers of gravity for electron clouds, and their movement dictates the material's electrical personality. However, calculating these positions for complex, messy materials like liquid water or amorphous silicon has long been a bottleneck. The standard methods require immense computing power, making it nearly impossible to simulate the vast, disordered systems needed to understand real-world phenomena like how water absorbs infrared light or how glass conducts electricity under stress.

A team of researchers has now introduced a new approach that bypasses these computational hurdles by using a type of artificial intelligence designed to learn the shape of electronic charge. Instead of trying to solve the complex equations of quantum mechanics for every single atom in a massive simulation, they trained a model to predict where these electron centers should be based on the arrangement of the atoms around them. The researchers call their method a universal diffusion model. In this process, the computer starts with a guess about where the electron centers are located and then iteratively refines that guess, removing random noise step by step until the predicted positions settle into the correct, stable configuration. It is similar to how a sculptor might start with a rough block of stone and chip away the excess to reveal the form within, except here the "chipping away" is a mathematical process that guides the model from chaos to order.

The team tested this framework on three very different types of matter: liquid water, amorphous silicon, and a crystalline ceramic called barium titanate. For liquid water, they trained the model on a small set of thirty-two molecules and then used it to predict the behavior of a massive system containing five hundred and twelve molecules. The model successfully reconstructed the movement of the electron centers, allowing the researchers to calculate the material's infrared spectrum. This spectrum, which shows how the material absorbs light, matched experimental data with high precision, capturing the specific vibrations of the hydrogen bonds that hold water together. This achievement is significant because simulating such a large system with traditional methods would have been prohibitively expensive and time-consuming.

In the case of amorphous silicon, a non-crystalline form of the element used in solar panels, the researchers applied the model to a system of five hundred and twelve atoms. Again, the model accurately predicted the positions of the electron centers, enabling the calculation of the material's infrared response. The results reproduced the characteristic vibrational features of disordered silicon, confirming that the model could handle the lack of a regular crystal structure. The researchers also applied the technique to barium titanate, a material known for its ability to generate electricity when squeezed. By predicting the electron centers, the model allowed them to calculate the material's piezoelectric properties—the electrical response to mechanical deformation—with an accuracy that matched established first-principles calculations.

Perhaps the most striking demonstration of the model's power was its ability to generalize to materials it had never seen before. The researchers trained a single, universal version of the model on a dataset of nearly three thousand crystalline insulators spanning sixty-seven different chemical elements. They then tested this model on one hundred completely new materials that were not part of the training set. Despite never having encountered these specific compounds, the model predicted the electron centers with high accuracy, showing that it had learned the fundamental rules of how electrons localize in different chemical environments. This suggests the model has captured a deep, transferable understanding of electronic structure that goes beyond memorizing specific examples.

The implications of this work extend to the future of materials science. By providing a fast and accurate way to determine electron positions, the model opens the door to simulating complex, large-scale systems that were previously out of reach. This includes studying how electrons behave in disordered liquids, how they move in amorphous solids, and how they respond to temperature changes in functional materials. The researchers envision that this tool could be integrated into the workflow of designing new materials, allowing scientists to screen thousands of candidates for specific electrical properties without the heavy computational cost of traditional methods. While the current work focuses on predicting the position and spread of electron centers, the framework is flexible enough to potentially predict other electronic properties in the future, offering a scalable route to understanding the electronic heart of complex matter.

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 →