PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential
This paper introduces the Perturbation Augmented Space group structure Sampling (PASS) method to generate a representative dataset for training a transferable, accurate, and computationally efficient machine learning interatomic potential based on the atomic cluster expansion framework, which successfully models the complex reactive dynamics of iron oxidation across bulk, surface, and interface systems.
Original paper licensed under CC BY 4.0 (https://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
Iron rusting is a process we see every day, a slow transformation where metal meets air and gradually turns into a brittle, reddish crust. While this phenomenon seems simple on the surface, the atomic world beneath it is a chaotic and complex landscape. When iron atoms react with oxygen, they do not just sit still; they shuffle, bond, and rearrange into different crystal structures, some of which are magnetic and others that are not. Scientists have long wanted to simulate this process on a computer to understand exactly how the rust forms, how it spreads, and how to stop it. To do this, they need a set of rules, known as an interatomic potential, that tells a computer how every single iron and oxygen atom should behave when they bump into one another. The challenge has always been that iron and oxygen can form many different types of structures, and creating a single set of rules that works for all of them without breaking down has been nearly impossible.
For years, researchers have tried to build these rules by hand, guessing which atomic arrangements might happen during rusting, or by letting computers search for them one by one, a slow and expensive process. A new study by Zixiong Wei, Fei Shuang, and Poulumi Dey at Delft University of Technology offers a different path. They developed a method called PASS, which stands for Perturbation Augmented Space Group structure Sampling. Instead of guessing or searching endlessly, this method starts with a library of all possible geometric patterns that iron and oxygen atoms can form, known as space groups. The researchers then take these patterns and systematically shake and stretch them, creating thousands of slightly distorted versions. This process generates a vast collection of atomic snapshots that cover both stable structures and the messy, unstable states that occur when atoms are moving or reacting. By feeding this diverse collection into a machine learning system, the team trained a new computer model that can predict how iron and oxygen interact with high accuracy.
The researchers tested this new model on a wide range of scenarios to see if it could handle the complexity of real-world rusting. They checked if it could correctly predict the properties of pure iron, including how it expands when heated and how it breaks under stress. They also tested it on various iron oxides, the different chemical compounds that make up rust, such as the red iron oxide found in hematite and the black magnetic oxide found in magnetite. Even though the model was trained only on tiny clusters of atoms containing fewer than ten atoms, it proved capable of describing much larger and more complicated structures, including the surfaces where rust begins to form and the interfaces where metal meets oxide. The model successfully predicted how oxygen atoms dissolve into iron and how they move through the metal, matching results from expensive, high-precision calculations and real-world experiments.
To prove that their method worked for the most difficult part of the problem, the team used their new model to simulate a large-scale oxidation event. They created a virtual block of iron and exposed its surface to oxygen at a high temperature of 873 Kelvin. The simulation ran for one nanosecond, a tiny fraction of a second, but long enough for the atoms to rearrange themselves. The result was a clear formation of an oxide layer on the surface. The model showed that oxygen atoms diffused inward, creating a distinct region where the iron and oxygen were mixed in nearly equal amounts. This new layer had a structure very similar to wüstite, a specific type of iron oxide that forms at high temperatures. The simulation captured the entire process, from the initial adsorption of oxygen on the surface to the growth of the oxide layer and the diffusion of atoms into the bulk metal, all without the model ever having been explicitly taught what rust looks like.
This work demonstrates that it is possible to create a reliable, universal set of rules for iron and oxygen by training on small, systematically generated samples rather than trying to manually design every possible scenario. The new model is not perfect; it does not explicitly account for the magnetic properties of the atoms, which can influence how they behave in certain conditions. However, it provides a robust and practical tool for studying the early stages of corrosion and the growth of oxide layers. By showing that a model trained on tiny, perturbed structures can accurately predict the behavior of large, evolving systems, the researchers have opened a door for more efficient and accurate simulations of chemical reactions. This approach suggests that the key to understanding complex material changes lies not in building bigger models from the start, but in teaching the computer to recognize the fundamental patterns of atomic arrangement through a comprehensive and systematic sampling of possibilities.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.