Scalable long-range machine learning interatomic potentials with electrical response for atomistic material modeling
This paper introduces DPA4C-LR, a scalable machine learning interatomic potential framework that analytically separates long-range electrostatics from local latent charge derivatives to enable efficient, distributed atomistic simulations of electrical responses and field-dependent material properties across millions of atoms.
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
Materials science has long faced a fundamental divide. On one side sit the laws of quantum mechanics, which can predict how individual atoms behave with perfect precision but are too slow to simulate more than a few thousand atoms at a time. On the other side are the massive simulations needed to understand real-world devices, like batteries or computer chips, which contain millions or billions of atoms but often rely on simplified rules that miss the subtle details of how atoms actually interact. The missing link has been a way to simulate huge collections of atoms with the accuracy of quantum physics, especially when those atoms are subjected to electric fields. In the real world, electricity drives ions to move, switches materials between different states, and generates heat at the microscopic level. To design better devices, scientists need to see these processes unfold across vast distances, but the math required to track the invisible electric forces between every single atom usually makes such large simulations impossible.
A team of researchers at the National University of Singapore has bridged this gap by creating a new computational framework that allows them to simulate millions of atoms with high accuracy while accounting for long-range electrical effects. Their work, published recently, introduces a method called DPA4C-LR. This system combines a highly efficient, local model that predicts how atoms push and pull on their immediate neighbors with a separate, global calculation that handles the long-distance electric forces. By splitting the problem this way, the researchers avoided the computational bottleneck that usually stops scientists from simulating large systems under electric fields. They demonstrated that their approach can run on just four powerful graphics processors to model a system containing 5.2 million atoms, a scale previously unreachable for this level of detail.
The core challenge the team solved was how to include electrical response without slowing everything down. In many materials, the way atoms respond to an electric field depends on the collective behavior of the entire system, not just the atoms nearby. Traditional methods that try to calculate these global effects for every step of a simulation require all the data to be held in one place, which prevents the work from being split up across multiple computers. The new method bypasses this by treating the long-range electric forces as a separate step that can be solved efficiently using established mathematical techniques, while the complex, local interactions are handled by a machine learning model that can be distributed across many processors simultaneously. This allows the simulation to scale up to millions of atoms without losing the ability to predict how the material will react to an electric current or field.
To prove their system works, the researchers tested it on three very different materials, each requiring a large number of atoms to reveal the truth. First, they looked at liquid water. Water molecules are polar, meaning they have a positive end and a negative end, and they constantly reorient themselves in response to electric fields. The team simulated a massive block of water and applied electric fields of varying strengths. Their model successfully predicted how the water's infrared spectrum—the unique fingerprint of how it absorbs light—shifted as the field changed. This confirmed that the system could accurately infer how the atoms would move and respond to electricity, even though it was only trained on basic energy and force data, without ever being explicitly taught about electric fields.
Next, the team tackled a complex ceramic material known as a relaxor ferroelectric, specifically lead magnesium niobate. These materials are famous for their ability to switch polarization, a property used in sensors and actuators, but their behavior is notoriously difficult to simulate because it depends on tiny, disordered regions of atoms that only appear in very large samples. When the researchers simulated small chunks of this material, the results looked like a standard, orderly crystal. However, when they scaled up to a simulation containing nearly 30,000 atoms, the picture changed completely. The larger simulation revealed a "slim" hysteresis loop, a specific signature of the material's ability to switch states without getting stuck, which is the defining characteristic of a relaxor. This result showed that the finite size of previous simulations had been hiding the true nature of the material, and that only by reaching the scale of millions of atoms could the correct behavior be observed.
Finally, the researchers applied their method to a solid-state electrolyte, a glass-like material used in next-generation batteries, specifically lithium niobium oxychloride. In these materials, ions must move freely through a disordered structure to conduct electricity. The team simulated a large glass sample and applied a steady electric field to watch the lithium ions drift. By tracking the movement of these ions over time, they calculated the material's electrical conductivity. The result was a precise value of 6.8 ± 0.7 millisiemens per centimeter at room temperature. This measurement was far more reliable than previous estimates derived from simulations without an applied field, which rely on tiny, random thermal fluctuations that are hard to measure accurately. The new method provided a clear, direct view of how the material conducts electricity, offering a reliable benchmark for designing better battery components.
The implications of this work extend beyond these specific examples. By proving that long-range electrical effects can be calculated efficiently across millions of atoms, the researchers have opened the door to simulating entire device layers rather than just tiny fragments of them. This means scientists can now model how heat and electricity flow across interfaces in real devices, such as the boundary between a silicon chip and a diamond heat sink, with a level of detail that was previously impossible. The framework is not just a theoretical advance; it is a practical tool that runs on existing hardware and can be used to explore the behavior of materials under the exact conditions they will face in the real world. As the demand for faster, more efficient electronic and energy devices grows, the ability to simulate these systems at the scale of millions of atoms provides a crucial foundation for innovation, allowing researchers to predict and optimize material performance before a single physical prototype is built.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.