← Latest papers
🔬 materials science

Fine-tuned atomistic foundation model for accurate sodium-carbon chemistry in Na-ion batteries

This paper presents a fine-tuned MACE-based atomistic foundation model that overcomes the limitations of pretrained models in capturing complex sodium-carbon interactions, enabling accurate DFT-level predictions of hard carbon storage mechanisms and establishing a direct link between closed-pore size and plateau voltage for Na-ion battery design.

Original author: Andres Felipe Usuga (Centre of Excellence ENSEMBLE3 Sp. z o. o), Ihor Radchenko (Centre of Excellence ENSEMBLE3 Sp. z o. o), Yuxin Tang (Qingyuan Innovation Laboratory, College of Chemical Engineering
Published 2026-11-16
📖 6 min read🧠 Deep dive

Original author: Andres Felipe Usuga (Centre of Excellence ENSEMBLE3 Sp. z o. o), Ihor Radchenko (Centre of Excellence ENSEMBLE3 Sp. z o. o), Yuxin Tang (Qingyuan Innovation Laboratory, College of Chemical Engineering - Fuzhou University), Oleksandr I. Malyi (Qingyuan Innovation Laboratory, Centre of Excellence ENSEMBLE3 Sp. z o. o)

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 author. For technical accuracy, refer to the original paper. Read full disclaimer

Batteries are the silent engines of modern life, powering everything from smartphones to electric cars. At the heart of many next-generation batteries is a simple chemical dance: sodium ions, which are abundant and cheap, moving in and out of a carbon-based material to store and release energy. For this process to work well, scientists need to understand exactly how these sodium atoms behave when they encounter carbon. The challenge lies in the fact that carbon is not a uniform block; in the materials used for batteries, it is a complex landscape of flat sheets, curved surfaces, and tiny, sealed pockets. Sodium atoms interact with this landscape in different ways. Sometimes they stick to the surface like dust on a table, and other times they clump together to form tiny metallic droplets inside the sealed pockets. Predicting exactly how these atoms arrange themselves and how much energy is required to move them is crucial for designing better batteries, but it is a task that has long been too difficult for standard computer simulations to handle efficiently.

For years, researchers have relied on a powerful method called density functional theory to simulate these atomic interactions. This approach is highly accurate, acting as a gold standard for predicting how atoms behave, but it is also incredibly slow and expensive to run. To simulate a large, complex battery material, it would take a supercomputer an impractical amount of time. To speed things up, scientists have turned to artificial intelligence, specifically a type of model known as a foundation model. These are large, pre-trained computer programs that have learned the general rules of chemistry by studying millions of different materials. The hope was that these models could act as a fast, accurate shortcut, predicting how sodium and carbon interact without needing the heavy computational cost of the traditional method. However, a new study reveals that while these AI models are powerful, they are not yet perfect for this specific job.

The researchers, working across institutions in China and Poland, set out to test whether these existing AI models could accurately describe the tricky behavior of sodium in hard carbon, a leading candidate for battery anodes. They focused on a specific, critical problem: the competition between two different ways sodium likes to bond. On one hand, sodium can give up an electron to the carbon, creating a weak, ionic attraction. On the other hand, sodium atoms are naturally drawn to each other, preferring to clump together into metallic clusters. In a real battery, both of these forces are at play, and the balance between them determines how the battery charges and discharges. When the team tested the best available pre-trained AI models, they found a significant flaw. The models were good at predicting how sodium sits on flat carbon sheets, but they failed completely when it came to the metallic clumping. The AI predicted that sodium atoms would repel each other and stay spread out, whereas the more accurate, traditional calculations showed that they should attract and form stable clusters. This error meant the AI models could not correctly predict the energy stability of the battery material, which would lead to wrong predictions about the battery's voltage and performance.

To fix this, the team did not start from scratch. Instead, they took one of the existing foundation models and gave it a targeted lesson, a process known as fine-tuning. They created a specialized database of examples that specifically highlighted the missing physics: scenarios where sodium atoms were close together, forming clusters, and scenarios where the carbon structure was slightly stretched or defective. By feeding these specific examples back into the model, they taught it to recognize the subtle difference between sodium sticking to carbon and sodium sticking to itself. The result was a dramatic improvement. The new, fine-tuned model could now predict the energy of these complex arrangements with an accuracy that matched the slow, gold-standard calculations, but at a fraction of the time. It successfully learned to distinguish between the dispersed sodium atoms and the metallic clusters, capturing the correct energetic balance that the original models had missed.

With this reliable tool in hand, the researchers turned their attention to a practical question that has puzzled battery designers for years: how do the tiny, sealed pores inside hard carbon affect the battery's voltage? In a hard carbon electrode, these pores are like microscopic caves where sodium can hide. It was generally assumed that all these pores filled up at the same low voltage, creating a flat line on the battery's charging graph. However, using their new fine-tuned model, the team simulated sodium filling pores of different sizes. They discovered that this assumption was incorrect. The size of the pore matters immensely. Smaller pores, which are tighter and more confined, fill up at a higher voltage, while larger pores fill up at a lower voltage. This means that in a real battery, which contains a mix of pore sizes, the charging process is not a single flat line but a complex curve where different pores contribute at different times. The smallest pores provide the strongest drive to fill up first, while the largest ones require a lower energy state to accept sodium.

This finding changes how scientists might think about designing battery materials. It suggests that by controlling the size distribution of these tiny pores, engineers could potentially tune the voltage profile of a battery more precisely than previously thought. The study demonstrates that while artificial intelligence foundation models are a powerful starting point, they often need to be taught the specific, competing rules of the chemical environment they are meant to simulate. By correcting the model's understanding of how sodium atoms interact with each other versus how they interact with carbon, the researchers created a tool that is both fast and accurate. This approach bridges the gap between the slow, perfect simulations of the past and the need for rapid, reliable design in the future, offering a clearer path toward better sodium-ion batteries. The work confirms that to truly understand complex materials, an AI must not just memorize data, but learn the underlying physical rules that govern how atoms compete and cooperate.

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 →