🔬 materials science

Predicting large-supercell defect formation energies from machine-learning charge density models trained on small supercells

This paper proposes a machine-learning charge density (MLCD) approach that, by optimizing training sets with mixed-size small supercells, achieves high-accuracy prediction of large-supercell defect formation energies with significantly greater data efficiency and lower error rates compared to traditional machine-learning interatomic potentials.

Original authors: Junjie Zhou, Menglin Huang, Shiyou Chen

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Junjie Zhou, Menglin Huang, Shiyou Chen

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

Imagine you are trying to understand how a single broken tile affects the entire floor of a massive, ancient castle. In the world of materials science, these "broken tiles" are called defects—missing atoms or atoms in the wrong place inside a crystal. These tiny mistakes can change how a material conducts electricity, glows, or holds up under pressure. To figure out exactly what these defects do, scientists use a powerful computer simulation tool called Density-Functional Theory (DFT). Think of DFT as a super-accurate digital microscope that calculates the behavior of every electron in the material. However, there's a catch: to get a true picture of a single defect, you need to simulate a huge chunk of the crystal so that the "broken tile" doesn't accidentally bump into its own reflection from the other side of the screen. This requirement for huge, expensive simulations makes studying these defects incredibly slow and costly, like trying to map a whole city just to find one pothole.

Recently, scientists have tried to speed things up using machine learning, teaching computers to guess the answers instead of calculating them from scratch. Usually, these computer models are trained to predict the total energy of a system, kind of like guessing the weight of a suitcase by looking at its size. But this paper suggests a different approach: instead of guessing the weight, let's teach the computer to see the "charge density." If you imagine the electrons in a material as a fog that swirls around the atoms, charge density is the map of how thick or thin that fog is in every single spot. The authors propose that if a computer learns to predict this fog map accurately, it can figure out the energy of a defect much more efficiently, even if it only sees small pieces of the puzzle during its training.

The researchers, Junjie Zhou, Menglin Huang, and Shiyou Chen from Fudan University, tested this idea on Gallium Nitride (GaN), a material used in everything from blue LEDs to power electronics. They focused on four specific types of "broken tiles" (defects) in this material. Their main goal was to see if they could train a machine-learning model on small, cheap simulations and then use it to predict the behavior of defects in massive, expensive simulations.

Here is the clever twist they discovered: they didn't just train the model on one size of simulation. Instead, they used a "mixed-size" strategy. They fed the computer a dataset containing a mix of small supercells (with 16 to 32 atoms) and a few medium-to-large ones (up to 96 atoms). They found that the small cells were great at teaching the model what the "bulk" material looks like far away from the defect, while the slightly larger cells were necessary to show the model how the defect actually distorts the nearby atoms. By mixing these different sizes, they created a model that could accurately predict the energy of a defect in a giant 360-atom supercell.

The results were striking. Using a dataset of only 96 different structures, their new model predicted the formation energy of defects with an error of less than 0.05 electron volts (eV). To put that in perspective, a standard machine-learning model trained on the same small data but trying to guess the total energy directly (instead of the charge density) made mistakes of over 1 eV—more than 20 times worse. The authors showed that this charge-density approach allows the computer to "extrapolate" or stretch its knowledge from small cells to large ones much better than previous methods.

They also demonstrated that this method works for more than just energy. Because the model predicts the charge density map, they could use it to calculate other electronic properties, like the energy levels where electrons get trapped (defect levels) and how the material spins (spin polarization). For instance, they correctly predicted the spin-splitting in a specific defect, a detail that smaller simulations completely missed because the "broken tile" was too close to its own reflection in those tiny boxes.

The paper argues that this success happens because charge density is a local property. If the computer learns what the electron fog looks like around a missing atom in a small box, it can recognize that same pattern in a huge box, provided it has seen enough examples of how the fog behaves at different distances. This is unlike total energy, which is a global sum of everything in the box and gets messed up if the box is too small.

In conclusion, the authors suggest that by training on a smart mix of small and medium-sized supercells and focusing on the charge density map rather than just the total energy, scientists can drastically reduce the cost of simulating defects. They found that this mixed-size approach reduced the computational cost by a factor of 6.2 compared to using only large supercells, while maintaining high accuracy. This doesn't mean the problem is solved for every material in the universe, but it offers a promising, data-efficient pathway for predicting how defects behave in large supercells, potentially speeding up the discovery of better materials for our electronics.

Technical Summary: Predicting Large-Supercell Defect Formation Energies from Machine-Learning Charge Density Models

Problem Statement
First-principles defect calculations using Density-Functional Theory (DFT) are computationally expensive because accurate results require large supercells to suppress artificial image interactions between periodic defects. The computational cost scales approximately as $O(N^3)$ with system size, limiting the application of DFT to large systems. While Machine-Learning Interatomic Potentials (MLIPs) offer an alternative, training them for defect systems typically requires thousands of large-supercell structures, a process that can take weeks of data generation. Furthermore, standard MLIPs trained on small supercells often fail to extrapolate accurately to larger cells due to systematic errors arising from strong defect-image interactions in the training data.

Methodology
The authors propose a Machine-Learning Charge Density (MLCD) approach as a more data-efficient pathway for predicting defect formation energies in large supercells. The core workflow involves:

Training: An MLCD model (based on the EAC-Net framework using equivariant atom-probe representations) is trained to predict charge densities ( $\rho$ ) directly from atomic structures ( $R$ ).
Inference: For a target large supercell, the model predicts the charge density $\rho'$ .
Property Extraction: Electronic structure properties, including total energy, forces, and eigenvalues, are obtained via a single non-self-consistent-field (NSCF) DFT calculation using the predicted $\rho'$ as a fixed input. This bypasses the iterative self-consistent loop, reducing the effective cost to nearly $O(N)$ .

To address the challenge of cross-size transfer, the authors introduce a mixed-size training strategy. They analyze the spatial extent of defect-induced charge perturbations and define a criterion: the half-lattice length of the training supercell must be comparable to or larger than the defect influence radius to avoid distorted local environments caused by periodic image interactions. Based on this, they construct training sets that mix small supercells (16–32 atoms, providing configurational diversity) with intermediate-sized supercells (up to 96 atoms, acting as "anchors" to capture the correct local defect structure).

Key Contributions and Results
The study focuses on four intrinsic defects in Gallium Nitride (GaN): nitrogen vacancy ( $V_N$ ), gallium vacancy ( $V_{Ga}$ ), gallium on nitrogen site ( $Ga_N$ ), and nitrogen on gallium site ( $N_{Ga}$ ).

Superiority of MLCD over MLIPs: When trained on the same dataset, MLCD models significantly outperform MLIPs (specifically Allegro) in cross-size transfer. While MLIPs trained on small supercells exhibited errors exceeding 1 eV for 360-atom targets, the MLCD approach achieved much higher accuracy.
Limitations of Single-Size Training:
- Training solely on 96-atom supercells achieved the target accuracy (Mean Absolute Error, MAE < 0.05 eV) but required 240 structures.
- Training solely on 16–32-atom supercells resulted in a systematic error floor around 0.1 eV, regardless of dataset size (up to 720 structures), due to the inability of small cells to represent the non-local charge response of certain defects (e.g., $Ga_N$ ).
Success of Mixed-Size Training: By combining 16–96 atom supercells, the authors developed a model ( $\gamma(16-96)$ ) that reached the target accuracy of < 0.05 eV using only 96 structures. This represents a reduction in the number of required configurations by roughly 60% compared to the single 96-atom strategy.
Electronic Structure Recovery: The MLCD-predicted charge densities successfully recovered complex electronic properties in 360-atom supercells, including band structures, defect levels, and spin polarization. Notably, the mixed-size training was essential for capturing spin-splitting in $Ga_N$ defects, which was suppressed in the small training cells due to image interactions.

Significance and Claims
The paper claims that learning charge density directly, rather than global energy and forces, enables more robust cross-size transfer because charge density is a spatially resolved, local quantity. Local patterns learned in small cells can be mapped to identical environments in larger cells, provided the training set samples the relevant defect and bulk regions.

The proposed mixed-size training strategy is presented as a novel data augmentation paradigm that treats the prediction of large-cell properties as a "continuation problem." Intermediate-sized supercells serve as anchors that constrain the extrapolation path from small to large cells, reducing uncertainty. The authors conclude that this approach offers a highly cost-effective route for accurate first-principles defect prediction, reducing the computational cost of dataset preparation by a factor of 6.2 compared to training exclusively on large supercells, without sacrificing accuracy.

More like this