Machine learning helps track semiconductor defects under more realistic temperature conditions

Machine learning helps track semiconductor defects under more realistic temperature conditions Sadie Harley Scientific Editor Robert Egan Senior Editor All materials contain inherent defects. These are not always harmful, since they can be used to engineer and introduce certain properties into a material. However, it is crucial to understand how they function and how they affect the material's properties. To understand and model defects in different materials, researchers' standard calculations are usually based on zero Kelvin, minus 273 degrees Celsius, the theoretically lowest possible temperature at which all atoms and molecules are completely stationary. But for materials intended for use in real-world products and everyday applications, it is essential to know how they behave at ordinary temperatures. Researchers at Chalmers University of Technology have now demonstrated a significantly more effective framework for calculating the properties of defects while accounting for different temperature effects, enabling results that are much closer to reality. The findings are presented in an article published in npj Computational Materials. Defects affect the performance and function of materials Julia Wiktor, a professor at the Department of Physics and Astronomy, leads the research group behind the study. Their work focuses on studying and modeling new materials using advanced computer simulations and machine learning. "We usually model materials as perfect crystals where every atom occupies a fixed position in a lattice, but real materials are not so perfect. An atom may be missing, there may be an extra atom between lattice sites, or two atoms may have swapped places. We need to understand these imperfections because they affect the performance and function of materials," says Wiktor. Tracking defects across different temperatures The materials the research group typically investigates are mainly within solar cell technology, but this time they used a broader set of benchmark systems to test whether the method could be applied to other material classes and emerging technologies. "We modeled defects in the materials by studying a missing atom in the crystal structure and observing how temperature affects the behavior of this point defect," says doctoral student Tobias Hainer. "This is something that most models and simulations based on zero Kelvin fail to capture, and it is an important step toward understanding how different materials actually behave under real operating conditions. Our results match experimental observations, but unlike experiments we do not just obtain an initial and a final measurement—we can follow the defect throughout the entire temperature range." The position determines the impact of the defect Defects can often introduce new energy levels into materials. A semiconductor, for example, has a so-called band gap, meaning an energy interval without available electronic states, and defects can introduce different levels within this gap. If a level lies deep within the band gap, this is in most cases undesirable for applications. But if it lies closer to the edge of the band gap, it can be engineered to tune the semiconductor's electrical properties. "The position of a defect level is highly significant when designing applications based on a material. If at zero Kelvin we predict that the defect lies near the edge, we may believe it to be harmless or even beneficial, but if at another temperature it shifts deeper into the gap it may become harmful, or vice versa. The position is therefore crucial," explains Hainer. Temperature matters in modeling The research group can now demonstrate a more efficient way of carrying out calculations at temperatures above 0 Kelvin, in a manner that is practically feasible for many different types of materials and without excessive computational cost. "The most important aspect we investigated is how these energy levels associated with a missing atom change with temperature, and we found quite significant shifts in several materials, both quantitatively and qualitatively. These changes would be missed in a model based on zero Kelvin. We have therefore shown that temperature really does matter in modeling and is necessary in order to describe these systems correctly," says Wiktor. Machine learning enabled the calculations The study is part of broader efforts at the Department of Physics and Astronomy to use machine learning to make materials modeling more accurate and computationally efficient, an area in which Wiktor's group collaborates with professor Paul Erhart's group. "The fact that we can now draw these conclusions is closely linked to developments in machine learning. Ten years ago, we probably would not have been able to do this on such a scale, because the calculations are extensive and time-consuming. It is exciting that these new developments make it possible to carry out studies like this in a comprehensive way," says Wiktor. "The main hope is that the method will be useful for researchers dealing with concrete material problems where the behavior of defects or their importance is unknown, allowing them either to collaborate with us or to use the method themselves." Publication details Tobias Hainer et al., Thermal stabilization of defect charge states and finite-temperature charge transition levels, npj Computational Materials (2026). DOI: 10.1038/s41524-026-02369-3 On arXiv DOI: 10.48550/arxiv.2512.15463 Journal information: npj Computational Materials , arXiv Provided by Chalmers University of Technology