Semiconductors are central to modern technology. They are used in computer chips, solar cells, sensors, LEDs, and communication devices. Before researchers make new semiconductor materials in the lab, they often test them first using quantum mechanical simulations.
One of the main tools for this work is density functional theory, or DFT, a computer-modelling method that skips tracking every single electron and instead uses their overall cloud density to quickly calculate a material's atomic structure and energy.
DFT gives researchers a practical way to calculate how electrons behave in materials and is widely used because it balances accuracy with computational cost. But it has a long-standing weakness: it generally underestimates the material’s band gap.
The band gap is the energy difference that controls how easily electrons move through a material. It helps determine whether a material acts as a metal, a semiconductor, or an insulator. It also shapes how a material absorbs light. This matters for technologies such as solar cells, photodetectors, and other electronic devices.
The problem becomes more serious for narrow-gap semiconductors. These materials have small band gaps and are important for optoelectronics, thermoelectrics, and spintronics. Standard DFT methods can wrongly predict that some of them behave like metals. That can send researchers down the wrong path when screening new materials.
A new method: DD-r²SCANH
A team of researchers at EPFL, Stefan Riemelmoser, Xun Xu, and Alfredo Pasquarello, has now developed a method called DD-r²SCANH that addresses this problem.
The innovation of DD-r²SCANH is that it combines two mathematical tools: First, a modern density functional, called r²SCAN. This is a math formula that quickly predicts how electrons interact locally, without having to look at the entire electron density of a material. Second, a smart algorithm that automatically tweaks the r²SCAN formula by mixing in long-range electron-electron interactions.
In such “dielectric-dependent hybrid functionals”, the amount of mixing depends on the material’s dielectric constant, which describes how strongly the material responds to an electric field. If this value is wrong, the final prediction can also be wrong.
Better dielectric constants and more accurate band-gap predictions
The researchers found that r²SCAN predicts dielectric constants much more accurately than the PBE functional, the most widely used density functional in physics and materials science for predicting how atoms bond and materials behave.
Across 39 well-characterized semiconductors and insulators, r²SCAN reduced the average error in dielectric constants from 26% to 9%.
These improvements in electronic structure are also inherited by the dielectric-dependent hybrid functionals. DD-r²SCANH reached an average band-gap error of 7%, compared with 27% for the standard PBE-based dielectric-dependent hybrid functional. It performed especially well for difficult narrow-gap semiconductors such as germanium and indium arsenide, where PBE can predict a false metallic state.
Performance across multiple properties
The team also tested other properties, including ionization potentials, effective masses, and crystal structures. DD-r²SCANH improved ionization-potential predictions while keeping good accuracy for the other properties. Its performance approached that of advanced many-body methods, which are highly accurate but much more computationally demanding.
The work provides materials scientists with a more reliable tool for predicting semiconductor properties. Since r²SCAN predicts the dielectric constants accurately, the mixing parameter in DD-r²SCANH can by determined from first principles. This means that the method can also be applied for materials discovery. It could help researchers screen materials more efficiently for future electronic and energy technologies.
Other contributors
Beijing Computational Science Research Center
Funding
CSEA-EPFL, CSCS
Reference
Stefan Riemelmoser, Xun Xu, Alfredo Pasquarello. Dielectric-dependent hybrid functional based on meta-GGA. Nature Communications, 04 July 2026. DOI:
10.1038/s41467-026-75146-x