Machine learning methods improve semiconductor band gap prediction

Imagine you’re cooking. They are trying to develop a unique flavor by mixing spices they have never paired before. It can be difficult to predict what this will look like. You want to make something tasty, but it will taste badly: a waste of time and ingredients.

But what if you have a machine that can tell you exactly what your formulation will look like? This is a type of technology developed by researchers at Kyoto University for the band gap in semiconductor materials. This work is featured in the Journal Computational Materials Science.

Such materials are constantly in demand for new devices and improved performance. Accurate prediction is essential as band gaps are the most important factor in determining semiconductor properties.

Unfortunately, traditional methods of calculating bandgaps are expensive and are based on the properties of absolute zero materials, so they are not accurate enough at room temperature. For this reason, researchers have been trying to develop machine learning methods to achieve faster and more accurate predictions.

The Kyoto team has begun developing a machine learning model that is integrated with neural networks. This new ensemble learning method uses data based on measurements of known compounds to predict the physical properties of unknown materials.

“Our model allows predictions based solely on the composition of the compounds,” says corresponding author Tanabe katsuaki.

The researchers used data from almost 2,000 semiconductor materials tested in six different neural networks. They found that the incorporation of conditional generated enemy networks, or CGANs, and message-passing neural networks (MPNNs), contribute significantly to improving prediction accuracy. The resulting model achieved the highest prediction accuracy of any existing model developed for the same purpose.

“The computational load of ensemble learning models is lightweight and can be run within hours on a typical laptop,” Tanabe continues. “And I can confidently say that this method allows for quick and very accurate predictions.”

On the other hand, the more accurate the machine learning model, the more internal mechanisms will sound. They are powerful for ad hoc calculations and prediction, but they are neither versatile nor scalable, so they require more work.

“We are also developing other ways to interpret the correlation between the properties of different materials and band gaps,” adds Tanabe.

Nevertheless, this integrated model demonstrates that ensemble models using neural networks are promising in this field and potentially useful for the development of new generations of semiconductors.