Machine learning method improves semiconductor band gap predictions

Imagine you’re cooking. You’re trying to develop a unique flavor by mixing spices you’ve never combined before. Predicting how this will turn out could be tricky. You want to create something delicious, but it could end up tasting awful: a waste of time and ingredients.

But what if you had a machine that could tell you exactly how your concoctions will turn out? That’s the kind of technology that researchers at Kyoto University have developed for the band gap of semiconductor materials. The work is published in the journal Computational Materials Science.

Such materials are constantly sought after in the development of new devices and improved performance. The most important factor in determining the properties of semiconductors is the band gap, so accurate predictions are essential.

Unfortunately, the conventional method of calculating a band gap is expensive and not accurate enough at room temperature, since it is based on a material’s properties at absolute zero. For this reason, researchers have been trying to develop a machine learning method to achieve more rapid and precise predictions.

KyotoU’s team set out to develop a machine learning model integrated with neural networks. This new ensemble-learning method predicts the physical properties of unknown materials, using data based on measurements of known compounds.

“Our model enables prediction based solely on the composition of a compound,” says corresponding author Katsuaki Tanabe.

The research team used data from almost 2,000 semiconductor materials tested on six different neural networks. They found that the incorporation of conditional generative adversarial networks, or CGAN, and message passing neural networks, or MPNN, contributed significantly to an improvement in forecast accuracy. The resulting model has achieved the highest prediction accuracy among existing models that have been developed for the same purpose.

“The computational load of the ensemble learning model is light and can be performed within a few hours on a typical laptop PC,” continues Tanabe. “And we can confidently say that this method enables fast and highly accurate forecasting.”

On the other hand, the higher the accuracy of machine learning models, the murkier their internal mechanisms become. Although they are powerful for ad hoc calculations and predictions, they are not versatile or scalable, so more work is required.

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

Still, this integrated model has demonstrated that ensemble models utilizing neural networks are promising for this field, and potentially useful for developing a new generation of semiconductors.