To create new materials, scientists need to understand their electronic structure - how electrons are arranged and behave in a substance. Calculating such structures usually requires expensive simulations on supercomputers. That is why researchers train artificial intelligence (AI) models on millions of simulated points to obtain fast predictions.
However, recent work led by Associate Professor Susanta Ghosh from the Department of Mechanical and Aerospace Engineering at Michigan Technological University (Michigan Tech) shows that so much data is not needed. Together with colleagues from the University of California, Los Angeles (UCLA), Ghosh found that AI models need far less information for accurate predictions.
The study was published in the journal npj Computational Materials. The authors report that the models could discard 90% of the training data with virtually no loss of accuracy. Even after removing 99% of the data, most predictions remained highly accurate, and training was at least three times faster. The results apply to a wide range of materials, including metals, water and complex alloys.
How can a model accurately predict electronic properties with less data? According to Ghosh, the secret is in a minimal set that researchers carefully select to cover the diversity of the data. Most of the electronic structure data repeats the same information, while a small, well-chosen sample contains almost everything AI needs for training, Ghosh explained.
These results call into question the need for extensive datasets for predicting electronic structure and could make AI-driven materials development faster and cheaper. Ghosh, a member of the Center for Artificial Intelligence in the Institute of Computing and Cybersystems at Michigan Tech, believes this could strongly influence the field of materials development and modeling.
By identifying representative electronic structure data and cutting training time by at least a factor of three, this work could accelerate the search for complex alloys with AI across a vast space of compositions, he noted.
It is important to understand that this concerns a statistical relationship between the volume of training data and model accuracy, not a proven cause. The study shows that data selection can preserve accuracy, but it does not claim that any dataset can be reduced without consequences. In addition, these are so far the results of scientific work, not a ready-made commercial product.