Materials AI can shed 90% of training data with little accuracy loss, study finds
The key was selecting data that preserved its diversity. Even a deeper cut left most predictions highly accurate, with training at least three times faster.
Loading page…
The key was selecting data that preserved its diversity. Even a deeper cut left most predictions highly accurate, with training at least three times faster.
Listen to this story
Michigan Tech associate professor Susanta Ghosh and UCLA collaborators report that selecting representative electronic-structure examples can preserve most of the information AI models need while shrinking their training sets. With 90% of the data removed, prediction accuracy was nearly unchanged; at a 99% reduction, most predictions remained highly accurate and training was at least three times faster. The results suggest a way to reduce the data burden in materials prediction, though the gains depend on careful sample selection, not simply using any smaller dataset.
The study covers metals, water and complex alloys, rather than testing only one material.
Electronic-structure training examples contain substantial repetition, which researchers used to select a smaller, more diverse sample.
Ghosh says the approach could speed searches across many possible complex-alloy compositions; that discovery benefit is a proposed application, not a measured result.
AI models used in materials research retained nearly the same prediction accuracy after researchers removed 90% of their training data, according to Michigan Technological University. The study points to a less data-intensive way to predict how electrons behave in materials—a computational task that typically requires expensive supercomputer simulations.
The work, published in npj Computational Materials, was led by Michigan Tech associate professor Susanta Ghosh in collaboration with researchers from UCLA. Its findings cover materials including metals, water and complex alloys, rather than a single material.
Materials development often begins with electronic structure: how electrons are arranged and behave within a substance. Researchers use AI to make faster predictions of that structure, training models on millions of data points generated by simulations. The simulations supply the examples from which the models learn.
Ghosh’s team instead used a smaller dataset carefully selected to capture the diversity of the full collection. The finding is not that any small sample will work equally well. The researchers’ explanation rests on choosing representative examples from electronic-structure data that contain substantial repetition.
Much of the electronic structure data repeats the same information. A small, well-chosen sample captures nearly everything the AI needs to learn.
Susanta Ghosh, associate professor, Michigan Technological University
The researchers also pushed the reduction further. After removing 99% of the training data, most predictions remained highly accurate, while the models trained at least three times faster, according to the university’s account. That is a more qualified accuracy result than the virtually unchanged performance reported for the 90% reduction.
The measured speed improvement concerns model training. The potential benefit Ghosh identifies goes further: using representative electronic-structure data could accelerate the search for complex alloys across many possible compositions. Faster materials discovery is the proposed payoff of the training result, rather than the same measurement.
For researchers developing these prediction models, the study challenges the need to retain extensive training datasets. Its practical lever is the selection step: preserve the range of information the model needs, while removing examples that repeat it.
Loading discussion...
Join the conversation
Explain where you would draw the line.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.