Reproduce “Foundation-Model Surrogates Enable Data-Efficient Active Learning for Materials Discovery” on OpenResearch

About this paper

Jeffrey Hu, Rongzhi Dong, Ying Feng, Ming Hu, Jianjun Hu

The study introduces In-Context Active Learning (ICAL), which employs the TabPFN foundation model as a surrogate for data-efficient materials discovery. This framework consistently reduced the number of experimental evaluations needed to find optimal materials by providing superior uncertainty quantification in small-data settings.

Key insights

• Foundation models like TabPFN can effectively address the trade-off in active learning by offering both high predictive capacity and reliable uncertainty quantification in data-scarce domains.
• Improved active learning efficiency stems primarily from the surrogate model's superior uncertainty calibration rather than solely higher regression accuracy.
• Feature engineering and the choice of acquisition function significantly influence active learning performance, with normalized and hybrid functions enhancing stability in low-data regimes.

Summary by alphaXiv.