Deep Principle Launches MPA, a 'Materials AlphaFold' Achieving SOTA on 40 Industrial Tasks

Deep Principle launched MPA, a materials property prediction foundation model built with LLM-style pre-training, mid-training, and fine-tuning.

MPA uses a hybrid readout architecture and physics-guided alignment to better match real experimental data.

The model reportedly outperformed leading molecular property models on 40 tasks, including tough scaffold-split evaluations.

Deep Principle has integrated MPA into its Agent product and sciclaw platform, signaling broader use in real-world materials discovery.