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

Deep Principle, a Beijing-based AI for science startup, has launched MPA (Materials Property Axiom), a foundation model for materials property prediction that achieves state-of-the-art results across 40 real-world industrial tasks. Described as a "Materials AlphaFold," MPA represents a breakthrough in applying large language model training methodologies to the physical sciences. MPA addresses a fundamental challenge in AI-driven materials science: models that achieve high accuracy on computational benchmarks consistently fail when confronted with real experimental data. The gap between theoretical prediction and laboratory reality has limited the practical adoption of AI in materials research. Deep Principle's solution borrows the three-stage training paradigm from large language models — pre-training, mid-training, and fine-tuning — adapted for the materials domain. The mid-training stage is particularly innovative, using physics-guided alignment to build "physical intuition" that bridges the gap between computational idealizations and real-world experimental conditions. The model's architecture features a novel "Hybrid Readout" design with two parallel pathways: an attention-based pooling pathway for properties like boiling point that depend on overall molecular "character," and an atom-wise summation pathway for properties like enthalpy of formation that scale with molecular size. A learnable parameter α dynamically weights the two pathways depending on the property being predicted. In extensive benchmarking, MPA outperformed five leading molecular property prediction models including ChemBERTa, ChemProp, Chemeleon, Uni-Mol2, and Suiren. It achieved SOTA on 38 out of 40 properties in random split tests (14.0% average error reduction) and 35 out of 40 in the more challenging scaffold split scenario (14.6% reduction). The strongest improvements came in scaffold split scenarios where test molecules have entirely novel structures never seen in training — precisely the scenario most common in real research. MPA has been integrated as a skill into Deep Principle's Agent product and is available through the company's sciclaw platform. The technical report is published on Deep Principle's website, demonstrating a practical path for aligning AI models with the physical laws governing real materials behavior.