Inside Deep Principle: The AI That Runs Its Own Chemistry Experiments to Discover New Materials

HANGZHOU, CHINA / ACCESS Newswire / October 29, 2026 /Inside a lab in Hangzhou, an AI system named Mira recently pushed a single-step reaction's selectivity from 51.82% to 85% in three months, without a chemist manually running the experiments. That result is exactly what Deep Principle 's autonomous lab, the AI Materials Factory, was built to produce, AI-generated predictions tested, corrected, and improved in a continuous loop until they become real, usable materials.

Inside the AI Materials Factory

Deep Principle calls it an L4 lab, meaning experiments run with minimal manual intervention from start to finish. Inside, Mira designs each experiment, operates the equipment, and collects the resulting data on its own. The company describes this as a dry-wet closed loop, meaning AI prediction in the "dry" lab and high-throughput physical testing in the "wet" lab continuously feed each other, with results flowing back automatically to improve the next round of predictions.

"Now we run the whole thing in-house," said Dr. Haojun Jia, co-founder and CEO of Deep Principle. "The client gives us a target, our AI generates candidates, our lab synthesizes and tests them, and we hand back a real, verified compound with real data. That data then trains our model again, so the next project is faster."

The lab runs around the clock and has produced measurable results for clients. In one specialty chemicals project, Deep Principle says it improved production yield from 72% to nearly 90%. In another, screening the top 20 candidate molecules in two weeks improved active ingredient retention rates by 20% across the board.

The Models Powering the Loop

The lab's throughput depends on the models feeding it. Deep Principle's MPA model has reached state-of-the-art performance on 35 of 40 real experimental property-prediction tasks, cutting average prediction error by 14.6% under the testing condition closest to real-world R&D. Its SAGA model handles complex materials generation tasks with what the company describes as stronger multi-objective optimization than comparable models built by much larger technology companies.

The clearest example of what the combination makes possible is OA-ReactDiff and React-OT, a pair of models that compress chemical reaction transition-state generation, a step that traditionally takes days or months, down to roughly 0.4 seconds.

Deep Principle says it is now extending the same lab and model architecture into new frontier materials, including computing-infrastructure materials, superconductors, and liquid coolants for data centers.