Deep Principle Launches MPA for Materials Science Breakthroughs

🔑 Enhanced Key Takeaways

•MPA employs a three-stage training paradigm (pre-training, mid-training, fine-tuning) adapted from large language models, with an innovative physics-guided alignment in the mid-training stage to bridge the gap between computational models and real experimental data.
•The model's architecture includes a novel "Hybrid Readout" design featuring two parallel pathways: an attention-based pooling pathway for properties dependent on overall molecular character and an atom-wise summation pathway for properties scaling with molecular size, dynamically weighted by a learnable parameter α.
•Deep Principle, a Beijing-based AI for science startup founded in 2024, has secured significant funding, including a Pre-A round of 100 million yuan in March 2025 and a Series A round of over RMB 100 million in November 2025, with investors like Lenovo, Baidu, Alibaba Entrepreneurs Fund, and Ant Group.
•Beyond MPA, Deep Principle's product portfolio includes the ReactiveAI platform, Agent Mira (an intelligent agent for autonomous materials discovery), and the development of an automated laboratory called the "AI Materials Factory."
•Deep Principle has also developed other generative AI models, such as the diffusion model React-OT, which achieves transition-state structure predictions significantly faster than traditional quantum chemistry methods.

🛠️ Technical Deep Dive

MPA utilizes a three-stage training paradigm: pre-training, mid-training, and fine-tuning, adapted from large language model methodologies.
The mid-training stage incorporates physics-guided alignment to instill "physical intuition," aiming to reconcile computational idealizations with real-world experimental conditions.
The model features a "Hybrid Readout" architecture with two distinct parallel pathways:
- An attention-based pooling pathway designed for properties that depend on the overall molecular "character," such as boiling point.
- An atom-wise summation pathway for properties that scale with molecular size, like enthalpy of formation.
A learnable parameter, α, dynamically weights these two pathways based on the specific property being predicted.
In extensive benchmarking, MPA demonstrated state-of-the-art performance, outperforming five leading molecular property prediction models (ChemBERTa, ChemProp, Chemeleon, Uni-Mol2, and Suiren) on 38 out of 40 properties in random split tests.
Deep Principle also employs diffusion models, such as React-OT, which can predict transition-state structures in 0.4 seconds with high accuracy.
The company integrates large language models for tasks like the generative design of transition metal complexes.

🔮 Future ImplicationsAI analysis grounded in cited sources

MPA will significantly accelerate the discovery and development cycle for new materials.

By bridging the gap between theoretical predictions and experimental reality and achieving state-of-the-art performance across numerous industrial tasks, MPA can drastically reduce the time and cost associated with traditional materials research and development.

Deep Principle's integrated AI platforms will lead to the widespread adoption of autonomous materials discovery laboratories.

With products like Agent Mira and the development of the 'AI Materials Factory,' combined with MPA's predictive capabilities, Deep Principle is building a full-stack solution for automated R&D, moving AI for science from concept to industrial reality.

The application of LLM techniques to physical sciences, as demonstrated by MPA, will become a standard approach for developing foundation models in other scientific domains.

MPA's success in adapting LLM training methodologies for materials science suggests a transferable paradigm for creating robust, general-purpose AI models that can handle complex scientific data and challenges in fields beyond chemistry and materials.

⏳ Timeline

2024

Deep Principle founded

2025-03

Secured Pre-A funding round of 100 million yuan from funds supported by LENOVO and Baidu

2025-11

Raised over RMB 100 million in Series A round, co-led by Alibaba Entrepreneurs Fund – Greater Bay Area and Ant Group

2026-03

Closed Series A2 funding round led by Golden Ant Investment

2026-06-02

Launched MPA (Materials Property Axiom) AI foundation model

Deep Principle Launches MPA for Materials Science Breakthroughs

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (10)

👉Related Updates