Cohort: 2 

Location: Hangzhou, China 

Application Title: Agentic AI for materials discovery 

Application Subtitle: 

Team Members: Shuming Zhang, Yuping Li, Luyang Zhang, Chenru Duan,Haojun Jia 

The challenge 

Scientific discovery in chemistry is often slowed by fragmentation. Experimental chemists, computational modelers and machine learning specialists typically work with different tools and methods. As a result, collaboration can be slow and progress uneven. 

Many areas of chemistry still rely on quantum mechanical simulations to understand how reactions occur. One commonly used approach is density functional theory (DFT), which models the behaviour of electrons to predict reaction outcomes. While DFT is highly accurate, it is also computationally intensive. Simulating a single reaction pathway can take hours or days. Exploring large reaction spaces or optimising reaction conditions can take much longer. 

Because of these constraints, innovation in areas such as materials science and pharmaceutical development often depends on manual trial and error. This limits the speed at which new compounds or processes can be identified. 

The solution 

A coordinated AI system for chemical research 

Deep Principle developed a research platform designed to connect computational chemistry, machine learning and experimental design within a single workflow. 

At the core of the platform is a multi-agent system that mirrors how research teams operate. Different AI agents are responsible for specific tasks. Some handle computational chemistry. Others focus on machine learning, literature analysis or experiment planning. These agents coordinate their work, allowing simulations, data analysis and knowledge retrieval to proceed in parallel rather than sequentially. 

The system can plan research steps, run simulations, train predictive models and combine results into actionable outputs. Researchers interact with the platform throughout the process rather than receiving results only at the end. 

How it works 

Predicting reaction mechanisms directly 

A key component of the platform is a reaction mechanism prediction model. Chemical reactions proceed through short-lived configurations known as transition states. Accurately identifying these structures is essential for understanding how reactions unfold, but they are difficult and expensive to compute using traditional methods. 

Deep Principle’s model generates transition-state structures directly using a custom SE(3)-equivariant graph neural network. This type of model is designed to respect the physical symmetries of molecules, such as rotation and spatial relationships between atoms. An optimal transport framework is used to ensure chemically plausible pathways. 

This approach allows reaction pathways that previously required hours or days of computation to be generated in seconds. Results are presented in an interpretable way, and researchers can refine outputs or explore alternatives using natural language queries. 

How it differs from conventional approaches 

Covering the full research lifecycle 

Most AI tools in chemistry focus on individual tasks, such as predicting reaction yields or screening candidate molecules. Deep Principle’s platform is designed to support reasoning across the full research lifecycle. 

The system adapts as new data becomes available and updates its recommendations as experiments progress. Because expert reasoning from different disciplines is embedded into the platform, research teams do not need to assemble rare combinations of specialised skills to move projects forward. 

This allows research efforts to scale without increasing reliance on highly specialised cross-disciplinary roles. 

Impact 

• More than 50% of materials simulations automated, compared with a near-zero baseline 
• 500–1,000× faster identification of optimal candidates through virtual screening 
• 30–50% reduction in experimental formulation costs within three to six months 
• 2× improvement in reaction prediction accuracy compared with expert judgement 
• Early pilot projects showing validated R&D cost savings and new revenue from material innovations 
• Approximately 50 users across internal and partner R&D teams, with around 20 API calls per user per day 

Why it matters 

Reducing bottlenecks in chemical discovery 

Chemical discovery increasingly depends on the ability to explore large design spaces efficiently. Deep Principle’s approach shows how AI systems can help reduce bottlenecks created by slow simulation cycles and fragmented workflows. 

By integrating computation, data analysis and experimental planning within a single platform, the system supports faster iteration while maintaining scientific transparency. This can shorten development timelines in materials and chemical research, where progress has traditionally been constrained by computational and coordination limits.