Summary of Key Points
This article focuses on the startup "Caikeyuan Tu," which specializes in the use of AI to develop new materials. By leveraging two key technological barriers—a global database of energy materials experiments and a cluster of physical models—the company has created a closed-loop system that combines AI with automated experiments, enabling them to quickly identify materials with improved performance. Their business model begins with providing customized research and development services to build a "materials bank," followed by establishing their own production lines to sell the materials. They emphasize practical outcomes over theoretical possibilities. While the industry is highly capital-driven, it is also fraught with bubbles. Caikeyuan Tu stands out by delivering actual products before even officially establishing the company. The company is optimistic about breakthroughs in areas such as solid-state batteries, and both China and the United States have their strengths in this field, with China placing a greater emphasis on practical applications.
Core Technology: Two Key Tools to Enable Precise Material Discovery
Caikeyuan Tu's technological advantages rely on two main components:
1. The world's largest database of energy materials experiments: It took seven years to manually collect data from academic papers, including data from failed experiments, as AI needs to understand what does not work in order to be more accurate. The database contains millions of data points (with the potential to grow to tens of millions by the end of the year). This dataset is invaluable; while the United States has attempted to build a similar database, the high cost of manual data collection has deterred them. AI systems, on their own, have an error rate of over 90% when trying to extract data from papers, making manual effort necessary. Moreover, this database contains real experimental data, which is more useful than theoretical calculations or databases of organic molecules.
2. A cluster of physical models: AI can easily generate speculative ideas (such as creating non-existent materials), but physical models provide a framework that forces it to think according to physical principles, addressing these issues. These models, combined with automated experiments, create a self-reinforcing cycle: a larger database leads to more accurate AI models, which in turn result in more precise experiments and better material properties. This iterative process, similar to that of AlphaGo, enables the company to quickly find the materials needed by its clients, such as overcoming long-standing challenges in catalytic material development for leading energy companies.
Business Model: Building a Materials Bank Before Selling Products
Caikeyuan Tu's business strategy consists of two phases:
1. Lean-capital phase: The company helps clients develop new materials and optimize processes while accumulating its own collection of high-quality materials, which it stores in a "materials bank." This bank allows clients to purchase materials or share in the profits. For example, in the case of volatile rare earth prices, having alternative materials in the bank ensures stability despite supply disruptions.
2. Heavy-capital phase: The company then establishes its own production lines to sell materials, such as those used in green electrochemical synthesis and solid-state batteries. They choose not to offer SaaS software, fearing that clients might steal the database (by purchasing the usage rights and exporting all the data). Selling materials is more profitable than providing services, as the concept of "AI for materials" without practical applications is merely theoretical.
Industry Context: Capital Boom with Many Hopes and Challenges
The field of AI for science is attracting significant capital, but it is also surrounded by bubbles, with many companies lacking tangible results. Caikeyuan Tu stands out by receiving orders before even officially founding the company. This approach is rare in the industry. Li Hao, the company's founder, asserts that their claim of a 100-fold increase in research and development efficiency is based on real data and not just empty promises.
Future Directions: Solid-State Batteries as a Potential Target for Disruption
Li Hao believes that solid-state batteries are likely to be the first area to be transformed by AI-driven material advancements. Currently, only 8% of solid-state electrolyte materials have been explored, leaving the remaining 92% as uncharted territory. Companies often claim to mass-produce solid-state batteries by next year, but this is often driven by fear of being outdone by competitors. He estimates that it will still take about two years for solid-state batteries to become widely adopted. Other potential areas for disruption include ammonia synthesis (where AI could find more cost-effective electrochemical methods using renewable energy) and traditional chemicals.
Competition between China and the United States
Both China and the United States have their strengths in the AI for materials field:
- China: Has larger teams (a research group of 20 people is considered large in the United States, but this is common in China), strong software development capabilities, and a focus on materials that can be immediately applied in industries such as batteries and catalysis, making their approach more practical and less flashy.
- United States: Possesses substantial computational power and funding, but their research often focuses on more niche areas (such as superconducting materials) that may not generate immediate industrial value. Li Hao believes that China's practical-oriented approach will drive faster industry progress.
In Conclusion
AI for materials is not about fancy concepts; it relies on real data, physical models, and closed-loop experiments to develop and sell new materials. Breakthroughs in areas like solid-state batteries could have a significant impact on our lives, similar to how lithium batteries transformed mobile phones.