The research team develops AI to perform chemical synthesis

Chemistry, with its complex processes and great potential for innovation, has always been a challenge for automation. Traditional computational tools, despite their advanced capabilities, often remain underutilized due to their complexity and the specialized knowledge required to operate them.

Now, researchers in Philippe Schwaller’s group at EPFL have developed ChemCrow, an AI that integrates 18 expert-designed tools, allowing it to navigate and perform tasks within chemical research with unprecedented efficiency. His research is published in Nature Machine Intelligence.

“You may be wondering why a crow?” Schwaller asks. “Because crows are known to use tools well.”

ChemCrow was developed by Ph.D. students Andres Bran and Oliver Schilter (EPFL, NCCR Catalysis) in collaboration with Sam Cox and Professor Andrew White from FutureHouse and the University of Rochester.

ChemCrow is based on a large language model (LLM), such as GPT-4, enhanced by LangChain for tool integration, to autonomously perform chemical synthesis tasks. The scientists augmented the language model with a set of specialized software tools already in use in chemistry, such as WebSearch for Internet-based information retrieval, LitSearch for scientific literature mining, and several molecular and reaction tools for chemical analysis.

By integrating ChemCrow with these tools, the researchers allowed it to autonomously plan and execute chemical syntheses, such as the creation of an insect repellent and various organocatalysts, and even help discover new chromophores, substances fundamental to in the dyes and pigments industries.

What sets ChemCrow apart is its ability to adapt and apply a structured reasoning process to chemical tasks.

“The system is analogous to a human expert with access to a calculator and databases that not only improve the efficiency of the expert, but also make them more factual in the case of ChemCrow, reducing hallucinations,” explains Andres Camilo Marulanda Bran, first author of the study.

ChemCrow receives a prompt from the user, plans ahead how to solve the task, selects the relevant tools, and iteratively refines its strategy based on the results of each step. This methodical approach ensures that ChemCrow not only works with theory, but is also grounded in practical application for real-world interaction with laboratory environments.

By democratizing access to complex chemical knowledge and processes, ChemCrow lowers the barrier to entry for non-experts while increasing the toolset available to veteran chemists. This can accelerate research and development in pharmaceuticals, materials science and beyond, making the process more efficient and safer.