Machine learning has become increasingly important in various fields, including chemistry and materials science. These areas often deal with small datasets, which has led to the development of advanced machine-learning techniques tailored to the specific needs of chemistry.
Scientists at EPFL demonstrate how GPT-3 can transform chemical analysis, making it faster and more user-friendly.
In a groundbreaking study, scientists led by Berend Smit at EPFL have tapped into the power of large language models like GPT-3 to simplify chemical analysis using artificial intelligence. GPT-3 is famous for its ability to understand and generate human-like text, forming the basis of popular AI ChatGPT.
Published in Nature Machine Intelligence, the study reveals a new approach that streamlines chemical analysis. Instead of directly asking GPT-3 chemical questions, the researchers fine-tuned it with a small dataset converted into questions and answers. This creates a specialized model capable of providing accurate chemical insights.
They fed GPT-3 a curated list of Q&As, like asking if a specific alloy is single-phase. By training it on known answers from the literature, they refined the AI to only respond with a yes or no.
In tests, the model answered over 95% of diverse chemical problems correctly, often outperforming state-of-the-art machine-learning models. What’s remarkable is how simple and fast this method is. While traditional models take months to develop and require extensive expertise, this approach takes just five minutes and needs zero prior knowledge.
The study’s implications are significant. It introduces a method as easy as a literature search, applicable to various chemical problems. Being able to ask questions and get accurate answers can revolutionize chemical research.
As the authors note, querying foundational models like GPT-3 could become a routine way to kickstart projects, leveraging collective knowledge encoded in these models. In short, as Smit says, “This is going to change the way we do chemistry.”
Journal Reference:
- Jablonka, K.M., Schwaller, P., Ortega-Guerrero, A. et al. Leveraging large language models for predictive chemistry. Nat Mach Intell (2024). DOI: 10.1038/s42256-023-00788-1