A team at EPFL have developed a framework that trains Large Language Models to create AI systems that can find the best possible setup or "recipe" for a scientific experiment.
Modern computational tools let scientists explore huge numbers of possible molecules, materials, and chemical reactions. But testing every combination in the lab is slow and costly. So how do researchers choose the best "recipe" for their experiment?
One method called "Bayesian optimization" learns from previous results, predicts which options look promising, and estimates how uncertain those predictions are. This allows researchers to focus their next experiments on the most useful options to test and avoid wasting time and money on less promising possibilities.
Nevertheless, this method doesn't carry over well between scientific fields. For example, an efficient model for choosing chemical reactions might not be suited to materials or molecular design, so each new problem effectively starts from scratch.
The rise of Large Language Models (LLMs) has opened up another route. LLMs already encode broad scientific knowledge and work with information expressed as text. But they too have a weakness: hallucinations. Their apparent confidence does not reliably reflect whether a suggestion is correct, which risks sending scientists down the wrong, time-consuming and expensive path.
Learning from uncertainty
Bojana Ranković and Philippe Schwaller at EPFL's Laboratory of Artificial Chemical Intelligence have developed a method that combines an LLM with a a Gaussian process, a kind of 'doubt detector", the probabilistic model commonly used in Bayesian optimization."
Called GOLLuM (Gaussian Process Optimized LLMs), the method was published in Nature Machine Intelligence. GOLLuM makes it possible to create a sample-efficient and human-interpretable optimizer for scientific discovery.
Rather than asking the LLM to choose experiments directly ("direct prompting"), GOLLuM trains it using the Gaussian process's way of scoring uncertainty and performance of each option. Because of this, as the LLM learns from past experiments, it adjusts how it "organizes" the search space, the "master list" of every possible choice for an experiment. Conditions that lead to similar results are placed closer together, while those with different results are pushed further apart.
"Language models are notoriously bad at knowing when they're wrong. In GOLLuM, that uncertainty becomes the very signal that trains them," says Bojana Ranković, who created the method during her doctoral research at EPFL. "The model reorganizes the search space until experiments that behave alike sit close together."
![As GOLLuM learns, its internal map of the design space reorganizes itself: reactions that behave alike drift together, and high-yielding conditions (yellow) separate from low-yielding ones (purple). Credit: Ranković et al. / EPFL]](https://actu.epfl.ch/public/upload/fckeditorimage/ef/a8/6cd3aad7.jpg)
Finding good options faster
The researchers evaluated GOLLuM across 23 benchmark tasks covering organic synthesis, analytical and process chemistry, materials and catalysis, and molecular property optimization. Each optimization run began with ten low-performing observations, and the team used the same GOLLuM configuration across all benchmark tasks rather than tuning it separately for each problem, showing its versatility.
Across the 23 benchmarks, GOLLuM ranked first on average, finding high-performing experimental conditions more consistently than Bayesian optimization with expert-designed descriptors. Within a budget of 50 experiments, 36.3% of the conditions it tested fell within the top 5% of all possible outcomes, compared with 29.7% for the traditional method. Overall, GOLLuM matched traditional methods using over 40% fewer experiments, indicating that it converges faster towards good solutions.
The researchers also tested what happens when LLMs choose experiments directly, without being coupled with the "doubt detector".
Without the GOLLuM framework, performance was inconsistent, with failure rates ranging from 10% to around 80%. Problems included invented chemical structures, repeated conditions and suggestions outside the permitted search space.
This method opens a different role for language models in experimental research: Rather than relying on their answers alone, researchers can combine their broad representations of scientific information with probabilistic models that explicitly account for uncertainty.
"It's a highly impactful technique that enables us to start experimental optimization campaigns from day one instead of discussing how to best describe experiments and compute descriptors for six months," says Philippe Schwaller. "The key is that we optimize directly on a plain English representation of the experimental procedure."
