A reaction can depend on the surrounding conditions
Knowing which molecules should react is only part of synthesis. Chemists must also choose a catalyst, ligand, solvent and other conditions. Trying combinations in the laboratory can consume material and time. [1]
Chemma was fine-tuned on 1.28 million chemistry question-and-answer pairs and reaction data. It predicted outcomes across a reaction space, then used each new experimental result to choose a useful next condition to test. [1]
Fifteen rounds with chemists in the loop
For an unreported Suzuki-Miyaura coupling, the human and AI team explored conditions over 15 laboratory runs. They identified tri(1-adamantyl)phosphine as the ligand and 1,4-dioxane as the solvent, and isolated the desired product at 67% yield. [1]
The yield belongs to this reaction under the selected conditions. It does not mean the model succeeds 67% of the time, or that 15 experiments will solve another synthesis problem. [1]
A focused experiment, not an autonomous laboratory
Researchers defined the reaction space, ran the experiments and supplied the measured results. The model proposed the next conditions within that setup. Safety review, practical handling and interpretation remained human responsibilities. [1]
The paper reports benchmark gains and one prospective reaction example. Wider evidence would require other reaction families, independent laboratories and comparisons that include the full cost of model training, failed experiments and human work. [1]
Sources & context
The arXiv manuscript and publisher DOI are versions of the same study.
Large language models to accelerate organic chemistry synthesis
Zhang and colleagues · Nature Machine Intelligence · July 1, 2025