Computer searches can propose enormous numbers of crystal structures. Before a laboratory spends time on one, researchers want to know whether a similar material has been made and which ingredients and process might work. ^[1]

This study turned crystal structures and synthesis records into text-like inputs for fine-tuned language models. Separate models estimated synthesizability, classified the broad preparation method and proposed precursor materials. ^[1]

The synthesizability classifier reported 98.6% accuracy on a comparison set of 1,512 structures that had experimental records, while running faster than the tested physics-based calculations. The method classifier reached 91.0% accuracy on its held-out set. ^[1]

That 91.0% average hides an imbalance. Accuracy was about 98% for solid-state synthesis, the dominant class in the training data, but about 51% for solution-based methods. A model can look strong overall while serving a smaller category poorly. ^[1]

The labels come from published materials records, which favor successful and commonly reported experiments. The study mainly evaluates predictions against held-out database entries. It does not show that every newly proposed crystal and recipe was made independently in a laboratory. ^[1]