Matbench Discovery

Start with a narrow scientific question for benchmark for evaluating machine-learning models on discovery of stable inorganic crystal structures. Write down the hypothesis, search scope, design objective, constraints, and the evidence that would count as validation before asking the AI system to explore the problem.

Use Matbench Discovery to generate a first set of candidates, analyses, structures, literature findings, segmentations, or experimental suggestions. Inspect the underlying evidence and intermediate outputs instead of accepting only the final result. Where possible, compare against a trusted baseline or an independent method.

Validate the result using the standards of the domain: experimental measurement for chemistry or biology, held-out benchmarks for predictive models, expert review for literature synthesis, or quantitative ground truth for imaging. Record failed predictions and negative results as well as successful examples.

Before scaling the workflow, confirm licensing and data-use terms, model and dataset versions, reproducibility, privacy or biosafety requirements, and the limits of the training distribution. Re-check the primary documentation when models or hosted services change.