The reinforcement learning-based method steers generative crystal design toward candidates that combine novelty and potential usefulness. This is reported by Nature Machine Intelligence in the description of the publication.
According to this description, generative machine learning methods have already aided crystal search, but are not able to fully explore the space of such materials. The new approach is intended to create new functional materials.
The practical significance of the result cannot yet be evaluated from the available data: the source does not disclose the method’s characteristics, verification results, or specific materials created.
editorial commentary
Why it matters
If results are confirmed by a full publication and experiments, the next observable signals will be specific material samples and independent verification of their properties. The main uncertainty is that the available description does not contain data on the quality of candidates, comparison with baseline methods, or practical applicability.