Algorithm predicts the compositions of new materials

A machine-learning algorithm that can predict the compositions of trend-defining new materials has been developed by RIKEN chemists1. It will be handy for finding materials for applications where there is a trade-off between two or more desirable properties.

Artificial intelligence has great potential to help scientists find new materials with desirable properties. A machine-learning algorithm that is trained with the compositions and properties of known materials can predict the properties of unknown materials, and save a lot of time in the lab.

But discovering new materials for applications can be tricky, as there is often a trade-off between two or more material properties. One example is organic materials for organic solar cells, where it is desired to maximize both voltage and current, notes Kei Terayama, who was at the RIKEN Center for Advanced Intelligence Project and is now at Yokohama City University. “There is a compromise between voltage and current: a material that exhibits a high voltage will have a low current, while one with a high current will have a low voltage.”

Materials scientists so often want to find “out-of-trend” materials that deserve the usual trade-off. But unfortunately, conventional machine-learning algorithms go much better at detecting trends than discovering materials that stand against them.

Now Terayama and his colleagues have developed a machine-learning algorithm, BLOX (BoundLess Object Free eXploration), which can find materials out-of-trend.

The team demonstrated the power of the algorithm by using it to identify eight out-of-trend molecules with a high degree of photoactivity from a drug discovery database. The properties of these molecules exhibited good agreement with those predicted by the algorithm. “We were concerned about the accuracy of the calculation, but were pleased to see that the calculation was correct,” says Terayama. “This demonstrates the potential of computationally driven material development.”

BLOX uses machine learning to generate a prediction model for key material properties. It does this by combining data for materials randomly selected from a materials database with experimental as calculation results. BLOX then uses the model to predict the properties of a new set of materials. Of these new materials, BLOX identifies the one that deviates most from the overall distribution. The properties of that material are determined by experiment or calculations and are then used to update the model for learning machines, and the cycle is repeated.

Importantly, unlike many previous algorithms, BLOX does not impose any restrictions on the range of material structures and compositions that can be explored. It can reach so far and wide in its search for foreign materials.

The team has made BLOX freely available online.