Exploitation and exploration are two concepts in materials development. The exploitation is to improve known materials towards higher performance, and the exploration is to know the performance of completely new materials, which are unimaginable based on previous knowledge. In the field of materials informatics, these concepts correspond respectively to interpolation and extrapolation, and the majority of papers is concerned with the interpolation. Here, we implemented a catalyst recommender system that can achieve both the interpolation and the extrapolation on the oxidative coupling of methane (OCM). The system is based on the evidence theory, which is a mathematical framework to quantify the certainty of a hypothesis by combining multiple pieces of evidence. It considers combinatorial effects in catalyst design based on similarity of combinations between elements and supports, and assigns probabilities (p_positive, p_non-positive, p_uncertain) to each catalyst. Note that p_positive is a probability that a catalyst shows a product yield higher than a predefined threshold and p_uncertain is a probability that the performance of a catalyst is uncertain. The system was trained on a past catalyst dataset, where the OCM performance of 376 quaternary catalysts (M1-M2-M3/Support) was recorded. New catalysts were recommended by using p_positive + xp_uncertain as the selection probability. They were prepared and evaluated by high-throughput experimentation. The obtained data were added to the dataset so as to update the system in the manner of active learning. Such the loop was recursively carried out, successfully leading to high-performance catalysts that were dissimilar to those contained in the original dataset.

Active learning loop based on the catalyst recommender system.

Sunao Nakanowatari

Appearances