ADVANCED MATERIALS & PROCESSES | MAY/JUNE 2024 1 7 In 2020, Antono et al.[9] at Panasonic and Citrine Informatics designed organic semiconductors for lighting and flexible electronic applications using an iterative ML framework. Organic semiconductors consist of small molecules that conduct electron holes efficiently. Molecular design is difficult because there are billions of possible molecules, so rapid screening is necessary to identify the most promising for investigation. In this work, the authors used ML trained on very expensive atomistic simulations to develop a model for predicting hole mobility as a function of molecular structure. Starting with an initial training dataset of 32 points, Citrine made ML predictions of high-value candidate molecules, selected based on the uncertainty of the ML predictions. Panasonic then applied their state-of-the-art simulations to assess the performance of the candidates (a handful at a time), with the results used to update the training data for the ML model. Over the course of several weeks in this active learning iterative approach[10], see Fig. 4, the ML model identified organic semi- conductors 26% more performant than the incumbent material, resulting in several patented molecules for commer- cialization. Furthermore, these materials belonged to a novel class of molecules that had not been considered as organic semiconductors before, providing the materials engineers with new areas to explore. The value in this ML success came from the replacement of expensive and time-consuming simulations with a cheaper ML surrogate model and accelerating the search for high- performance materials within a very large design space using uncertainty- driven active learning algorithms. SELF-DRIVING LABORATORIES The combination of high-throughput experiments, literature data, simulations, and advanced machine learning techniques has significant potential to accelerate materials discovery. This requires automation of experiments via robots and smart exploration, Fig. 3 — Changes in composition during optimization of pitting potential for fixed chloride ion concentration for a variety of alloy classes, based on an ML model. Reprinted with permission from Ref. 7 under the terms of the Creative Commons CC BY license. (a) (b) (c) (d) (e)
RkJQdWJsaXNoZXIy MTYyMzk3NQ==