Electronic structure calculations can provide insights and ideas for catalyst design that would be difficult to achieve with experimental techniques alone, and thus have seen a rapid increase in popularity in recent years. However, these calculations are inherently slow even with modern computational power, and system sizes and effects that can be studied are typically limited to very small, ideal systems. Recently, it has been hypothesized and demonstrated that machine-learning functions can learn the form of electronic structure calculations, thus avoiding the need to perform many of these expensive calculations by predicting the outcome of new calculations. This has the potential to drastically accelerate such atomistic calculations, opening the door to phenomena previously not thought to be accessible. However, we argue that these procedures will only have value if two conditions are met: (1) the training of machine-learning models becomes simple and intuitive for the average user, and (2) feedback can be provided on when to trust the machine-learning predictions, versus requiring a full electronic-structure calculation. In this talk, I will give a background on machine-learning acceleration for atomistic calculations in general, along with several strategies that are employed. Examples will be given of how these calculations can accelerate challenging atomistic calculations, while providing trust to the results. An emphasis will be put on current bottlenecks and roadblocks that prevent the widespread adoption, focusing on the two conditions described above.

Possibilities and roadblocks in the machine-learning acceleration of atomistic calculations

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.