Chemical space is so large as to make a brute force search for molecules with improved properties infeasible. Bayesian optimization methods can accelerate the discovery process by sequentially identifying the most useful experiments to be performed next. However, existing methods have shortcomings that limit their applicability to the molecule search problem. First, they lack scalability to the large amounts of data that are required to successfully navigate chemical space. Second, they are unable to learn feature
representations for the data, which reduces their statistical efficiency in the large data scenario. Third, they cannot collect data using very large batch sizes, which is required when many experiments can be performed simultaneously and finally, they often fail when the search space is discrete as is the case of chemical space. In this talk I will give a brief introduction to Bayesian optimization methods and then I will present different contributions that aim to solve or at least alleviate the aforementioned problems.
Back to Workshop I: Machine Learning Meets Many-Particle Problems