Toward Machine Learning-Enhanced High-Throughput Experimentation

Recent research in machine learning for chemistry includes developments in areas particularly relevant for integration with high-throughput experimentation platforms: synthesis planning, experimental design, and processing of analytical data. Automated, reconfigurable experimentation systems capable of exploring chemistry using ever-smaller quantities of material are under development, allowing efficient exploration of broad regions of chemical space. High-quality experimental endpoints can now be achieved as part of high-throughput experimentation workflows thanks to the automation of powerful analytical instrumentation. Initiatives designed to capture and store the enormous quantities of data that chemists generate are positioned to transform machine learning predictions for chemistry discovery and development. Recent literature suggests that the fields of machine learning (ML) and high-throughput experimentation (HTE) have separately received considerable attention from chemists and engineers, leading to the development of powerful reactivity models and platforms capable of rapidly performing thousands of reactions. The merger of ML with HTE presents a wealth of opportunities for the exploration of chemical space, but the integration of the two has yet to be fully realized. We highlight examples of recent developments in ML and HTE that collectively suggest the utility of their integration. Our analysis highlights the complementarity of the two fields, while exposing a number of obstacles that can and should be overcome to take full advantage of this merger and thereby accelerate chemical research. Recent literature suggests that the fields of machine learning (ML) and high-throughput experimentation (HTE) have separately received considerable attention from chemists and engineers, leading to the development of powerful reactivity models and platforms capable of rapidly performing thousands of reactions. The merger of ML with HTE presents a wealth of opportunities for the exploration of chemical space, but the integration of the two has yet to be fully realized. We highlight examples of recent developments in ML and HTE that collectively suggest the utility of their integration. Our analysis highlights the complementarity of the two fields, while exposing a number of obstacles that can and should be overcome to take full advantage of this merger and thereby accelerate chemical research. a type of probabilistic surrogate model structured akin to a traditional NN that is trained using Bayesian inference (as opposed to stochastic gradient descent). the (abstract) region of input space in which a model’s predictions can be trusted. chemistry performed in a vessel into which chemical species are continuously fed and from which products and unreacted species are continuously removed. a nonparametric distribution over functions; a type of probabilistic surrogate model. an automated, platform-based style of experimentation in which large numbers of experiments are performed rapidly, often thanks to massive parallelization and accelerated and/or simplified analytics. a suite of mathematical approaches for pattern recognition, data analysis, and modeling. a type of surrogate model comprising layers of weights whose values are determined by data. a type of surrogate model comprising an ensemble of decision trees.