(192b) Accelerated Process and Product Development with Small Data and Machine-Learning

Data-driven machine-learning (ML) offers the possibility to dramatically reduce the resources needed for reaction, formulation, and process optimization through predictions of behavioral trends, variable effects, and optimal conditions. However, most ML algorithms require large datasets (i.e. 100’s-1000’s of points) to provide accurate predictions, and experimental data collection is expensive in human and material resources. The challenge of building large datasets can be averted with Bayesian optimization (BO) frameworks. These guide experimental campaigns and carefully build training datasets that ultimately reduce the number of experiments required to achieve reliable understanding of variable effects and reaction behavior. However, their implementation is not straightforward in many cases.

Developing an enhanced BO approach, we elucidate a reaction-agnostic pathway for the design and implementation of intelligent experimental campaigns using material formulations, as well as catalytic and electrochemical transformations as case studies. Working at the intersection of ML and chemistry, we created SuntheticsML, an online, user-friendly, ML platform available to researchers that does not require coding or ML expertise. It leverages small datasets (starting with only 5 data points) to enable the development of new chemical processes, materials, and formulations with unprecedented efficiencies and up to 32x less time and experiments.

The insights from this work innovate on the future of reaction engineering, material design, and process optimization while lowering the barrier to the implementation of new chemical technologies. SuntheticsML accelerates lengthy and complex traditional optimization campaigns, harnessing the power of ML and small data to fast-track innovation, sustainability, and digitalization in the chemical industry, further reducing development waste, emissions, and resource consumption by up to 97%.