How to Accelerate R&D and Optimize Experiment Planning with Machine Learning and Data Science

Abstract

Accelerating R&D is essential to address some of the challenges humanity is currently facing, such as achieving the global sustainability goals. Today’s Edisonian approach of trial-and-error still prevalent in R&D labs takes up to two decades of fundamental and applied research for new materials to reach the market. Turning around this situation calls for strategies to upgrade R&D and expedite innovation. By conducting smart experiment planning that is data-driven and guided by AI/ML, researchers can more efficiently search through the complex - often constrained - space of possible experiments and find or hit the global optima much faster than with the current approaches. Moreover, with digitized data management, researchers will be able to maximize the utility of their data in the short and long terms with the aid of statistics, ML and visualization tools. In what follows, we describe a framework and lay out the key technologies to accelerate R&D and optimize experiment planning

Bayesian Optimization as a Sustainable Strategy for Early-Stage Process Development? A Case Study of Cu-Catalyzed C–N Coupling of Sterically Hindered Pyrazines

“…In particular, we believe that BO is well-suited for investigating large to very large reaction spaces. Therefore, it is appropriate for early-stage process optimization programs. , For smaller reaction spaces (typically less than 5000 reactions), other techniques may be competitive.…”

“…Therefore, it is appropriate for early-stage process optimization programs. 108,109 For smaller reaction spaces (typically less than 5000 reactions), other techniques may be competitive.…”

Highlights from the 56th Bürgenstock Conference on Stereochemistry 2023

“…This industry-focused session was complemented by research endeavours of industrial JSP fellows, whose work covered innovations in scalable pyrazole synthesis, 28 synthesis with spirocycles, 29 and machine-learning enabled reaction optimisation. 30 The same aernoon saw a return to a rapid relay of short talks, further expanding the conference coverage to areas including innovations in iron catalysis, 31 Lewis base catalysis, 32 and literature-based synthetic yield predictions in synthesis. 33 11 Back to natural products…”

Combining Bayesian optimization and automation to simultaneously optimize reaction conditions and routes

“…As R&D undergoes continuous digitization and is fueled by the abundance of untapped data and impressive advancements in articial intelligence (AI), there arises a growing imperative to adopt a data-driven methodology. [19][20][21][22] This adoption will empower scientists and researchers to utilize AI and to learn from past experiments, recognize patterns in the data, and ultimately suggest the next best experiments expediting R&D. 23 One prominent trend in this data-driven approach involves the rising popularity of Bayesian optimization 24 (BO) and reinforcement learning in recent years. [25][26][27][28][29][30][31][32] Notably, deep reinforcement learning techniques were used along with domain knowledge of chemistry to improve the yield of four reactions carried out in microdroplet reactors.…”

“…As R&D undergoes continuous digitization and is fueled by the abundance of untapped data and impressive advancements in artificial intelligence (AI), there arises a growing imperative to adopt a data-driven methodology. 19–22 This adoption will empower scientists and researchers to utilize AI and to learn from past experiments, recognize patterns in the data, and ultimately suggest the next best experiments expediting R&D. 23…”