Accelerated Exploration of Heterogeneous CO2 Hydrogenation Catalysts by Bayesian Optimized High-throughput and Automated Experimentation.

Abstract

Automated high-throughput platforms and Artificial Intelligence (AI) are already accelerating discovery and optimization in various fields of chemistry and chemical engineering. However, despite some promising solutions, little to no attempts have targeted the full heterogeneous catalyst discovery workflow, with most chemistry laboratories continuing to perform research with a traditional one-at-a-time experiment approach and limited digitization. In this work, we present a closed-loop data-driven approach targeting the optimization of catalysts’ composition for the direct transformation of carbon dioxide (CO2) into methanol, by combining Bayesian Optimization (BO) algorithm, automated synthesis by incipient wetness impregnation and high-throughput catalytic performance evaluation in fixed bed mode. The BO algorithm optimized a four-objective function simultaneously (high CO2 conversion, high methanol selectivity, low methane selectivity, and low metal cost) with a total of 11 parameters (4 supports, 6 metals salts, and one promoter). In 6 weeks, 144 catalysts were synthesized and tested, with limited manual laboratory activity. The results show a significant improvement in the objectives at the end of each iteration. Between the first and fifth catalyst generation, the average CO2 conversion and methanol formation rates have been multiplied by 5.7 and 12.6 respectively, while simultaneously reducing the methane production rate by 3.2 and dividing the metal cost by 6.3 times. Notably, through the exploration process, the BO algorithm rapidly focuses on copper-based catalysts supported on zirconia doped with Zinc and/or Cerium, with the best catalysts, according to the model, showing an optimized composition of 1.85wt% Cu, 0.69wt% Zn, and 0.05wt% Ce supported on ZrO2. When changing the objective, i.e. removing the metal cost as a constrain, the BO algorithm suggests compositions centered on Indium-based catalysts, highlighting an alternative family of catalysts, testifying of the algorithm adaptability and the reusability of the data when targeting different objectives. In only 30 days, the BO, coupled with automated synthesis and high-throughput testing, has been able to replicate the major development stages in the field of heterogeneous catalysts research for CO2 conversion to methanol, made over of the last 100 years with a conventional experimental approach. This data-driven approach proves to be very efficient in exploring and optimizing catalyst composition from the vast multi-parameter space towards multiple performance objectives simultaneously and could be easily extrapolated to different parameter spaces, objectives, and be transposed to other applications.

Chemspyd: An Open-Source Python Interface for Chemspeed Robotic Chemistry and Materials Platforms

“…Arguably, the most prominent such systems have come from companies such as Chemspeed Technologies and Unchained Labs, and have shown the enormous potential to enable highly complex discovery workflows across various fields in chemistry and materials science. Examples include the discovery of battery electrolytes, 13 new catalysts, [14][15][16] organic laser materials, 17,18 polymer formulations, [19][20][21] or stereoselective synthesis. 8 The current phase in the evolution of automated laboratories involves the transition from static, predefined automation workflows to modular and flexible labs where decisions about the next experimental steps are adaptively made in real time (Fig.…”

Cost-Informed Bayesian Reaction Optimization

“…BO-driven reaction optimization has seen significant success in the last few years, especially in the automated laboratory and high-throughput experimentation (HTE) setting. 3,[9][10][11][12][13][14][15][16] Therein, all necessary materials (i.e., substrates, catalysts, additives, solvents) to be considered are typically procured prior to experimentation, and BO is used to find the best reagents and reaction conditions. 1,3,17,18 Yet, the implementation of BO and other ML frameworks in traditional laboratories is still limited.…”

“…To overcome this limitation, here we introduce cost-informed Bayesian optimization (CIBO), a BO framework that incorporates cost into the decision-making process for practical and rational batch experimentation planning. Note that we do not aim to optimize the overall cost of individual reactions nor to include constraints, as pursued in other works, 16,18,23,24 but rather expedite reaction optimization by proposing experiments that are as informative and promising, yet as cost-effective, as possible. CIBO's acquisition function prioritizes experiments with a high benefit-to-cost ratio, enabling minimization of both the number of experiments to carry out and their total cost (Figure 1).…”