Direct Design of Catalysts in Oxidative Coupling of Methane via High‐Throughput Experiment and Deep Learning

Abstract

The combination of deep learning and high‐throughput experiments is proposed for the direct design of heterogeneous catalysts in the oxidative coupling of methane (OCM) reaction. Deep learning predicts 20 active catalysts from high‐throughput 12,708 OCM experimental data where 19 of the predicted 20 catalysts have not been previously reported. The predicted 20 catalysts are then evaluated through high‐throughput experiments where a highly active unreported catalyst Ti−Na2WO4/TiO2 is discovered. Ti−Na2WO4/TiO2 results in a high C2 yield of 18.8 % where the maximum C2 yield is reported to be approximately 20 % within the 12,708 OCM data. Furthermore, the experimental conditions predicted for Ti−Na2WO4/TiO2 are also reproduced by high‐throughput experiment. Thus, deep learning demonstrates that both catalysts and experimental conditions can be simultaneously explored for designing catalysts. More importantly, deep learning assisted catalysts search is found to dramatically increase the chances of finding active catalysts where 9 out of 20 predicted catalysts result in a C2 yield over 15 %. Therefore, the combination of deep learning with high‐throughput experiments is proposed to be an effective strategy for the direct design of catalysts.

A chemistry‐inspired neural network kinetic model for oxidative coupling of methane from high‐throughput data

“…It should be noted that Ti is more expensive than Mn. A similar observation was also made by Nguyen et al 25 and Sugiyama et al 23 The effect of M2 on the methane consumption rate is actually small locally, with most catalysts being within 80% (for M2 = Ba) to 105% (for M2 = Zn) of the reference. The effect on C 2 yield is substantial, with a variation from 60% to essentially 100%.…”

“…Several works have recently built on this idea to formulate classification and regression models of both literature and HTE datasets to identify promising catalyst composition. [16][17][18][19][20][21][22] While such methods are indeed quite promising in identifying better catalyst formulations and even nudging experimentalists to look for a particular material subspace for further examination, 19,23,24 the underlying models are often black box, without taking into consideration the intrinsic chemistry or physics that are already known. We proffer, on the other hand, that machine-learned models should include domain knowledge that ranges from rather simple concepts like mass balance and non-negativity of quantities (such as conversion, yields, and concentration) to more complicated aspects such as thermodynamic constraints at or near equilibrium, known or expected periodic trends identified from computational chemistry, and so on.…”

High-throughput screening and literature data-driven machine learning-assisted investigation of multi-component La2O3-based catalysts for the oxidative coupling of methane

“…Indeed, within the last a few years, the combined use of tailor-made experimental datasets obtained with a high throughput screening (HTS) machine and data analytic techniques such as multi-output ML and network profiling have presented new avenues for the design and elucidation of the nature and catalytic activity of catalysts, in particular OCM. [22][23][24][25][26][27][28][29][30][31][32] Several parallel reactor systems for effective data collection in a fixed bed catalytic reaction have become commercially available 33,34 and others have been proposed in the literature. 22,[35][36][37] Nevertheless, these systems still require special skills and entail high costs for operation and construction similar to the conventional modes of catalyst investigation.…”

“…Indeed, within the last a few years, the combined use of tailor-made experimental datasets obtained with a high throughput screening (HTS) machine and data analytic techniques such as multi-output ML and network profiling have presented new avenues for the design and elucidation of the nature and catalytic activity of catalysts, in particular OCM. 22–32…”

Designing Catalyst Descriptors for Machine Learning in Oxidative Coupling of Methane

“…On the other hand, catalyst compositions are generally expressed as atomic symbols and therefore must be converted to numerical variables for machine learning. In the OCM reaction, catalysts have been converted into atomic numbers as well as binary variables using one hot encoding. − However, descriptors for determining C 2 yield remains uncertain; thus, physical quantities from the periodic table should be explored. Here, catalyst descriptors for C 2 H 4 /C 2 H 6 selectivity in the OCM reaction are performed using machine learning and the information from the periodic table, after which the designed catalysts are then experimentally tested.…”