Revisiting Machine Learning Predictions for Oxidative Coupling of Methane (OCM) based on Literature Data

Abstract

Machine learning (ML) predictions for the oxidative coupling of methane (OCM) are evaluated under experiment situation. The ML protocol has sparked new motivation for trial runs of 96 kinds of metal‐supported catalysts based not only on scientists’ experiences but also on data presented in earlier reports of the literatures and obtained during verification. Our protocol discovers unreported catalyst combinations for OCM reactions from data expanding upon three decades of research, where various numbers of catalysts are predicted and confirmed to perform better than blank data. Nevertheless, the target on C2 yield for the OCM reaction remains as a challenging subject: i. e. higher than 30 %. Revisiting data reported in the literature reveals that different reactor systems and/or specific methods are used in the original data for achieving higher than 30 % C2 yield. Such specialties are attributed to the inadequacy of a literature‐data‐driven ML approach at the present situation. Furthermore, classification of experimental data has indicated target C2 yield values and trends toward CH4 and O2 conversion and product selectivity in high dimensions can improve future ML prediction. These findings are greatly beneficial for the next stage of development to find a global descriptor to improve ML prediction accuracy beyond interpolation filling.

Sulfur oxidative coupling of methane process development and its modeling via machine learning

“…This protocol treats the elemental features of a catalyst as inputs, without the need of directly inputting the catalyst, resulting in high prediction accuracy. Over the last lustrum , the group of Takahashi have published several papers on ML applied to OCM 24,28–31 . For example, they used different ML techniques to predict the effect of the OCM reaction conditions on the C 2 yield, revealing the nonlinearity between experimental conditions and C 2 yield 31 .…”

Representing Catalytic and Processing Space in Methane Oxidation Reaction via Multioutput Machine Learning

“…One can consider that this behavior is an example of process dependency on the OCM reaction, which, in past studies, has been widely attempted to achieve a high C2 value. 19 Given that the concentration of the diluent (in this case, Ar gas) correlates to the reduced CO and CO 2 selectivities and the increases in Although the behaviors of maximum CO and CO 2 selectivity are different between Clusters 1 and 3, Cluster 3 tends to have higher CO and CO 2 selectivities at its best. Meanwhile, Clusters 2 and 4 are found to have higher CO 2 selectivities.…”

“…To represent the complex behaviors of selectivities and CH 4 conversion, multioutput supervises machine learning has been introduced, where multiple target variables are predicted. The existing oxidative coupling of methane (OCM) catalyst data, such as literature data, have a disadvantage in the implementation of machine learning because of their inconsistency, such as careful controls for the preparation of catalysts or experimental conditions to achieve a high C2 value . Thus methane oxidation via a high-throughput (HTP) experiment is chosen as it generates homogeneous and consistent data, which is an important factor for designing accurate machine-learning models. , Here multioutput supervised machine learning is introduced to simultaneously evaluate the selectivities and the CH 4 conversion upon the change in experimental conditions and catalysts combined with HTP experimental data.…”

“…On the contrary, as seen in Figure b,c, the trends for C 2 H 4 and C 2 H 6 selectivity when the Ar flow is 1.5 mL/min are different, where C 2 H 4 is produced in a constant quantity at all temperatures, whereas the production of C 2 H 6 decreases as temperature increases. One can consider that this behavior is an example of process dependency on the OCM reaction, which, in past studies, has been widely attempted to achieve a high C2 value . Given that the concentration of the diluent (in this case, Ar gas) correlates to the reduced CO and CO 2 selectivities and the increases in C 2 H 4 and C 2 H 6 selectivities when temperature is increased, data where the Ar flow is 14.0 mL/min is extracted.…”

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

“…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.…”