Leveraging machine learning engineering to uncover insights into heterogeneous catalyst design for oxidative coupling of methane

Abstract

Machine learning (ML)-assisted catalyst investigations for oxidative coupling of methane (OCM) are assessed using published datasets that include literature data reported by different research teams, along with systematic high-throughput screening (HTS) data. Support vector regression (SVR) is performed on the selected 2842 data points. The first SVR leads to eight catalysts with C₂ yields higher than 15.0% under the current reaction conditions, but the second attempt with the updated dataset including the first validation results does not improve the prediction because of spatial shrinkage. The Bayesian optimization processes also start with datasets of 3335 data points, and are considered for three cycles using the updated dataset. Repeating the Bayesian processes certainly improves the C₂ yields observed in the validation results, but the convergence of the elements presents another issue. Accordingly, data-driven catalyst investigations involve a different set of defect issues from the conventional style of catalyst investigations. The unveiling of issues in the highly active OCM catalyst investigation by ML engineering conducted for this study is intended to clarify future challenging subjects for ML-assisted research innovations. Actions to proactively discover the encounters with serendipity to broaden the scope of the material survey area using ML approaches and/or working with the researcher's intuition can increase the possibility of fortuitous discoveries and the achievement of desired outcomes.