(2gf) Probabilistic Prediction Model-Based High-Throughput Screening for Discovering Feasible and Effective Catalysts for Dry Reforming of Methane

Research Interests: Using machine learning, computational fluid dynamics, and process simulation for multidisciplinary applications

My research interests are problem-solving using machine learning, CFD, and process simulation. I developed machine learning-based prediction models of distillation column temperature and composite resin properties using machine learning and applied them to the commercial industry. I also optimized the sootblower inside the recovery boiler and water mist system through 3D modeling and case studies using computational fluid dynamics. Recently, I conducted to develop a novel high-throughput screening method for discovering promising catalysts using machine learning.

The world is currently facing critical energy and environmental challenges, such as increased energy consumption due to population growth and industrialization [1,2], global warming caused by greenhouse gas emissions [3–5], limited fossil fuel resources [6,7], and energy supply issues due to geopolitical conflicts. To address these problems, eco-friendly energy solutions are urgently needed. One promising approach is the methane reforming process, which converts methane, a potent greenhouse gas, into hydrogen, a next-generation energy source. Among the various methane reforming technologies, dry reforming of methane (DRM) has emerged as a particularly attractive option because it converts both methane and carbon dioxide, which are the main components of greenhouse gases, into useful syngas [8].

However, the development of DRM catalysts remains challenging as most of the highly active catalysts discovered to date are based on noble metals, which are expensive and scarce. Moreover, non-noble metals such as cobalt and nickel-based catalysts tend to deactivate rapidly due to carbon formation or sintering, making them unsuitable for industrial use [9]. Traditional catalyst discovery methods, which rely on trial-and-error experiments, are time-consuming and costly [10]. To accelerate the discovery of effective DRM catalysts, we propose a new methodology called probabilistic prediction machine learning-based high-throughput screening (PP-HTS). The approach uses a probabilistic prediction model trained on a dataset of published DRM experimental results to provide predicted values for catalyst performance and the probability of the predicted results. The probability values represent the degree of confidence in the model's predictions and can be used to assess whether the results are trustworthy.

We validated the model performance and analyzed the impact of variables related to catalyst performance using feature importance and SHAP value. Our methodology provides a reliable and efficient way to discover promising DRM catalysts and can be extended to other catalytic reactions. The successful application of this methodology could help overcome obstacles to the industrialization of DRM and advance the development of sustainable energy solutions.