Machine Learning Approaches to the Exploration of Ammonia Synthesis Catalysts Utilizing Experimental Data
Abstract
Traditional heterogeneous catalyst discovery process relies upon human expert-guided experiments based on domain knowledge and human intuition supported by catalyst characterization. However, this process is time and resource-consuming and suffers from human bias. Machine learning (ML) models trained on experimental data have the potential to significantly accelerate this discovery process by guiding experiments to optimize the catalyst performance and increase the understanding of the catalysts. Ammonia synthesis catalysts are a class of catalysts widely investigated in the literature, which has generated decades of experimental data. Moreover, ammonia has been gaining attention since it can be used as a carbon-free next-generation fuel and a storage medium for hydrogen generated via renewable sources such as wind and solar. Due to the high geological dispersion, intermittency, and variability of these renewable sources, modular ammonia synthesis under milder conditions (1-3 MPa, 573-673 K) compared to the current Haber-Bosch process (10-25 MPa, 723-873 K) is required. The current iron (Fe) catalysts are unsuitable for modular ammonia synthesis due to their low activity and longer activation time. Ruthenium (Ru) catalysts on metal oxide supports have risen as a promising alternative due to their high activity, ease of synthesis, and quick activation. However, novel catalysts with high activities and low Ru weight loadings are still required due to the high cost of Ru compared to Fe. This work focuses on machine learning approaches to the exploration of these Ru catalysts. Experimental data for the catalyst was mined from the literature, cleaned, and featurized. The ML model trained on these data guided the lab experiments for catalyst synthesis and testing via active learning. Consequently, it optimized the ammonia synthesis rates of the catalyst search space designed via domain knowledge. The Ru(1 wt.%), Ba (2 wt.%), Cs (2 wt.%) supported on Pr2O3 had the highest activity in the search space, which was able to exceed the activity of most of the state-of-the-art catalysts for thermocatalytic ammonia synthesis in the literature. The number of experiments required to reach the best catalyst in the search space was reduced by 50 percent, even when no information related to the best promoter (Ba) was in the literature dataset initially in the ML model. The analysis of the experimental pathways suggested by the ML model suggested that the best catalysts in the search space have promoters with lower first ionization energies and higher atomic radii. A separate ML was developed to understand the effects of metal oxide supports on the ammonia synthesis rate. The Shapley additive explanations (SHAP) analysis of the model revealed that the formation energies of metal nitrides and metal hydrides formed on the metal oxide surface have a volcano-type relationship with the ammonia synthesis rate. The relationship with the metal nitride formation energy was verified via ammonia temperature programmed desorption (TPD) experiments, and the one with hydride formation energy was found to be complex according to the hydrogen-TPD experiments.