Graph Neural Networks (GNNs) provide an efficient surrogate for computationally intensive density functional theory calculations in catalytic material discovery. However, they often struggle with prediction reliability and out-of-domain generalization. These limitations necessitate reliable prediction uncertainty quantification for informed catalyst discovery. While Gaussian Processes (GPs) offer principled Bayesian uncertainty quantification, their cubic time complexity, high memory requirements, and inability to learn from graph structures limit their application in high-throughput discovery. We introduce Deep Graph Kernel Learning (DGKL), a scalable framework that combines a GNN backbone with a sparse variational Gaussian Process (SVGP) for uncertainty quantification in adsorption energy prediction. We benchmark DGKL against state-of-the-art methods, including ensemble/query-by-committee, evidential, and Monte-Carlo dropout approaches. DGKL consistently outperforms existing methods across ranking-based metrics (negative log-likelihood, expected normalized calibration error, miscalibration area) and error-based metrics (RMSE vs. RMV and error vs. standard deviation) while maintaining computational efficiency. Specifically, DGKL achieves the lowest expected normalized calibration error (0.06-0.10), lowest miscalibration area (0.04-0.07), and highest Spearman correlation coefficient (0.34-0.51) across diverse datasets and GNN backbone combinations. Qualitatively, DGKL's RMSE vs. RMV plots demonstrate superior calibration compared to competing methods. Additionally, we propose a DGKL variation capable of predicting atomic-level uncertainty - a feature absent in existing methods — offering fine-grained insights into out-of-domain data. DGKL can be incorporated into active learning frameworks to efficiently explore catalytic material space, accelerating the discovery of novel catalysts.
