SVGD-HMC Hybrid Inference for Microkinetic Modeling of Novel NOx Adsorption Process

System identification remains a bottleneck in digital modeling of emerging chemical processes and catalytic materials, where limited observability, high model complexity and incomplete mechanistic knowledge hinder predictive modeling. In this work, hybrid Bayesian inference framework combining Stein Variational Gradient Descent (SVGD) and Hamiltonian Monte Carlo (HMC) was presented for identifying NOx adsorption microkinetics model under high-dimensional parameter space. The method begins with SVGD to efficiently explore the posterior via particle-based, repulsion-driven dynamics. To address search stagnation caused by model stiffness, a particle reassignment strategy is introduced. Additionally, Root Mean Square Propagation (RMSProp)-based stepping is employed to improve convergence robustness. Consequentially, this hybrid approach leverages the global search efficiency of SVGD while incorporating HMC to enhance local posterior resolution in regions where searching stagnate. The presented approach was applied to model a novel NOx adsorption system with Pd/Beta-zeolite being adsorbent, as well as a published Pd/SSZ13-zeolite NOx capture model as benchmark. The proposed SVGD-HMC pipeline offers a scalable and interpretable inference strategy for data-constrained models, enabling richer uncertainty-aware predictions. These results highlight the need for integrated digital-experimental workflows where robust Bayesian methods complement limited experimental information, particularly for next-generation materials and non-steady-state reactor systems.