(329f) Data-Driven Discrepancy Quantification in Microkinetic Modeling of Catalytic Processes Via Bayesian Calibration

Computational modeling has become a powerful tool in catalyst design and reaction engineering, facilitating a shift from empirical trial-and-error approaches to predictive mechanistically grounded ones that guide experimentation and accelerate the discovery of efficient catalytic systems [1-3]. First-principles-based models—particularly those based on density functional theory (DFT) and transition state theory (TST)—are instrumental in elucidating elementary reaction mechanisms and estimating kinetic parameters at the atomic level. However, these models often diverge from experimental observations due to idealized representations of surface structure and thermodynamic conditions, oversimplified microkinetic schemes, and insufficient treatment of the multiscale complexity and inherent uncertainty in catalytic systems. These models offer mechanistic insight into adsorption, surface reactions, and activation barriers, essential for constructing microkinetic models of heterogeneous catalytic processes [4,5].

Bridging the gap between theoretical predictions and experimental data requires hybrid modeling frameworks that integrate mechanistic understanding with empirical observations. This work introduces a Bayesian discrepancy correction framework that couples physics-based models with experimental data to improve predictive accuracy and model reliability. Unlike conventional methods that neglect measurement uncertainty, the proposed approach incorporates independent probabilistic error terms for both model predictions and experimental measurements. Initial implementations that assumed negligible experimental noise were found to have limited generalizability when Gaussian process (GP) regression was applied to correct model discrepancies. By relaxing this assumption and employing noise-aware GP modeling, the revised framework captures model-inform errors more effectively across a wide range of operating conditions. This probabilistic calibration approach enhances both the interpretability and robustness of kinetic models, facilitating data-driven model refinement and systematic uncertainty propagation—key capabilities for model-based design, optimization, and control. Although the framework incurs additional computational costs due to probabilistic inference and hyperparameter tuning, these are offset by the substantial improvements in model fidelity and predictive performance [6,7].

The methodology is demonstrated using dry reforming of methane (DRM) as a representative case study. DRM is a catalytically driven process that converts CH₄ and CO₂ into synthesis gas (H₂ and CO), offering a promising pathway for CO₂ utilization. However, practical implementation is limited by challenges such as carbon deposition and catalyst deactivation. A mean-field microkinetic model is developed based on DFT-derived energetics for Ni surfaces, encompassing key elementary steps including CH₄ and CO₂ adsorption, sequential C–H bond dissociation, hydrogen spillover, and metal-support interactions. The influence of support materials—specifically ceria (CeO₂) and alumina (γ-Al₂O₃)—is incorporated by modeling CO₂ activation pathways involving Lewis acid-base interactions and oxygen vacancy sites [8-10]. Recent experimental findings have demonstrated that Ni catalysts supported on ceria outperform those on alumina, particularly under high space velocities and elevated temperatures. These improvements are attributed to ceria’s capacity to enhance CO₂ activation, inhibit side reactions such as the reverse water-gas shift and methanation, and stabilize Ni particles through strong metal–support interactions [11]. The case study illustrates how integrating mechanistic models with data-driven corrections can reconcile DFT-informed microkinetic models with experimental behavior, enabling more accurate predictions and informing rational catalyst design under realistic conditions.

References:

[1] Nørskov, J.K., Bligaard, T., Rossmeisl, J., Christensen, C.H. (2011). Towards the computational design of solid catalysts. Nature Chemistry, 1(1), 37–46.

[2] Butler, K.T., Davies, D.W., Cartwright, H., Isayev, O., Walsh, A. (2018). Machine learning for molecular and materials science. Nature, 559(7715), 547–555.

[3] Hongliang X., “Catalyst design with machine learning.” Nature Energy 7.9 (2022): 790-791.

[4] Chorkendorff, J. W. Niemantsverdriet, Concepts of modern catalysis and kinetics, John Wiley & Sons, 2017.

[5] Motagamwala A.H., Dumesic J.A., Microkinetic modeling: A tool for rational catalyst design, Chemical Reviews 121 (2) (2020) 1049–1076.

[6] Kennedy M. C., & O’Hagan A. (2001). Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63(3), 425-464.

[7] Brynjarsdóttir J., & O’Hagan A. (2014). Learning about physical parameters: The importance of model discrepancy. Inverse Problems, 30(11), 114007.

[8] Zhu Y.A., Chen D., Zhou X.G., Yuan W.K., DFT studies of dry reforming of methane on Ni catalyst, Catalysis Today 148 (3-4) (2009) 260–267.

[9] Fan C., Zhu Y.A., Yang M.L., Sui Z.J., Zhou X.G., Chen D., Density functional theory-assisted microkinetic analysis of methane dry reforming on Ni catalyst, Industrial & Engineering Chemistry Research 54 (22) (2015) 5901–5913.

[10] Liu S., Zhou Z., Chen J., Fu Y., Cai C., Adsorption and decomposition of CO₂ on γ-Al₂O₃ (1 0 0): First-principles investigation, Applied Surface Science 611 (2023) 155645

[11] Ighalo J.O., et al. “Dry reforming of methane at high space velocities on CeO₂-supported Ni catalysts.” Chemical Engineering Journal 508 (2025): 160707.