Catalytic hydrogen combustion microreactors involve strong coupling among surface kinetics, flow organization, heat transfer, and hydraulic loss, making data-efficient and interpretable design screening challenging when only sparse high-fidelity simulations are available. This study proposes a physics-guided machine learning (PGML) framework for small-sample performance prediction and multi-objective design screening of catalytic hydrogen combustion microreactors. The framework uses a low-fidelity kinetic calculation as a physics anchor and learns the discrepancy between the kinetic baseline and CFD responses through supervised residual modeling. A 40-sample CFD dataset covering U-bend numbers \(N_U = 0, 1, 2,\) and 5, two inlet configurations, equivalence ratios of 0.40-0.50, and mass-flow rates of \(9.0\times10^{-5}\)-\(1.8\times10^{-4}\) kg s\(^{-1}\) was used to evaluate wall temperature, hydrogen conversion, pressure drop, and heat-transfer power. In sparse-sample temperature prediction, the PGML strategy reduced the average mean absolute error from 27.51 K for direct machine learning to 12.57 K. In the reconstructed design-screening case, the PGML-selected candidate showed a wall-temperature error of 2.85 K against CFD, compared with 29.02 K for the direct-ML selected candidate, while the pressure-drop error decreased from 66.52 Pa to 38.05 Pa. The regime analysis further indicates that \(N_U = 0\)-1 is mainly mechanism-limited, \(N_U = 2\) provides a transition region with balanced thermal enhancement and hydraulic cost, and \(N_U = 5\) gives high conversion but excessive pressure penalty. The results demonstrate that PGML can improve prediction robustness and provide interpretable decision support for microreactor optimization under sparse CFD data.

Catalytic hydrogen combustion,Machine learning techniques,Multi-objective optimization,Microreactor design