Knowledge-informed reduced-order modeling for erosion prediction in pipe bends

Abstract

Solid particle erosion in pipeline elbows poses a persistent challenge in the energy and process industries, where accurate yet efficient prediction methods are urgently needed. While computational fluid dynamics–discrete element method (CFD–DEM) simulations provide high-fidelity erosion predictions, their computational demands severely limit practical deployment. To bridge this gap, this study proposes a knowledge-informed reduced-order modeling framework that couples proper orthogonal decomposition (POD) with Kriging interpolation. The surrogate model is enhanced by numerically validated, physically motivated correlations between erosion ratios and key impact parameters, enabling improved extrapolation and interpretability. Validation against full-order CFD–DEM results demonstrates that the enhanced POD–Kriging model accurately reproduces spatial erosion fields while achieving speedups exceeding 2000 × . Compared to the conventional POD-based surrogate, the proposed approach reduces prediction errors by up to 76 %, with local error at high-risk elbow regions reduced to within 4 %. These results highlight the model's robustness and generalizability across both single- and multi-parameter operating conditions. The framework offers a computationally efficient and physically consistent alternative for erosion assessment and design optimization in industrial pipeline systems.