Surrogate-Based Optimization of an Adaptively Refined-Mesh Multiphase Granular Flow Model for Landslide-Induced Waves

Subaerial landslides can generate destructive tsunami waves that threaten coastal and reservoir environments. High-fidelity multiphase CFD models provide an accurate representation of these events but demand considerable computational resources, especially when simulating granular flows and complex wave interactions. Adaptive mesh refinement (AMR) has shown promise in improving simulation efficiency; however, its effectiveness is highly sensitive to parameter settings that are often tuned heuristically. This study introduces a surrogate-based sequential approximation optimization (SAO) framework using Radial Basis Function (RBF) models to systematically optimize AMR parameters within a multiphase granular flow solver. The method is applied to two benchmark cases representing small- and large-scale laboratory scenarios. In the small-scale case, SAO achieves a 56% reduction in time with negligible loss in accuracy. The optimized configuration is successfully transferred to the large-scale case, yielding a 30% reduction, and is further validated through independent optimization. The framework demonstrates stable convergence across different sampling sets and remains sample-efficient under limited simulation samples. A detailed analysis highlights the influence of mesh dynamics, domain complexity, and early-stage refinement behavior on AMR performance. The results confirm that surrogate-assisted AMR optimization offers a scalable and efficient alternative to traditional sensitivity analysis. The proposed framework offers a practical approach for improving the efficiency of CFD-based landslide-induced tsunami simulations and shows potential for future extension to large-scale, more realistic topographies and early-stage reconstructing studies