High-dimensional predictions for impact-based risk analysis of geohazards

Rapid flow-like geohazards pose acute threats to communities and infrastructure, yet physics-based runout simulators remain computationally prohibitive for operational impact-based risk analysis. Even if high-resolution datasets with extensive coverage are available, high computational costs direct the decision makers into using scenario-based assessments, which can significantly miscalculate the expected risk given the highly uncertain nature of such events. This study investigates Gaussian-process (GP) emulators for extremely high-dimensional outputs (exceeding 10³ to 10⁴ spatio-temporal grid points), systematically quantifying the trade-offs introduced by dimensionality reduction (DR). We compare three GP variants—Parallel Partial Gaussian Process (PPGaSP), Batch-independent GP (BiGP), and Multitask GP (MTGP)—and apply an established DR–GP workflow to assess the impact of different DR approaches on emulation accuracy and efficiency. This workflow first compresses spatio-temporal fields into low-dimensional latent representations, then performs GP emulation in latent space, and finally reconstructs predictions with uncertainty quantification in the original grid space. Three benchmark cases, synthetic and real-world problems, are used to validate the framework. Our findings provide actionable guidance for selecting appropriate emulation models in high-dimensional geohazard problems. We also investigate the balance between computational efficiency and prediction fidelity for risk-informed early-warning integration.