Multivariate design storms using vine copulas and Gaussian process regression

Multivariate design storms represent events associated with a defined Joint Return Period (JRP) while preserving observed spatial dependence. They enable the synthetic generation of storms that have not yet been observed but are plausibly likely, thereby expanding the scenario space for design and risk assessment. Two practical challenges persist: (i) embedding JRPs within storm-generation procedures beyond three dimensions, which is computationally demanding; and (ii) representing spatial intermittency, which has so far been treated only with a bivariate Multivariate Mixture Model (MMM). We present a method that addresses both challenges concurrently. We compare five copula-based approaches for high-dimensional dependence, employ Gaussian process regression (GPR) to accelerate JRP estimation, and extend the MMM to represent intermittency in multiple dimensions. The framework is applied to the Besaya River basin (northern Spain). Results show that GPR reduces computational cost by an order of magnitude while maintaining approximation errors on the order of 10 − 3 , and that the MMM captures intermittency in multidimensional settings. Flexible vine copula–based approaches best preserve spatial and multidimensional dependence in multivariate storm generation, whereas Gaussian-based and univariate benchmarks systematically underestimate storm totals. • Multivariate design storms preserve spatial dependence to explore plausible unobserved extremes for a given return period. • Gaussian process regression cuts joint return period computation by 10 × in high dimensions, with approximation errors near 10 − 3 . • Multivariate Mixture Models capture spatial intermittency patterns in multidimensional precipitation scenarios. • Flexible vine copulas outperform Gaussian methods in preserving spatial dependence and estimating compound storm magnitudes.