911½ñÈÕºÚÁÏ

, UCL
Title- Novel climate modelling merging Machine Learning and High Performance Computing
This talk covers new advances in the design, development and deployment of Machine Learning (ML) tools for climate modelling. We first present how using ML can improve convection within a climate model. The climate model’s temperature and humidity profiles are stochastically perturbed every 6 hours at runtime based on Gaussian process emulators trained on data from a high-resolution model with realistic convection. This hybrid approach on a simplified climate model’s output (SPEEDY) improves the precipitation pattern globally, with the largest reductions in the tropics by around 20%. A follow-on project examines the engineering challenges of compiling the Gaussian process deployment into a Fortran-based state-of-the-art climate model (CESM) at runtime with minimal overhead. Running on an NVIDIA GH200 device reduces the overheads from 3x to only 3% in compute time and yields large modelling improvements. We then discuss the first steps towards including coastal wave modelling into a climate model through ML using this hybrid strategy, emulating Smoothed Particle Hydrodynamics modelling of wave breaking for air-sea exchanges representations. We then present ongoing work on the modelling of weather and climate with diffusion models to reflect better uncertainties, with a focus on acceleration strategies of distillation, and impact on downstream risk quantification. Finally we introduce our current investigation on the modelling and a possible early detection of tipping points in the North Atlantic Subpolar Gyre, requiring decadal multiphysics ML emulation to enable the assessment of uncertainties for robust warnings.

911½ñÈÕºÚÁÏ

Title- Transforming covariates to enhance spatio-temporal predictions in climate applications
In multivariate spatio-temporal statistics our starting point is often a linear model like multiple linear regression or vector autoregression. However, sometimes the cross-sectional interactions between variables are somewhat more subtle, and exist only in the tails of the distribution, or in some other nonlinear sense. In this talk we provide three practical case studies where transforming covariates provides improved models and predictions. The first looks at the nonlinear effects of extreme precipitation on deforestation in Nepal, using novel approaches from functional data analysis. The second looks at the nonlinear effects of ocean currents on the abundance of Antarctic krill, using novel approaches from spectral analysis. Finally, the third introduces a new methodology for performing causal discovery in a nonlinear multivariate time series setting, using novel approaches from extreme value analysis.

Schedule