Title: Physics-Informed Deep learning for Uncertainty Quantification in Localized Climate Projections

Speaker: Björn Lütjens, HSL Graduate Student

Climate models forecast a global warming of 1 to 5 °C until 2100 (CMIP6 model ensemble, 2019). Local decision makers in e.g., risk management or  policy, have started to take actions, but climate change can vary strongly from the expected 1-5° global change on a local level. An ensemble of high-resolution climate model projections could quantify the magnitude and uncertainty of localized climate projections. High resolution physics-based climate models, however, use up to 600MWhr per simulated year and are computationally too expensive to accurately quantify the uncertainty in localized climate projections (Fuhrer et al, 2018). Recent works propose deep neural networks (DNNs) to learn 30-15k-times faster copies or surrogates, of climate models, but DNNs are known to extrapolate poorly, rendering purely DNN-based climate projections untrustworthy. This thesis presentation will combine physical sciences with DNNs, coined physics-informed neural networks (PINNs), to lay the foundations for a computationally lightweight ensemble of climate models and enable localized climate risk management with quantified uncertainties.