A Digital Twin Application: Climate Extremes Detection and Characterization using Deep Learning

A Digital Twin Application: Climate Extremes Detection and Characterization using Deep Learning Christian Pagé & Anne Durif (CERFACS (CECI), France) [email protected] https://linkedin.com/in/pagechristian https://www.researchgate.net/profile/Christian_Page CECI, Université de Toulouse, CNRS, CERFACS, Toulouse, France Download Poster! I Impacts of Climate Change Poster in PDF: https://rebrand.ly/xtclim-extremes-clivar The interTwin project is funded by the European Union - Grant Agreement Number 101058386 II interTwin Digital Twin Engine (DTE) 2020 Hurricane Delta causes damage to Louisiana's Gulf Coast 2021 Germany Erftstadt, southwest of Cologne ➤Urgent needs of impact assessments ➤Characterize changes of climate extremes ➤Multiple domains: infrastructures, urban, agriculture, transportation, etc. ➤Compound Events ➤Uncertainty Quantification: ensemble approach interTwin Project: https://www.intertwin.eu xtclim software: https://github.com/cerfacs-globc/xtclim icclim software: https://github.com/cerfacs-globc/icclim ➤interTwin DTE conceptual model ➤Open-source integrated platform, based on ➤Open standards ➤APIs, Protocols ➤Offers the capability to integrate with applicationspecific Digital Twins (DTs). Blueprint Architecture ➤Possible huge gain of performance ➤Efficient parallel execution ➤Use of GPU architectures ➤More generic approach ➤Novel techniques in climate data analysis Take Home Messages 1.Generic and unsupervised anomaly detection and characterization 2.Coherent results 3.Handles high amounts of data 4.First CVAE for climate projection analyses III Using Machine Learning to Detect and Characterize Climate Extremes V Perspectives ➤Robustness with more members ➤Validation with icclim ➤n-day input (“video”) ➤further integration with interTwin architecture Data ➤Climate variables ➤1950-2100 (historical and projected) ➤~125*125km grid ➤SSP: 1-2.6, 2-4.5, 3-7.0, 5-8.5 ➤Daily max temperature, precip, wind ➤Climate model: CMCC-ESM2 ➤NetCDF files preprocessed to ndarrays ➤32*32 square of Western Europe ➤Season split ➤Normalization ➤Climate indices (icclim) for validation Method: Variational Auto Encoder Method: High-Level Workflow Summer Results Generic detection algorithm ○Intense rainfall ○Drought ○Heatwave ○Cold spell ○High wind Characterization ○Frequency of occurrence ○Spatial extent ○Intensity (if relevant) ○Duration IV Why use AI? Yearly Results Scenario Test data SSP1 -2.6 SSP2 -4.5 SSP3 -7.0 SSP5 -8.5 Unusual days 12 391 749 1016 1610 Proportion of unusual days 100,33% 494,19% 946,66% 1284,13% 2034,88% Maximum spike 0,0040 0,0042 0,0067 0,0064 0,0069 Average maximum 0,0030 0,0031 0,0034 0,0035 0,0036 Maximum duration (days) 5 26 43 54 55 Average duration (days) 23,833333 6,040323 6,271605 8,00995 Number of spikes 6 102 124 162 201