Facilitating the development of Machine Learning-based Digital Twin applications for extreme weather events

With the increasing availability of higher-resolution weather and environmental data as well as advances in Machine Learning (ML) algorithms, data-driven approaches have emerged over the last few years as innovative and fast-computing solutions for addressing detection and prediction of extreme weather events, like storms and wildfires. 

Designing, training and deploying ML models is not trivial and can result in a time consuming process. An integrated software infrastructure for supporting and automating the different steps of the workflow, from weather/climate data gathering and preparation, ML model configuration and training, to deployment of the trained model for detection and prediction applications is required. In this regard, solutions for tracking training metrics and provenance information are crucial components for reproducibility of the results. Besides the software components, HPC infrastructures for handling distributed training over multiple GPUs are also needed to speed up the process.

In the context of the EU-funded interTwin project we are implementing ML-powered Digital Twin (DT) applications for the analysis of extreme events (i.e., Tropical Cyclones and wildfires). The interTwin project is designing and developing a generic Digital Twin Engine (DTE) for supporting DTs from different scientific domains. The DTE provides a software and computing infrastructure for simplifying the creation and management of complex DT workflows. 

This contribution, in particular, will present how the interTwin DTE is supporting the different workflow stages, from model training to their execution, of ML-based DT applications for the detection and prediction of extreme events.