In recent years, Climate Change has been already manifesting widespread effects on the environment; the exacerbation of extreme weather events is raising major concerns in terms of increased intensity, frequency and duration.
Extreme events such as tropical storms and wildfires represent significant natural hazards that can disrupt ecosystems, and cause economic losses and casualties. Phenomena like these often reach their full destructive potential by leveraging other natural processes. An example of this was Katrina, a tropical cyclone that entered the Gulf of Mexico and rapidly intensified into a Category 5 hurricane due to the presence of a warm eddy that was stationed in the gulf.
Similarly to tropical cyclones, fire plays a crucial role in shaping ecosystems, but climate change can alter its long-term carbon-neutral status by worsening the fire-related weather factors and increasing global fire activity. The expansion of fires in evergreen forest regions could weaken their ability to act as carbon sinks, releasing stored carbon into the atmosphere and providing feedback to climate change. Therefore, it is essential to improve understanding and anticipation of fires in the Earth system, particularly by assessing the likelihood of occurrence of large events, as climate change and human activity continue to influence fire regimes globally.
Detecting and predicting extreme events is challenging due to the rare occurrence of these events, and the consequent lack of related historical data. Recent advances in Machine Learning (ML) provide cutting-edge modeling techniques to deal with detection and prediction tasks, offering also cost-effective and fast-computing solutions.
Digital twins, based on data-driven models, for the analysis of extreme and natural events on climate projections, will help understand how climate change is affecting the frequency and location of extreme events, according to different future projection scenarios based on different socioeconomic changes.
A Digital Twin for projecting wildfire danger due to climate change
Background
Challenge
This Digital Twin aims to help understand how climate change is affecting the probability, location and spread of fires (e.g., burned areas). Advances in Machine Learning (ML) are currently looked at as novel modeling solutions for predicting extreme events like wildfires, by using historical observations to extract useful patterns in the data.
Solution
In interTwin, a team of climate and computer scientists from CMCC, the University of Trento, and IPSL is developing a set of use cases related to the analysis of climate projections of extreme events.
The DTs rely on harnessing knowledge from the huge amount of historical climate and environmental data to train deep neural networks (DNN) capable of predicting and detecting extreme events like wildfires and tropical cyclones. Different DNN architectures will be evaluated to address the various use cases, including Convolutional and Graph networks. The trained models will be used as a base for understanding how different climate scenarios can affect extreme events frequency, locations, probability, etc. Future projection data from large experiments, like the Climate Model Intercomparison Project 6th phase (CMIP6), will be used as input for the trained models.
A set of thematic modules implemented within the project will provide the backbone of the DTs and allow developers to set up additional DTs targeting different applications. On the other side, end-users (e.g., domain scientists and policy makers) will be able to run the DTs for the analysis of extreme events under different climate scenarios. The results will be produced as NetCDF files or as customizable visualizations through easy-to-use solutions like dashboards or Jupyter Notebooks.
The DTs for climate change future projections of extreme events will provide end-users with novel analysis solutions exploiting cutting-edge data-driven approaches. The InterTwin project will support the execution of complex workflows for data management, pre/post-processing and ML model training/inference as well as increase automation of the overall process from data gathering to result production.
Donatello Elia, (Scientist, CMCC)