Natural hazards, such as earthquakes, volcanoes, tsunamis, hurricanes, and results of climate change are of great scientiﬁc and societal interest. The Symposium focuses on deep learning and inverse problems, including monitoring, enabling the study of the processes associated with natural hazards, forecasting, natural disaster mitigation and early warning, exploiting the latest developments in data acquisition, inference, and modeling. The challenge in modeling accurately the mentioned processes is not only one of computational power. Indeed, one needs to develop new approaches to describe the complexities of these natural systems and to understand the robustness and predictive capabilities of models that operate within complete information on the contributing processes. Often a cascade of events drives their behavior. While supervised learning assumes a ﬁxed data distribution and task to be performed (that is, the same during training and testing), a growing problem of interest is the case of non-stationary environments, whereby both data and the tasks to be performed are no longer assumed ﬁxed. Under such changing conditions, effective learning strategies must therefore be able to extract structural regularities across different data distributions and tasks – possibly thanks to physics-aware machine learning models. Causal inference is quickly becoming a major area of deep learning research, highlighting the need to go beyond passive observational models.
In the context of the Symposium, the interest is in exploring the role of causal inference in improving models for critical geoscience phenomena. Moreover, deep learning algorithms may advance ways to prepare relief efforts ahead of projected disasters. Complementary to early warning, deep-learning-based methodologies for planning processes, with different lead time scales for different types of disasters, could save lives and property.