Hazardous phenomena such as landslides, debris flows, lava flows, tsunamis, earthquakes or floods, are pervasive in the Earth system and cause severe losses of life and property worldwide, every year. These phenomena tend to be extremely complex and may occur in relative isolation but also compounded with other phenomena, thus posing significant challenges for single- and multi-hazard assessments. If the interaction between hazard and the human environment is considered (e.g. water pumping and the availability of fresh water), the hazard quantification becomes even more complicated.
Generally, the data needed to quantify hazard are scarce (particularly for large, infrequent events) and as such purely data-driven methods have limited application. Hence, a comprehensive mapping of unobserved, yet likely or possible, events to properly quantify the hazard commonly relies on the use of computer models, or simulators. These simulators tend to be sophisticated, to capture the underlying physics, and, consequently, computationally-demanding. Thus, strategies to reduce the total number of simulations required for different purposes, like uncertainty quantification, model calibration, optimisation and/or forecasting, are constantly sought.
One popular strategy is the use of ‘surrogate models’ (e.g. statistical emulators, machine-learning techniques, etc.), which can be defined as computationally-cheaper, statistical models aimed at reproducing the behaviour of the simulator as closely as possible, so they can be used as a (fast) substitute of the latter. This turns, for instance, uncertainty quantification for probabilistic hazard assessment into a computationally-tractable problem.
In this session, we would like to collectively explore the recent application of surrogate models in quantitative single- and multi-hazard assessments, to better understand common and unique issues arising from different hazardous phenomena and/or local-to-regional contexts (e.g. diverse topographic and/or bathymetric configurations). Another key goal is discussing future developments in the area, for example, challenges in modelling systems with high-dimensional input-output spaces, with a vision of comprehensive hazard assessments contributing to effective and efficient risk management of these phenomena.
Generally, the data needed to quantify hazard are scarce (particularly for large, infrequent events) and as such purely data-driven methods have limited application. Hence, a comprehensive mapping of unobserved, yet likely or possible, events to properly quantify the hazard commonly relies on the use of computer models, or simulators. These simulators tend to be sophisticated, to capture the underlying physics, and, consequently, computationally-demanding. Thus, strategies to reduce the total number of simulations required for different purposes, like uncertainty quantification, model calibration, optimisation and/or forecasting, are constantly sought.
One popular strategy is the use of ‘surrogate models’ (e.g. statistical emulators, machine-learning techniques, etc.), which can be defined as computationally-cheaper, statistical models aimed at reproducing the behaviour of the simulator as closely as possible, so they can be used as a (fast) substitute of the latter. This turns, for instance, uncertainty quantification for probabilistic hazard assessment into a computationally-tractable problem.
In this session, we would like to collectively explore the recent application of surrogate models in quantitative single- and multi-hazard assessments, to better understand common and unique issues arising from different hazardous phenomena and/or local-to-regional contexts (e.g. diverse topographic and/or bathymetric configurations). Another key goal is discussing future developments in the area, for example, challenges in modelling systems with high-dimensional input-output spaces, with a vision of comprehensive hazard assessments contributing to effective and efficient risk management of these phenomena.