GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.2 GNGTS 2024 An AI-based emulator to enhance SPH lava fows simulatons E. Amato 1,2 , V. Zago 1 , C. Del Negro 1 1 Isttuto Nazionale di Geofsica e Vulcanologia, Osservatorio Etneo, Catania, Italy 2 Department of Mathematcs and Computer Science, University of Palermo, Palermo, Italy Lava fows are complex fuids, exhibitng non-Newtonian rheology with temperature-dependent viscosity and phase transiton, capable of overcoming barriers and forming channels and tunnels (Cordonnier et al., 2016). While lava fows are generally not hazardous to nearby populatons due to their slow velocites, their passage through towns can cause complete destructon. Therefore, reliable predictons of the areas likely to be inundated by lava fows are of obvious interest to hazard managers during a volcanic erupton (Del Negro et al., 2020). The main factors that govern lava-fow length include the discharge rate of lava at its vent, the lava compositon, erupton temperature, cooling rate and the ground topography over which the lava fows. As a result, numerical simulatons that consider the key factors infuencing the extent of lava fow propagaton are crucial for forecastng efusive scenarios (Del Negro et al., 2016). Due to the complex physics of volcanic phenomena and the unique characteristcs of lava, mathematcal models can assist in simulatng the evoluton of lava, providing accurate predictons of the spato-temporal dynamics of the fuid. These kinds of simulatons consttute a challenge for Computatonal Fluid Dynamics (CFD) (Anderson and Wendt, 1992). Smoothed Partcle Hydrodynamics (SPH) (Monaghan, 2005) is a potent CFD method partcularly suited for simulatng lava fows (Zago et al., 2017, 2018). It is a Lagrangian mesh-free numerical method based on a discrete approximaton of the Navier-Stokes equatons. It is a partcle-based method, where the fuid is discretzed using partcles, and it has the capability to handle specifc fuid details, such as viscous and thermal efects. In additon, it is parallelizable and executable on Graphics Processing Units (GPU), thereby acceleratng simulatons. However, SPH simulatons stll require extended run tmes and substantal computatonal resources, typically taking weeks to obtain a few minutes of simulaton. This poses challenges for achieving real-tme applicatons, especially in the context of volcanic hazard monitoring. While speed-ups of the simulatons can be achieved by simplifying the model or increasing computatonal resources, a simplifed version of the model may not capture the complex physical dynamics present in real world applicatons. These limitatons can be addressed by introducing the use of Artfcial Intelligence (AI) (Goodfellow et al. 2016, Bonaccorso, 2017) to reduce the required computaton. The combinaton of CFD and AI allows for the enhancement of fuid modeling performance and the extension of functonalites (Bortnik and Camporeale, 2021). AI algorithms can be trained on SPH simulated data to rapidly learn the behavior of the CFD reference model. Models of this nature are referred to as emulators