GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.2 GNGTS 2024 AI-based emulators for data driven CFD model reconstructon V. Zago 1 , E. Amato 1,2 , 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 Computatonal Fluid Dynamics (CFD) models have become fundamental tools for the study of fuids and the design of their applicatons. Their use becomes crucial when the fuid exhibits complex behaviors, which can be due to fow conditons or the physical propertes of the fuid. Geophysical fows are good examples of these fows. Among these, a fuid with a high intrinsic complexity is Lava, which include non-Newtonian rheology, a behavior strongly dependent on temperature, and the coexistence of the three phases, solid, fuid and gas (Cordonnier et al., 2016). Numerous CFD approaches have been developed to simulate lava fows. The Smoothed Partcle Hydrodynamics (SPH) has proven partcularly suited for this applicaton, thanks to its Lagrangian and mesh-free formulaton (Zago et al., 2017, 2018). However, despite the good level of descripton that these models can provide, one of the main limitatons in practcal applicaton to complex fuids comes from the epistemic uncertainty, which is due to the poor observability of these fows for which detailed studies and measurements are hard or impractcal. A soluton to this lack of informaton can come from CFD models themselves, which can be used to get informaton about the real phenomena. For example, one can create the so-called digital twins of the studied system or perform reverse engineering. This potental becomes even higher if we consider the growing fusion between CFD methods and Artfcial intelligence, which have highly increased the potentals of physical modeling (Kasim et al., 2021, Amato 2023, Amato et al, 2023A). Here we present an illustratve applicaton of a combined CFD-AI model adopted to retrieve informaton about the viscous model of Lava. The model that we use is an AI based emulator of the SPH method (Zago et al., 2023, Amato et al, 2023B), which is applied to a lava fow simulated with SPH, with a known acceleraton feld. The emulator can recreate pressure and viscous interacton forces, to reconstruct the SPH model originally adopted to generate the simulaton, and to generate new simulatons, emulatng the behavior of the unknown SPH model. These results provide a concrete step toward the development of physically accurate models involving a minimal knowledge of the real phenomena.