Venerdì 28 novembre – ore 11:00 CET
Aula Grandori (Edificio 4)
Stanza virtuale: https://politecnicomilano.webex.com/meet/alberto.corigliano

Il seminario intitolato “Generative strategies to empower physics-based wave propagation with deep learning. Applications to earthquake engineering and non-destructive testing.” sarà tenuto dal Dr. Filippo Gatti, Maître de Conférences presso CentraleSupélec.

Abstract

In this study, we present a quantitative evaluation of how elastodynamics simulations can substantially benefit from deep learning–based generative methods integrated with traditional numerical approaches. Two primary frameworks are explored: conditional generative models and neural operators. In the first, a diffusion model is trained within a time-series super-resolution framework. Its objective is to enhance the outputs of 3D elastic wave numerical simulations—accurate up to 5 Hz in large-scale applications [1,2]—by extrapolating them to higher frequencies (5–30 Hz) through learning the low-to-high frequency mapping from globally recorded seismograms [3,4]. The generation process is conditioned on synthetic time histories from numerical simulations, allowing rapid inference of new wave propagation solutions [4].
The second framework demonstrates the effective application of neural operators as full replacements for computationally expensive 3D elastic wave propagation simulations in heterogeneous and anisotropic (polycrystalline) media [5,6]. This approach paves the way toward real-time, large-scale digital twins for complex elastodynamics problems [7,8], including earthquake early warning and ultrasonic non-destructive testing in additive manufacturing.
Finally, we present the results of two hybrid strategies that combine diffusion models and neural operators. In these schemes, zero-shot predictions from neural operators either guide the conditioning and fine-tuning of the super-resolution diffusion model described above or enable transfer learning using a pre-trained GenCFD diffusion model designed for generating solutions to the 3D Navier–Stokes equations [9].

Speaker’s bio

Filippo Gatti is Maître de Conférences at CentraleSupélec since 2019, affiliated to Laboratoire de Mécanique des Sols, Structures et Matériaux (MSSMat) until 2021 and to Laboratoire de Mécanique Paris-Saclay (LMPS) since 2022.
He holds a PhD in Civil Engineering from Université Paris-Saclay and Politecnico di Milano (2017), as well as a MSc (2014) and BEng (2011) from Politecnico di Milano. He has been JSPS post-doctoral fellow at the Disaster Prevention Research Institute at Kyoto University (2018) and visiting researcher at the Earthquake Research Institute at The University of Tokyo (2021).

Research appointments
Since 2022, Filippo Gatti is in charge of Research Operation Jumeaux Hybrides: Simulation, Apprentissage within the LMPS team OMEIR, as well as LMPS representative for research valorization strategies.

Research Activities
Filippo Gatti’s research interests cover the physics-based simulation of wave propagation phenomena, focusing on earthquake engineering and ultrasound wave propagation for non-destructive testing. Since 2014, Filippo Gatti co-develops the high-performance software SEM3D and maintains its open-source release since 2023. SEM3D is CPU/GPU parallel implementation of the spectral element method applied to 3D elastoacoustic wave propagation in highly heterogeneous and non-linear media. SEM3D has been widely employed for blind earthquake simulation and seismic risk assessment of critical infrastructures, such as Kashiwazaki-Kariwa nuclear power plant (Japan) and the Cadarache ITER facilities. In this context, in 2023, Dr. Gatti’s spin-off project, EASYRISK, has been selected as recipient of the POC’UP spin-off program by the SATT Paris-Saclay, aiming at enabling non-expert users with the realism of SEM3D, towards achieving physics-based seismic risk assessment.
Since 2018, Dr. Gatti research activities encompass the use of advanced machine learning to improve the fidelity of the numerical solutions of complex wave propagation problems, by blending physics-based approaches with complex features learned from data.