Speaker: Dr. Maurizio Carbone
Title: Backtracking turbulent scalar fields: comparison with classic heuristics
Abstract:
We evaluate a backtracking-based approach for odor source localization, which relies on learning the odor particle propagator via a normalizing flow and sampling it through an associated Fokker-Planck equation. The resulting model is trained on a passive scalar field emitted from a point source and advected by a two-dimensional turbulent flow in the inverse cascade regime. We show that the agent can effectively backtrack odor detections, reconstructing trajectories that resemble passive scalar plumes, thus efficiently localizing the source. We explore how the mean arrival time and success rate depend on characteristic parameters, such as the mean wind speed relative to the agent speed and turbulence intensity. Finally, we compare the arrival times with well-established heuristics, including infotaxis and castingand-surging strategies.
Poster
