Christian Kehl

Date: 16:00 – 16:30, Thursday, 11.11.2021
Location: MS Teams ICS Colloquium

Title: Uncertainty Quantification in Lagrangian Fluid Models using ML Stochastic Methods

Abstract:

The modeling of oceanic currents is central to the research on climate change and oceanic plastic pollution, among others, which directly relate to UU’s research agenda on the ‘Pathways to Sustainability’. Lagrangian ocean models, which are predominantly deterministic in their approach, are used in this regard to study ocean currents and the transport of particulate matter by those fluidflow patterns. Conversely, when stakeholders and the public engage with, for example, the problem of ocean plastics and its transport, common questions arise about the future distribution of an existing plastic distribution, as well localized confidence intervals of the prediction, to enact policy reactions. In my talk, I show how we approach this question, I will address why the available deterministic modeling approach is insufficient in addressing those questions, and how we can employ, extend and adapt stochastic modeling approaches from within machine learning to provide future physics-related predictions together with global- and local uncertainty approximations. A specific incarnation to our approach is an adaptive convolutional LSTM network with Monte-Carlo dropout (MCdropout) that deducts particle distributions in variable future timesteps, as well as a confidence map and bayesian-based prediction probability metrics as global and local uncertainty estimates. Furthermore, I highlight the relation of our current research insights to our recent ‘AI & Sustainability’ Labs proposal on explainable AI to guide plastic cleanup strategies in the Galapagos archipelagos