BALLAST: Bayesian Active Learning with Look-ahead Amendment for Sea-drifter Trajectories under Spatio-Temporal Vector Fields

We introduce a formal active learning methodology for guiding the placement of Lagrangian observers to infer time-dependent vector fields – a key task in oceanography, marine science, and ocean engineering – using a physics-informed spatio-temporal Gaussian process surrogate model. The majority of existing placement campaigns either follow standard `space-filling’ designs or relatively ad-hoc expert opinions. A key challenge to applying principled active learning in this setting is that Lagrangian observers are continuously advected through the vector field, so they make measurements at different locations and times. It is, therefore, important to consider the likely future trajectories of placed observers to account for the utility of candidate placement locations. To this end, we present BALLAST: Bayesian Active Learning with Look-ahead Amendment for Sea-drifter Trajectories. We observe noticeable benefits of BALLAST-aided sequential observer placement strategies on both synthetic and high-fidelity ocean current models. In addition, we developed a novel GP inference method – the Vanilla SPDE Exchange (VaSE) – to boost the GP posterior sampling efficiency, which is also of independent interest.

💡 Research Summary

This paper addresses the problem of efficiently deploying Lagrangian drifters to infer time‑dependent ocean vector fields, a task central to oceanography, marine science, and engineering. Existing deployment strategies rely on space‑filling designs or expert intuition, which ignore the fact that drifters are advected by the flow and thus collect measurements along trajectories. The authors propose BALLAST (Bayesian Active Learning with Look‑ahead Amendment for Sea‑drifter Trajectories), a principled active‑learning framework that explicitly accounts for the future paths of placed drifters, and introduce a novel Gaussian‑process inference technique called Vanilla‑SPDE Exchange (VaSE) that makes posterior sampling tractable on large spatio‑temporal grids.

Surrogate model. The unknown vector field is modeled with a vector‑output Gaussian process built from a Helmholtz kernel for the spatial component and a Matérn 3/2 kernel for time, yielding a separable spatio‑temporal kernel k_tHelm(s,t; s′,t′)=k_helm(s,s′)·k_time(t,t′). This construction respects the divergence‑free and curl‑free structure of ocean currents while providing analytic expressions for posterior means and covariances.

Why standard active learning fails. Conventional Bayesian active learning selects the next measurement location by maximizing the expected information gain (EIG) based on the posterior predictive distribution at a single candidate point. For Lagrangian observers, this ignores the subsequent observations that will be gathered as the drifter moves, leading to sub‑optimal placements (e.g., near domain boundaries where the drifter quickly exits the region). The paper formalizes this limitation (Proposition 1) and demonstrates empirically that EIG can perform worse than a naïve uniform deployment.

BALLAST acquisition function. BALLAST augments the utility computation with a look‑ahead step. From the current posterior p(f|D_n) it draws J≈20 samples of the full vector field. For each candidate placement s at decision time t_n, each sampled field F^{(j)} is used to simulate a drifter trajectory P_T^{F^{(j)}}(s,t_n) until the terminal time T using a simple ODE integrator (Euler). The simulated trajectory yields a set of future observation locations and times, which are added to the existing data D_n. The expected information gain is then evaluated on the augmented dataset, leading to the acquisition function:

 s* = arg max_s (1/J) ∑_{j=1}^J log det