Work

Physics-informed offline reinforcement learning reduces worst-case fuel risk in maritime routing

Year: 2026
Type: article
Abstract: International shipping produces approximately 3% of global greenhouse gas emissions, yet voyage routing remains dominated by heuristic methods that adapt poorly to varying ocean conditions. Here we present Pier (Physics-Informed, Energy-efficient, Risk-aware routing), an offline reinforcement learning framework that learns fuel-efficient, safety-aware routing policies from physics-calibrated environments grounded in historical vessel-tracking and ocean-reanalysis data, requiring no online simulator. Validated on one full year (2023) of Automatic Identification System data across seven Gulf of Mexico routes, Pier reduces mean fuel consumption and, more strongly, its variance and worst-case tail. Across five routes, mean per-voyage CO 2 is 10.2% lower than great-circle routing (bootstrap 95% CI [2.9%, 16.7%]; Monte Carlo [9.3%, 11.0%]). Per-voyage CO 2 variance is 3.7 \(\times\) lower (Levene’s test \(p = 0.0020\) ), and voyages exceeding 2 \(\times\) the median CO 2 fall from 3.0% to 0.3% (2 of 577), with the worst single voyage reduced by 65%. Pier is forecast-independent: in an exploratory experiment, classical path optimization’s wave protection degrades as forecast noise grows ( \(\sim\) 2.2 \(\times\) at \(\sigma = 0.25\) ), while Pier maintains constant performance from local observations alone. Pier demonstrates that physics-informed offline reinforcement learning enables stable, risk-aware decision-making in safety-critical transportation domains where online simulation is unavailable.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7138957057
FWCI: 1.27
Cites: 37
Cited by: 1
Related to: 10