Every merchant ship already does a version of this problem each voyage: pick a route and a speed that arrive on time, inside weather limits, burning the least fuel. The question AI research is forcing is whether a learned model does that better than the deterministic optimization stack — Dijkstra variants, isochrones, dynamic programming — the industry has run on for decades.
The 2025 literature finally has controlled comparisons rather than demos. The honest answer turns out to be less flattering to AI than the headlines — and more useful.
Key findings
- In a controlled 2025 study, deep reinforcement learning beat a distance-based baseline by up to ~12% on fuel — but the best classical heuristic (Tabu search) still edged out the DRL by ~1%. Research demonstration
- AI's real contribution isn't raw optimality — it's adaptability: fine-tuning quickly when weather updates arrive mid-voyage. Research demonstration
- The field is fragmented and hard to compare — a systematic review found inconsistent terminology and evaluation across studies, making pooled conclusions weak. Established
- Weather routing is already an industry deployment — commercial services exist — but the ML-vs-classical question inside those products is proprietary and mostly unaudited publicly. Industry deployment
- The binding constraint at sea isn't optimization quality — it's forecast uncertainty. A perfect optimizer over a wrong forecast optimizes the wrong voyage. Maritime-GPT analysis
What the best evidence says
The cleanest recent experiment is a 2025 study in the Journal of Marine Science and Engineering that put two model-free deep RL agents — a Double Deep Q-Network and Deep Deterministic Policy Gradient — against classical and heuristic optimizers on the same voyage-optimization problem, with weather updating en route. Three results matter: Research demonstration
- Both DRL agents beat the naive distance-minimal route by up to roughly 12% fuel consumption.
- The continuous-action DDPG outperformed the discrete DDQN by about 4% — plausible, since speed and heading are continuous decisions.
- And the checkable caveat: a classical Tabu-search heuristic still beat the best RL agent by ~1%.
That last line is the discipline the field needed. RL's value wasn't producing a better route on a fixed forecast — classical methods already do that well — it was theoffline/online split: expensive pre-training before departure, then cheap re-fitting as updated forecasts arrive. Flexibility under moving information, not raw optimality, is the actual technical contribution. Maritime-GPT analysis
Why comparing papers is still hard
A systematic review of weather-routing optimization (Ocean Engineering; full text hosted by Chalmers) documents the problem bluntly: merchant-ship weather routing and autonomous-ship path research use overlapping terms for different problems, evaluation baselines differ between papers, and the post-2020 influx of ML methods arrived without a shared benchmark suite. Established A 2025 review reaches the same diagnosis — the field is active but fragmented, with limited real-time integration and weak explainability work. Established
Practical consequence: when a vendor claims "AI saves X% fuel," the defensible question is vs. what baseline, on which fleet, over what period, audited by whom? In published research, the honest baselines shrink the impressive numbers considerably. Maritime-GPT analysis
What this means aboard a real ship
Routing is a chain, and AI enters at the middle link. The chain is: forecast → ship response model → optimizer → master's decision → execution. ML can fit ship-response models (how much extra power for a given sea state, for your hull,your fouling state) better than generic curves — that's arguably the highest-value place for learning, because the physics is real but vessel-specific. Maritime-GPT analysis
The optimizer sees a forecast, not the weather. Ensemble spread on a 10-day forecast is huge; a pointwise-optimized route can be optimal against a scenario that never happens. Robust optimization under forecast uncertainty is where research should — and increasingly does — concentrate. Maritime-GPT analysis
Constraints are not negotiable tokens. A route that saves 4% fuel by clipping a heavy-weather margin the master wouldn't accept is not a saving — it's a risk transfer. Any deployable system needs the safety envelope enforced outside the learning layer, exactly the same architectural conclusion the collision-avoidance literature reached. Maritime-GPT analysis
Competing approaches, honestly compared
| Approach | Evidence | Strength | Weakness | Maturity |
|---|---|---|---|---|
| Classical optimization (isochrones, DP, Tabu) | Industry deployment | Deterministic, auditable, still competitive on fuel | Re-optimizing on new forecasts can be slow | Production |
| Deep RL (DDQN/DDPG) | Research demonstration | Adapts fast to updated weather; self-learning | ~1% behind best heuristic in the fairest test; opaque | Experimental |
| ML ship-response models inside classical optimizers | Industry deployment | Better fuel-vs-weather curves per vessel | Needs per-ship performance data | Deployed inside products |
| LLM/agentic voyage assistants | Proposed | Explains trade-offs, integrates text advisories | Not a numerical optimizer; hallucination risk | Early research |
What we know / don't / being tested
What we know
- Weather-aware routing saves fuel vs. distance-minimal routes — the direction of the effect is not in doubt.
- ML can match classical methods within ~1% and re-adapt faster mid-voyage.
- Ship-specific performance models (speed-power in waves) are learnable from operational data.
What we don't know
- Fleet-scale audited savings — almost no independent, at-scale verification exists publicly.
- How ML optimizers behave at distribution edges (cyclones, rogue conditions) where data is thinnest.
- How much of commercial "AI routing" is ML vs. rebranded classical optimization.
Being tested
- RL agents fine-tuned en-route as forecasts update.
- Stochastic/ensemble-aware objectives instead of point-forecast optimization.
- Digital-twin-coupled routing — the twin carries vessel state, the optimizer rides it.
Maritime-GPT analysis
The durable architecture is hybrid. Classical or RL optimizer for the numbers, learned ship-response model for the physics, deterministic safety envelope for the limits, and — this is where it gets interesting — a language-model layer as theinterface: turning an optimizer's trade-space into a briefing a watch officer can interrogate ("why 13.5 kn and not 15?"). Optimization chooses; the LLM explains. Systems that blur those two roles will produce confident nonsense; systems that separate them inherit the strengths of both. Maritime-GPT analysis
For operators evaluating vendors, the research suggests three discriminating questions: show me the baseline you beat; show me behaviour when the forecast was wrong; show me the safety constraints that sit outside the model. Maritime-GPT analysis
Open research problems
- A shared public benchmark: same routes, same forecast archives, same baselines — until then, cross-paper comparisons stay anecdotal.
- Uncertainty-aware objectives: optimizing expected fuel plus safety margin under ensemble forecasts, not single-run results.
- Explainable voyage optimization: an optimizer that can state its trade-offs in operational language remains an unsolved product problem.
- Learning under sparse extremes: tail-risk weather is exactly where training data is thinnest and consequences largest.
Evidence scorecard
| Dimension | Rating | Basis |
|---|---|---|
| Evidence strength | Moderate | Solid peer-reviewed comparisons exist but baselines and settings differ; no shared benchmark. |
| Research maturity | Field trial → deployed mix | Classical weather routing is production-grade; the ML variants are mostly experimental-to-field. |
| Maritime deployment | Limited / proprietary | Commercial routing services operate at scale; the ML content inside them is not independently auditable. |
| Confidence in conclusions | High | All figures quoted from peer-reviewed work; uncertainty ranges labelled. |
Related tools & reading
- Voyage time calculator
- Port-to-port distance calculator
- Canal vs. Cape route calculator
- Suez vs. Cape — what the diversion costs
- Route distance library
Sources
- Marine Voyage Optimization and Weather Routing with Deep Reinforcement Learning — J. Marine Science and Engineering (MDPI)Peer-reviewed (2025) — DOI 10.3390/jmse13050902
DDQN/DDPG voyage optimization: up to ~12% fuel improvement over a distance-based baseline; DDPG ~4% better than DDQN; Tabu-search heuristic still ~1% better than the best DRL. - State-of-the-art optimization algorithms in weather routing — ship decision support systems: challenge, taxonomy, and review — Ocean Engineering (authors' full text via Chalmers)Peer-reviewed review
Systematic taxonomy of optimization methods; documents the terminology confusion between merchant-ship weather routing and autonomous-ship research, and the post-2020 ML shift. - Smart Routing for Sustainable Shipping: A Review of Trajectory Optimization Approaches in Waterborne Transport — Sustainability (MDPI)Peer-reviewed review (2025)
Six-area survey of routing research; finds the field fragmented, with limited real-time integration and underdeveloped explainability.
Fact-check: percentages cited are the figures reported by the named studies (their baselines, their test conditions); they are not universal savings. Routing-model distances on this site are planning estimates. (Accessed 5 October 2026.) Not navigation or commercial advice.