Carbon intensity indicator rating-based ship retrofit strategy framework via deep reinforcement learning

Qi Wei, Ying Liu, Yue Dong, Dan M. FrangopolOpenalex Unknown 2025 · DOI · OpenAlex

This paper presents a novel approach to one of the maritime industry's most pressing financial and environmental headaches: how to keep aging ships compliant with tightening carbon regulations without breaking the bank.

TL;DR

Researchers have developed a Deep Reinforcement Learning (DRL) framework that optimizes ship retrofitting strategies over a vessel's remaining lifespan. By modeling the transition between Carbon Intensity Indicator (CII) ratings as a sequential decision problem, the system identifies the most cost-effective timing and combination of technologies—such as carbon capture and wind propulsion—to ensure regulatory compliance at the lowest life-cycle cost.

What Problem Does This Paper Address?

The International Maritime Organization (IMO) has introduced the Carbon Intensity Indicator (CII), a mandatory rating system (A through E) for ships. Vessels with poor ratings (E, or three consecutive Ds) face "corrective action plans," which translate to higher insurance premiums, asset depreciation, and potential port restrictions.

For owners of existing "aging" ships, the challenge is twofold:

  1. Complexity: There are many retrofit options (wind assistance, hull coatings, alternative fuels, carbon capture), each with different costs and emission-reduction potentials.
  2. Timing: CII requirements get stricter every year. A ship that is compliant today might be a "failing" ship in three years.

Existing research often looks at these technologies in isolation. This paper addresses the "fusion perspective"—how to stack and time these retrofits over a decade or more to minimize total costs while staying in the "green" rating zones.

Methods

The authors frame the retrofitting journey as a Markov Decision Process (MDP) and solve it using Deep Reinforcement Learning.

  • Dueling Double Deep Q-Network (Dueling DQN): This specific DRL architecture was chosen for its ability to handle multi-dimensional states. It separates the estimation of the "state value" (how good it is to be in a certain CII rating) from the "action advantage" (how much better one retrofit is over another).
  • The State Space: The agent monitors four variables: remaining ship life, acquired CII, current CII rating, and the required reduction ratio.
  • The Action Space: The agent can choose from five mature retrofit technologies:
  • Antifouling coatings (7.5% reduction)
    • Battery power systems (10% reduction)
    • Wind propulsion/Flettner rotors (15% reduction)
    • LNG fuel conversion (25% reduction)
    • Carbon capture systems (75% reduction)
  • Cost Model: The "reward" function for the DRL agent includes capital expenditure (CAPEX), operating costs (OPEX), fuel savings, and a quantized benefit/penalty for the CII rating (ranging from +5 for an 'A' to -10 for an 'E').

Datasets

The study uses a "digital twin" approach for evaluation:

  • Ship Model: A standard container ship (KCS) provided by the Korean Register of Shipping.
  • Performance Data: Retrofit costs and emission reduction percentages were synthesized from a meta-analysis of recent engineering literature (e.g., Blanco-Davis et al. for coatings, Chica et al. for wind propulsion).
  • Economic Data: The model incorporates EU carbon tax estimates to calculate the financial benefit of emission reductions.

Key Insights and Ideas

  • Front-Loading is Profitable: The DRL agent consistently chooses to implement high-impact retrofits (like carbon capture or LNG) very early in the ship's remaining life. The "aha moment" here is that the cumulative benefits of maintaining an 'A' rating—and the fuel savings over time—far outweigh the high initial CAPEX.
  • The "Late-Life" Taper: As a ship approaches the end of its service life (the last 5 years), the agent shifts to "no action" or minor retrofits. At this stage, the remaining time is too short to recoup the investment of major hardware overhauls.
  • Strategic Compliance: The model doesn't just aim for "passing" (C rating); it aggressively pursues 'A' and 'B' ratings because the framework assumes (based on IMO recommendations) that high-rated ships will receive preferential port treatment and lower insurance costs.

Key Results

  • Convergence: The DRL agent successfully learned an optimal policy, with the total cost curve converging after approximately 1,500 training episodes.
  • Scenario Analysis:
    • For a ship with 15 years of life remaining and a starting rating of 'E', the agent immediately triggers major retrofits to jump to an 'A' rating.
  • For ships with only 5 years left, the strategy is much more conservative, as there isn't enough time to reach a "break-even" point on expensive tech like carbon capture.
  • Economic Logic: The study demonstrates that a "combined" retrofit policy is essential; relying on a single technology is rarely the most cost-effective path over a 15-year horizon.

Limitations and Future Work

  • Independence Assumption: The authors assume that the five retrofit technologies are independent and their effects are additive. In reality, physical space on a ship is limited; you might not be able to fit both a massive carbon capture system and a fleet of Flettner rotors.
  • Data Generalization: The costs are based on specific engineering cases. Real-world costs vary wildly based on shipyard location and global supply chains.
  • Regulatory Uncertainty: The "benefit" of a high CII rating is currently modeled as a proxy value (e.g., +5). As global carbon markets and port incentives mature, these numbers will need to be replaced with actual dollar-value regulations.

Future work aims to integrate more "immature" technologies as data becomes available and to account for the physical conflicts between different hardware installations.

Summary updated

Abstract

The tightening CII regulations has brought challenges to the maritime industry. For a huge number of aging ships, retrofit is vital for stakeholders to achieve IMO’s targeted carbon emission reduction of 20-30 https://www.w3.org/1998/Math/MathML" display="inline"> % by 2030. At present, the retrofitting approaches generally includes wind assisted ship propulsion, air lubrication biofouling management, electrification, low carbon emission fuel, etc. Considering emission reduction effect and economic efficiency, a combined retrofit policy is crucial for each ship in demand. Considering that the requirements of CII regulations are tightening gradually, adopting optimal retrofit approaches at the right time based on ship’s decarbonization conditions is vital for stakeholders. Deep reinforcement learning (DRL) is introduced to acquire the optimal ship retrofit policy for lowest life-cycle cost under the supervision of CII regulations. Herein, a retrofit strategy framework based on CII is established for decision-making support in the context of carbon reduction.
BibTeX
@inproceedings{wei2025carbon,
  author = {Qi Wei and Ying Liu and Yue Dong and Dan M. Frangopol},
  title = {{Carbon intensity indicator rating-based ship retrofit strategy framework via deep reinforcement learning}},
  booktitle = {Openalex Unknown},
  year = {2025},
  doi = {10.1201/9781003595120-171},
}

Loading related content...

Figure