1. Introduction
The international shipping industry serves as the backbone of global trade, carrying over 80% of international cargo and playing a critical role in global economic integration (Fan et al., 2025; World Economic Forum, 2024; Yan et al., 2024). However, this vital role comes at a significant environmental cost.
The sector is one of the largest consumers of fossil fuels, contributing approximately 3% of total global greenhouse gas (GHG) emissions (Transport & Environment, 2024). In response to the escalating climate crisis, the International Maritime Organization (IMO) has implemented stringent regulatory measures, such as the Energy Efficiency Index for Existing Ships (EEXI) and the Carbon Intensity Indicator (CII), to steer the industry toward a net-zero future (International Maritime Organization, 2023; Tadros et al., 2023).
Achieving these ambitious decarbonisation targets requires a profound transformation in how vessel energy efficiency is managed. While technological innovations are crucial, operational strategies, such as speed regulation, route optimisation, and trim adjustment, are recognised as the most cost-effective and immediately applicable solutions (Godet et al., 2024; Yan et al., 2024). The effectiveness of these strategies fundamentally depends on the accurate prediction and deep understanding of Ship Fuel Consumption (SFC). An accurate SFC model not only provides a quantitative foundation for evaluating energy-saving interventions but also forms the theoretical basis for intelligent energy management systems that will underpin a sustainable maritime future (Wang et al., 2023).
Despite recent advances in applying machine learning (ML) to SFC prediction, existing approaches face several fundamental limitations that constrain their practical applicability and scientific impact. This research identifies three core and interconnected challenges, which form the foundation of the proposed study.
Challenge 1: The “Homogeneity Assumption” and Operational Heterogeneity
Current modelling paradigms often treat the complex and dynamic process of ship navigation as a statistically homogeneous system, assuming that a single global model can represent SFC mechanisms across all operational scenarios (Feng et al., 2024). This “homogeneity assumption” neglects the intrinsic diversity of ship operations, where different conditions, such as cruising in rough seas, navigating calm waters, or manoeuvring in port, are governed by distinct physical dynamics and energy consumption behaviours (Fan et al., 2025). Such oversimplification obscures essential variations, limits model generalisability, and results in unreliable predictions under non-standard conditions, thereby impeding the development of targeted, condition-specific energy optimisation strategies (Iqbal et al., 2025).
Challenge 2: The “Black Box” Problem and the Correlation-Causation Fallacy
Although advanced ML models achieve high predictive accuracy, their “black box” nature hinders transparency, credibility, and adoption in safety-critical maritime operations (Wang et al., 2023). Most interpretability studies provide only a ranking of feature importance, failing to uncover the complex non-linear relationships, thresholds, and inflection points that truly describe how factors influence SFC (Xie et al., 2024). More importantly, these models reveal statistical correlations rather than causal mechanisms (Huang et al., 2024). This distinction is crucial: effective energy efficiency policies must be grounded in an understanding of causal drivers, not merely correlated variables.
Challenge 3: Methodological Fragmentation and High-Dimensional Complexity
Current research remains fragmented, with limited integration between predictive modelling, explainable AI (XAI), and causal inference methods (Li et al., 2024). Additionally, the fusion of multi-source data—such as noon reports, AIS, and environmental datasets—introduces challenges including feature redundancy, multicollinearity, and the curse of dimensionality, which can increase computational burden and lead to overfitting (Afshar & Usefi, 2020). Existing feature selection and hyperparameter optimisation techniques are often single-model-based, prone to bias, or computationally intensive, making them unsuitable for complex, high-dimensional maritime datasets (Muñoz et al., 2025).
These challenges lead to three central research questions that will guide this doctoral study:
RQ1: How can the intrinsic heterogeneity of ship operational states be systematically identified, modelled, and interpreted to develop more accurate, robust, and physically meaningful SFC prediction frameworks?
RQ2: How can we move beyond correlation to establish and quantify the true causal relationships between key operational, navigational, and environmental factors and SFC to inform effective and reliable intervention strategies?
RQ3: How can a unified analytical framework be developed to integrate advanced data fusion, robust feature selection, efficient model optimisation, and deep interpretability - ultimately creating transparent, trustworthy, and actionable decision-support tools for ship energy management?
This research aims to develop an AI-aided, interpretable, and causally informed framework for ship fuel consumption prediction, bridging the gap between predictive accuracy, scientific understanding, and practical applicability. By addressing the challenges of heterogeneity, causality, and complexity, this study will advance both the methodological frontier of maritime AI research and the practical goal of achieving low-carbon, energy-efficient shipping operations.
References:
Afshar, M., Usefi, H., 2020. High-dimensional feature selection for genomic datasets. Knowledge-Based Systems 206, 106370. https://doi.org/10.1016/j.knosys.2020.106370
Birchler, J.A., Veitia, R.A., 2007. The Gene Balance Hypothesis: From Classical Genetics to Modern Genomics. The Plant Cell 19 (2), 395-402. https://doi.org/10.1105/tpc.106.049338
Fan, A., Wang, Y., Hu, Z., Yang, L., Fan, X., Yang, Z., 2025. Multi-dimensional performance verification of ship fuel consumption prediction model under dynamic operating conditions. Energy 332, 137120. https://doi.org/10.1016/j.energy.2025.137120
Feng, Y., Wang, X., Chen, Q., Yang, Z., Wang, J., Li, H., Xia, G., Liu, Z., 2024. Prediction of the severity of marine accidents using improved machine learning. Transportation Research Part E: Logistics and Transportation Review 188, 103647. https://doi.org/10.1016/j.tre.2024.103647
Godet, A., Panagakos, G., Barfod, M.B., Lindstad, E., 2024. Operational cycles for maritime transportation: Consolidated methodology and assessments. Transportation Research Part D: Transport and Environment 132, 104238. https://doi.org/10.1016/j.trd.2024.104238
Huang, H., Li, B., Wang, Y., Zhang, Z., He, H., 2024. Analysis of factors influencing energy consumption of electric vehicles: Statistical, predictive, and causal perspectives. Applied Energy 375, 124110. https://doi.org/10.1016/j.apenergy.2024.124110
International Maritime Organization, 2023. EEXI and CII - ship carbon intensity and rating system. https://www.imo.org/en/mediacentre/hottopics/pages/eexi-cii-faq.aspx
Iqbal, S., Qureshi, A.N., Alhussein, M., Aurangzeb, K., Mahmood, A., Razalli Bin Azzuhri, S., 2025. Dynamic selectout and voting-based federated learning for enhanced medical image analysis. Machine Learning: Science and Technology 6 (1), 015010. https://doi.org/10.1088/2632-2153/ada0a6
Jia, Y., Hu, X., Kang, W., Dong, X., 2024. Unveiling Microbial Nitrogen Metabolism in Rivers using a Machine Learning Approach. Environmental Science & Technology 58 (15), 6605-6615. https://doi.org/10.1021/acs.est.3c09653
Li, S., Pu, Z., Cui, Z., Lee, S., Guo, X., Ngoduy, D., 2024. Inferring heterogeneous treatment effects of crashes on highway traffic: A doubly robust causal machine learning approach. Transportation Research Part C: Emerging Technologies 160, 104537. https://doi.org/10.1016/j.trc.2024.104537
Muñoz, V., Ballester, C., Copaci, D., Moreno, L., Blanco, D., 2025. Accelerating hyperparameter optimization with a secretary. Neurocomputing 625, 129455.