⚡ electrical engineering

Machine Learning-Based Prediction and Optimization of Ship Design Particulars for CO₂ Emission Reduction

This study develops a data-driven framework using four supervised learning algorithms to predict ship design particulars and optimize engine power for CO₂ emission reduction, demonstrating that Gaussian Process Regression offers the most consistent performance while a displacement-based reformulation enhances the physical consistency of block coefficient estimation.

Original authors: S M Rashidul Hasan, Md. Shariful Islam, Zobair Ibn Awal, Khandakar Akhter Hossain

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: S M Rashidul Hasan, Md. Shariful Islam, Zobair Ibn Awal, Khandakar Akhter Hossain

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ocean as a giant, churning highway where massive ships are the trucks and buses, carrying everything from your sneakers to your dinner. For decades, the people who design these ships—naval architects—have been like master chefs trying to perfect a recipe. They use old-school cookbooks (physics equations) and taste tests (trial and error) to figure out how big a ship should be, how deep it should sit in the water, and how much fuel its engine needs to burn. But the world is changing. The "climate police" (international regulators) are now demanding that these ships pollute less, and the old cookbooks are sometimes too slow or too simple to handle the messy, complex reality of different ship types, especially the smaller ones that zip around rivers and coastlines.

Enter Machine Learning, the digital apprentice that can taste a thousand recipes in a second and spot patterns humans might miss. Think of it as a super-smart detective that looks at a pile of data about how ships actually behave and tries to guess the perfect design before a single piece of metal is cut. The big question isn't just "Can we build a ship?" but "Can we build a ship that gets us where we need to go without burning the planet?" This is where the story gets exciting: using computers to learn from the past to design a cleaner future.

The Ship Designer's New Crystal Ball

In this study, a team of researchers from Bangladesh decided to give ship designers a serious upgrade. Instead of relying solely on traditional math, they built a digital "crystal ball" using Machine Learning. They fed their computer a massive library of data from three very different types of ships: cargo vessels (the heavy haulers), oil tankers (the liquid carriers), and passenger ships (the river buses). They wanted to see if the computer could look at a few basic facts—like how big the ship is, how fast it goes, and how much weight it carries—and then predict the perfect design details, specifically how much engine power is needed.

The researchers didn't just pick one guessing game; they set up a four-way race between four different types of AI detectives:

Artificial Neural Networks (ANN): Think of this as a brain with many layers, great at spotting complex, squiggly patterns.
Random Forests (RF): Imagine a crowd of experts voting on the answer; it's very stable and hard to trick.
Support Vector Regression (SVR): A sharp, precise tool that draws a line to separate the best answers from the rest.
Gaussian Process Regression (GPR): The cautious mathematician that not only gives an answer but also says, "I'm pretty sure about this, but here's the range of possibilities."

The Big Discovery: The Easy and the Hard
When the race started, the results were surprisingly clear. The AI detectives were amazing at predicting engine power. If you told them the ship's size and speed, they could guess the engine size with incredible accuracy. It's like if you could look at a car's weight and speed and instantly know exactly how big the engine needs to be.

However, they hit a wall with one specific detail: the block coefficient. In plain English, this is a number that describes how "boxy" or "streamlined" the ship's underwater shape is. The AI struggled to guess this directly. It's like trying to guess the exact shape of a secret object just by feeling its weight; it's too many variables interacting at once.

The Clever Workaround
Here is where the researchers got creative. Since the AI was bad at guessing the "boxiness" directly, they changed the game. Instead of asking the AI to guess the shape, they asked it to guess the displacement (how much water the ship pushes aside, which is basically its weight in water). The AI was fantastic at this! Once the computer knew the weight, the researchers used a simple physics rule to calculate the "boxiness" from there. It was like realizing you can't guess the shape of a cake directly, but if you guess the weight of the ingredients, you can figure out the shape easily. This two-step trick made the whole design process much more reliable and physically sensible.

The Winner of the Race
So, which AI detective won? It depends on the ship.

Gaussian Process Regression (GPR) was the most consistent all-rounder, performing well across all ship types.
Support Vector Regression (SVR) was the star for oil tankers.
Random Forests were the most stable for passenger ships.
Neural Networks were good at learning the complex curves but weren't the most consistent.

The paper suggests that there is no single "magic bullet" AI that works for every ship. Instead, the best approach is to pick the right detective for the specific job.

From Prediction to Optimization
The study didn't stop at just guessing. The researchers took their best-performing models and plugged them into a "design optimizer." Imagine a video game where you have to build a ship, but you have a strict budget for fuel. The computer used the AI's predictions to tweak the ship's design—changing its length, width, or draft—until it found the version that used the least amount of engine power while still meeting all the safety rules.

The results showed that for cargo ships, the model was incredibly accurate (98.9% match), and for oil tankers and passenger ships, it was still very strong (around 90% match). This means the tool isn't just a theoretical toy; it can actually help engineers design ships that burn less fuel and produce less CO₂ right from the very first sketch.

The Catch
The researchers are honest about the limits of their work. They admit that their "crystal ball" was trained on a specific set of ships from their region. While it works great for those, they can't be 100% sure it will work perfectly for a completely different type of ship they've never seen before. It's like a chef who is a master at cooking Bangladeshi cuisine; they might struggle if asked to cook a specific French dish they've never tried. The paper suggests that while this is a huge step forward for preliminary design, the models need to be tested on even more diverse data before they can be trusted for every single ship in the world.

In short, this paper shows that by combining smart AI with a little bit of clever physics tricks, we can design ships that are not only built faster but are also kinder to the planet. It's a promising start to a future where the ships of tomorrow are as efficient as they are powerful.

Technical Summary: Machine Learning-Based Prediction and Optimization of Ship Design Particulars for CO₂ Emission Reduction

Problem Statement
Shipping is a significant contributor to global greenhouse gas emissions, prompting the International Maritime Organization (IMO) to strengthen strategies for reducing carbon intensity. While traditional naval architectural methods rely on first-principles hydrodynamics and empirical correlations, they often struggle to capture complex nonlinear interactions among hull geometry, loading conditions, and operational parameters, particularly within heterogeneous inland and shallow-water vessel fleets. Existing machine learning (ML) studies in maritime engineering largely focus on ocean-going vessels or operational predictions (e.g., fuel consumption) without providing principled uncertainty quantification or interpretable frameworks for predicting principal design particulars of diverse inland fleets. There is a specific gap in developing transferable, uncertainty-aware ML frameworks for heterogeneous inland vessel classes, such as oil tankers and passenger ships, to support low-emission preliminary design.

Methodology
The study proposes a data-driven framework integrating supervised learning with constrained optimization to predict principal ship particulars and support energy-efficient design.

Data Assembly and Preprocessing: The research utilizes a harmonized dataset comprising three vessel classes: cargo vessels, oil tankers, and passenger vessels. Key variables include length of waterline (LWL), breadth (B), draft (T), block coefficient ( $C_B$ ), deadweight/gross tonnage (DWT/GRT), service speed ( $V_s$ ), and engine power ( $P_{ME}$ ). The data underwent rigorous preprocessing, including outlier detection using Z-score and Interquartile Range (IQR) methods, resulting in refined datasets of 262 cargo vessels, 281 oil tankers, and 434 passenger vessels.
Model Selection and Training: Four supervised learning algorithms were evaluated: Artificial Neural Networks (ANN), Random Forests (RF), Support Vector Regression (SVR), and Gaussian Process Regression (GPR).
- ANN: Utilized a two-hidden-layer architecture with ReLU activation and RMSProp/Adam optimizers to capture nonlinear relationships.
- RF: Employed ensemble averaging of decision trees to improve generalization and provide feature importance.
- SVR: Used kernel functions (RBF and polynomial) to handle high-dimensional feature spaces with controlled model capacity.
- GPR: Applied a probabilistic, non-parametric approach to provide predictions alongside uncertainty estimates.
- Hyperparameters were tuned via cross-validation for each vessel class.
Methodological Innovation (Block Coefficient): Recognizing that direct prediction of the block coefficient ( $C_B$ ) yielded low accuracy, the authors introduced a two-stage approach. Displacement was predicted first, and $C_B$ was subsequently derived using the hydrostatic relation between predicted displacement and principal dimensions. This converted a difficult ratio-estimation problem into a more stable regression task.
Constrained Optimization: Validated ML models were embedded into a MATLAB-based optimization framework using fmincon and genetic algorithms. The objective was to minimize engine power ( $P$ ) subject to practical design constraints (deadweight, speed, draft). Second-order polynomial objective functions were derived for each vessel class to represent the nonlinear relationship between design variables and power demand.

Key Results

Predictive Performance: Engine power was identified as the most predictable target across all vessel classes, with high coefficients of determination ( $R^2$ ). Length was also predicted reliably. Conversely, block coefficient remained the most difficult variable to estimate directly.
Model Comparison:
- GPR demonstrated the most consistent performance overall, particularly for block coefficient estimation, due to its ability to handle uncertainty and nonlinear variation.
- SVR performed strongly for oil tankers.
- RF provided stable predictions for passenger vessels.
- ANN effectively captured nonlinear patterns but showed variability depending on the vessel class.
Optimization Outcomes: The integration of predictive models into the optimization framework successfully identified feasible design configurations that reduce engine power. The objective functions achieved high accuracy ( $R^2$ = 0.989 for cargo, 0.906 for oil tankers, and 0.901 for passenger vessels), confirming the suitability of the models for design exploration.
Block Coefficient Improvement: The displacement-based reformulation significantly improved the physical consistency and robustness of block coefficient estimation compared to direct prediction.

Significance and Claims
The paper claims four primary contributions:

Dataset Harmonization: The development of a unified inland-vessel dataset covering cargo, oil tanker, and passenger classes.
Algorithmic Evaluation: A comparative analysis of ANN, RF, SVR, and GPR, demonstrating that no single algorithm is universally superior; model selection should be driven by vessel class and target variable.
Physically Consistent Modeling: The introduction of a displacement-based approach to improve the estimation of block coefficient, aligning with recommendations for physics-informed ML.
Design Integration: The successful linkage of predictive models with constrained optimization to create a decision-support tool for low-emission preliminary design.

The authors position this work as a step toward bridging the gap between data-driven prediction and practical naval architectural decision-making. They emphasize that while the framework is effective for preliminary design within the observed data domain, it relies on internal cross-validation and lacks external validation on entirely separate datasets. Consequently, the study advocates for caution when extrapolating results to vessel configurations outside the training range and calls for future work involving larger datasets, external validation, and physics-informed hybrid formulations to enhance generalizability.

The Ship Designer's New Crystal Ball

Technical Summary: Machine Learning-Based Prediction and Optimization of Ship Design Particulars for CO₂ Emission Reduction

More like this