← Back to Publications
Conference Document • Delft Hydrodynamics Database Series

Ship Residuary Resistance Prediction with Machine Learning

Poorya Khorsandy Department of Marine Engineering
Seyed Saeed Hayati Khorramshahr University of Marine Science and Technology
Download PDF View Source Code

Abstract

This paper explores machine learning techniques as a cost-effective, accurate, and relatively fast alternative to traditional methods for yacht residuary resistance calculation. Currently, solving CFD equations is a well-known method in residuary resistance calculation. However, solving CFD equations requires an iterative process which is highly time-consuming and demands high-performance processing units. Instead of solving CFD equations for every individual ship geometry, we propose applying machine learning algorithms to historical experiment arrays. We utilized the Delft yacht hydrodynamics dataset obtained from 308 experimental runs to predict the resistance coefficients corresponding to various hull parameters. Our results demonstrate that the XGBoost regressor is superior to other structural engines with an RMSE of 0.54.

1. Introduction & Background

Accurate ship residuary resistance prediction is vital in naval architecture to optimize hull design and maximize fuel efficiency. Traditional approaches depend heavily on physical tank scale tests or detailed computational fluid dynamics configurations. While these strategies deliver high precision, they introduce severe resource constraints into early conceptual design iterations.

By shifting towards data-driven predictive surrogates, design offices can analyze thousands of parametric variations in real-time. This paper evaluates the regression mapping capabilities of tree-based ensembles against classical statistical learning architectures.

2. Dataset & Hull Geometric Bounds

The research relies on the legacy Delft Yacht Hydrodynamics Dataset, containing 308 experimental records testing variations across 22 individual hull shapes derived from the Standfast 43 design. The models leverage 6 primary geometric parameters alongside the Froude number:

  • Longitudinal Position of Center of Buoyancy ($LC$)
  • Prismatic Coefficient ($PC$)
  • Length-Displacement Ratio ($L/\nabla^{1/3}$)
  • Beam-Draught Ratio ($B/T_c$)
  • Length-Beam Ratio ($L_{wl}/B$)
  • Froude Number ($Fr$) — tested within the range of 0.125 to 0.450.

3. Mathematical Framework

The core objective function seeks to map a multidimensional structural array directly to the residuary resistance coefficient ($C_r$). The total drag profile is captured via objective loss convergence mappings:

$$\text{Obj}(\theta) = \sum L(y_i, \hat{y}_i) + \sum \Omega(f_k)$$

Where $\Omega(f_k)$ measures model complexity penalties to eliminate overfitting inside sparse fluid boundaries.

4. Hyperparameter Tuning Matrix

To establish optimization parity across gradients, the XGBoost engine was structured with the following tuned constraints:

Hyperparameter Optimized Value Operational Bound
Learning Rate (eta) 0.05 [0.01 - 0.20]
Max Tree Depth 6 [3 - 10]
Subsample Ratio 0.80 [0.50 - 1.00]

5. Algorithmic Evaluation & Convergence

The algorithms were systematically benchmarked using a 70/30 train/test data split distribution. Performance metrics reveal that Gradient Boosted Tree architectures excel at mapping non-linear interactions across dimensional fluid bounds, while Support Vector Regression struggles with sparse data extremes.