AI Model Adjustment for Determination of Ship Design Characteristics

Abstract

This paper presents a detailed study of machine learning methods application for predicting ships key hydrodynamic parameters. The relevance of the work is due to the need to reduce the time and financial costs of traditional hydrodynamic calculations in ship design. In the study a comparative analysis of three machine learning methods – Linear Regression, Random Forest, Gradient Boosting is carried out. Experiments are performed on an original dataset including 52 existing vessels of Feeder and Panamax classes, with characteristics: length, width, draft, speed, wetted surface area and total completeness ratio. The focus is on predicting Froude and Reynolds numbers for preliminary drag estimation. The results demonstrate the superiority of linear regression: for Froude number, R² = 0.9986 was achieved at MSE = 5.15e-07 and for Reynolds number, R² = 0.9998 at MSE = 9.00e+13. The mean absolute percentage error (MAPE) was 0.15% and 0.57%, respectively. The final result is an AI-based calculation program that is able to read Fr and Re values. The practical significance of the work is due to the demonstration of the possibilities of the new calculation methods, which, with a proper approach, can improve the accuracy and optimize the process of classical calculation. The obtained results allow to recommend the method of linear regression for preliminary assessment of hydrodynamic characteristics at the early stages of design. The practical significance of the work is due to the possibility of improving the accuracy of hydrodynamic calculations and hull shape optimization based on the predicted data. Prospects for further research include expansion of the data set, consideration of additional parameters and testing of hybrid models.