This study investigates the influence of data preprocessing techniques on the development of an artificial neural network (ANN)-based surrogate model for predicting the hydrodynamic coefficients of a two-dimensional rectangular barge. Conventional potential-flow-based numerical simulations are associated with high computational costs and time constraints, motivating the need for ANN models that can provide fast and efficient predictions. However, the predictive performance of ANNs strongly depends on the statistical characteristics of the training data. In this work, three types of input datasets—raw data, cube-root-transformed data, and Yeo–Johnson-transformed data—were used to train a representative multilayer perceptron (MLP) ANN model, and their prediction accuracies were systematically compared. The results show that models trained with preprocessed data consistently outperformed those trained with raw data across all hydrodynamic coefficients. In particular, for coefficients exhibiting strong nonlinear behavior, such as added mass and radiation damping coefficients, the Yeo–Johnson transformation yielded the highest prediction accuracy. The findings highlight the essential role of appropriate data preprocessing in enhancing the reliability and accuracy of data-driven predictive models in ocean engineering and provide a foundation for the development of robust surrogate models for high-dimensional engineering problems.
Son et al. (Fri,) studied this question.