Resistance and Trim Modeling of a Systematic Planing Hull Series 62 (with 12.5°, 25°, and 30° Deadrise Angles) Using Artificial Neural Networks, Part 1: The Database

Abstract

Recent advances in high-speed computing, combined with the emergence of artificial neural network (ANN) techniques, for the analysis of large data sets have enabled researchers to provide the design community with higher-resolution mathematical models (MMs) for existing test data. Presently, one of the most popular planing hull prediction methods for resistance and trim are based on regressions of the Series 62 database. New MMs developed here address two major shortcomings of the original approaches; first, the equations are now continuous functions of volumetric Froude number (rather than separate regressions for each speed) and second, MM for trim is much more accurate, enabling designers to identify the double hump in trim that is associated with dynamic instabilities at higher speeds. This work includes not only the original David Taylor Model Basin (TMB) Series 62 data for 12.5° deadrise, but also the later extensions made by Delft University of Technology (DUT), including 25° and 30° deadrise. The present report, Part 1, explains the procedures used to streamline a large foundational database to prepare for the derivation of an MM. This step is usually not discussed in detail, but the success of the entire procedure depends on it. In large data sets collected over multiple decades, there are often outliers in the data and pockets of the test matrix with insufficient test data for successful fitting of MMs. Because of the lack of data in these pockets, fitting routines for MMs have no incentive to produce rational results in these areas, often leading to an unstable model. In the past, overly stiff models were used to fair through these regions, at the expense of reduced accuracy in regions where there were sufficient data. This report describes the addition of "virtual measurements" based on interpolation or engineering calculations, which enable the model to produce reasonable results in regions of limited data, while also remaining accurate in regions with sufficient data. Additionally, an iterative procedure, where preliminary MMs are used to identify and eliminate outliers and erroneous points is described. The techniques described here can be applied to improve fitting of many types of data sets in Naval Architecture.