1 INTRODUCTION The structural design is one of the fields where the applications of optimizations techniques have been most studied. The objective is to reduce the structural mass assuring the integrity of the structure. Although a heavier structural arrangement can be more resistant than a lighter structural arrangement, the lighter one may be more efficient if one consider that structural efficiency could be defined as the ratio between the maximum load the structure can withstand with no structural failure and the structure’s weight. Ships structures are typical examples of structural arrangements based on this concept of structural efficiency, as the reduction of the ship’s steel mass may be associated to the reduction of its fabrication cost and a better performance. Several works have been developed recently aiming to reduce ship’s steel mass, such as: (Huang, 2009), (Jastrzebski, 2005) e (Sekulski, 2003). In the traditional approach, which is based on the well-known idea of the design spiral (Evans, 1959), the conception of the initial structural arrangement, in early design phases, is done using ship’s classifications rules. In this approach, the main objective is to conceive a structural arrangement that satisfies the criteria required by these rules. Finding a configuration, which satisfies these criteria and is optimized regarding its structural weight, strongly depends on the engineer’s experience. Yet, it’s impossible to be sure that the found solution is the one with the lowest weight possible. Determining the structural steel weight and also the position of its center of mass is very important to a correct and coherent convergence of the design, as these characteristics affects many others attributes like stability or cargo capacity. In the early design stages, the main characteristics of the vessel are not known yet and that’s why the structural design using the classification rules may become a long and exhaustive process, as it has to be Optimal Structural Design of Small Ships with Neural Network Response Surface V. T. Chaves, T. P. Tancredi & B. L. R. de Andrade Laboratório de Otimização e Projeto Integrado Escola Politécnica da Universidade de São Paulo, São Paulo, Brazil ABSTRACT: This article presents the development of a prediction model based on neural networks that are trained with data of optimized structural arrangements and are capable of estimate the structural properties for different hull’s geometries. This model can only be applied for ships with less than 100 meters of length and built in steel. Therefore, this model is going to be able to give, yet in the early stages of the design, an estimative of a structural op- timized arrangement with the lowest mass possible that satisfies all the required specifications by the rules for ship classification. This estiŵatiǀe ǁill ďe ďasiĐally ďased oŶly oŶ a feǁ paraŵeters as ship’s leŶgth, ďeaŵ, draft, depth aŶd block coefficient and on certain design variable as the spacing between frames and the spacing between the longitu- dinals and girders. Afterwards, a parametric analysis for a better comprehension of the optimum structural arrange- ments will be done and will be presented practical recommendations for an optimum structural design for ships built in steel with less than 100 m of length.