Machine Learning as Applied to Shape Parameterization of Submerged Arch Structures

Submerged structures are usually designed under arch-type forms. Such forms may be performed through different parametric functions, and their design procedure usually involves several variables in order to obtain an efficient, economical and safe installation. Past works have found that for intermediate depth ratios, the momentless (or funicular) shape of the submerged arch is a geometrical form between the parabola and the ellipse curve. Alternatively, some authors have proposed the shape parameterization of the submerged arch’s centerline in order to approach an efficient design that takes into account not only the arch’s mechanical behaviour, but other design aspects such as its enclosed airspace or the arch’s serviceability. Such processes traditionally involve, regardless of an optimizer, the implementation of a finite element model and the corresponding parametric function. However, this may be greatly simplified through the use of machine learning. In this work, several regressors are trained in order to predict the fitness function that governs the multi-objective optimization of a submerged arch. To this aim, three different parametric functions are considered: conics, elliptics and Bézier curves, and a comparison of the regressor performance for each parametric shape function is presented.