DNN-based inverse design of line heating patterns for automated plate forming in shipbuilding using multi-start convex optimization

Abstract

Line heating is a widely used plate forming technique in the shipbuilding industry, where steel plates are heated along specified paths to achieve desired deformations. Traditionally, the design of these heating patterns relies on the expertise of skilled workers due to the complex and nonlinear relationship between the heating patterns and the resultant plate deformations. This reliance often results in inconsistent productivity and quality. This study presents a data-driven inverse design framework that automates the optimization of line heating patterns for specified plate deformations, addressing the need for rapid and systematic methodologies. A deep neural network (DNN) trained on finite element method (FEM) simulation data, which validated against experimental results, is employed to model the relationship between initial plate geometry, line heating patterns, and resultant deformations. This surrogate model enables rapid predictions of deformed plate shapes. Utilizing the trained DNN, a multi-start convex optimization process is employed to identify the optimal line heating patterns for any given initial plate geometry and desired deformation. The proposed framework demonstrates significant potential for various engineering inverse design applications requiring prompt and accurate results, as validated by designing line heating patterns for 800 different plate shapes, achieving desired deformations not included in the training data.