Leveraging machine learning in material and structural design: applications in heating pattern, composite configuration, and rib pattern optimization

Abstract

In recent years, the rapid advancement of computer performance has enabled the prediction and mapping of various complex problems previously unsolved by humans through deep learning. Consequently, surrogate model-based optimization, which involves training predictive models based on data and using them for design optimization, has gained significant attention. This dissertation presents various design optimizations in several industrial fields, including heavy industry, composite materials, and automotive parts, considering the size of the design space and the dimensions of the variables. The first study addresses the design problem of heating pattern for forming steel plates used in shipbuilding. Ships' exteriors are constructed by welding curved surfaces with varying sizes of twisting and bending. The process of creating desired curved surfaces from flat plates involves localized thermal deformation through line heating. However, designing heating line pattern based on trial and error by practitioners leads to considerable time and cost inefficiencies. This research predicts the angle of twist and bending displacement of the processed curved surface through geometric variables such as the spacing, length, and angle of the heating lines. An AI-based optimization methodology was employed to design these heating line variables to achieve the desired angle of twist and bending displacement. The design space was small and continuous with upper and lower bounds, requiring rapid solution provision. Thus, the model was trained with data collected through grid sampling, including boundaries, making the entire design space a seen domain. High-performance models derived in this way facilitated swift optimization. The second study focuses on the pattern optimization of grid composites. Grid composites, a fundamental form of composites, consist of materials with different properties arranged in a lattice structure. However, they suffer from digital design variables and an enormous design space. Hence, extrapolation in model prediction becomes inevitable. Notably, the toughness of composites, a property encompassing deformation from linear elasticity to final fracture through crack propagation, involves significant computational costs. This research employed hierarchical artificial neural network models to predict toughness and performed pattern optimization for grid composites through an algorithm that iterates extrapolation and model updating. The final study tackles the rib placement optimization problem for replacing traditional metal parts with reinforced plastics in automobiles for light-weighting. Reinforced plastics, being significantly lighter than metals but mechanically weaker, use rib reinforcement structures to compensate for their mechanical properties. However, as the number of ribs increases, so does the number of design variables, making the design space varying. Traditional surrogate model-based optimization cannot be performed in such cases since models cannot have varying inputs. This research utilized reinforcement learning algorithms to optimize rib patterns, accommodating the changing number of variables. Through these studies, this dissertation presents surrogate model-based optimization methodologies for three different design variable scenarios. The first scenario involves a small, continuous design space with upper and lower bounds. Here, the methodology improves the model's predictive performance through sampling training data that ensures interpolation within the design space, enabling rapid optimization. The second scenario deals with a digital and vast design space, where interpolation through the model is not guaranteed. A methodology is presented that enhances the model's extrapolation capabilities using physical input features, followed by repeated extrapolation and model updating for optimization. Lastly, for a varying design space with changing numbers of design variables, a methodology deviating from traditional surrogate model-based optimization using varying inputs is presented. It employs reinforcement learning algorithm-based optimization, allowing for efficient data use even as the dimensions of the design space increase. This dissertation aims to present efficient optimization strategies based on machine learning models for various engineering problems with diverse design variable scenarios.