Neural Network-based Topology Optimization for Manufacturing

<p dir="ltr">Subtractive and additive are two key technologies in manufacturing. Additive manufacturing has emerged as a powerful method for producing complex components. However, several often overlooked challenges around the key issues in AM, such as part removal from the build plate, support structure removal, and part post-processing, make AM highly cost and time-prohibitive for large-scale manufacturing. Subtractive manufacturing can also benefit from topology-optimized design. However, the complex, curved surfaces of topology-optimized designs typically necessitate long machining times using point milling for a faithful production of the part. Computational part design algorithms that focus solely on engineering requirements, such as weight and compliance, often overlook the cost associated with machining time and post-processing, thereby further hampering a wide-scale adoption of topology optimization. </p><p dir="ltr">In this thesis, we propose new criteria and associated algorithms designed to facilitate a more comprehensive modeling of part production. A key advance is our adoption of neural networks as a means to integrate manufacturing with multiple post-processing and shape objectives, enabling our computational topology design algorithms to optimize for these criteria jointly. Unlike conventional topology optimization, our approach computes analytical density gradients from the neural networks, enabling the creation of optimal support structures with minimal overhangs for additive manufacturing. Furthermore, we compute analytical curvature information to create developable surfaces to minimize machining time. We also demonstrate how neural networks seamlessly allow for user-controllable input dimension extension, convergence speed improvement, and a very compact and time-efficient modeling of multi-scale topology optimization by incorporating local and global spatial coordinates for shape design that can be manufactured with both subtractive and additive methods. </p><p dir="ltr">Toward these ideas, this thesis addresses manufacturing efficiency challenges by developing a tailored neural network-based topology optimization. Our approach focuses on three main objectives: (1) optimizing support structures and topology for post-processing in additive manufacturing, (2) concurrent topology and Gaussian curvature minimization with seamline feature scale control for efficient subtractive manufacturing with flank milling, and (3) efficient large multiscale and metamaterial design that can be manufactured with both additive and subtractive manufacturing.</p>