Geometric Reasoning and Machine Learning: A Set of Design and Manufacturing Problems
MetadataShow full item record
Research in engineering design and manufacturing domain has enabled significant advances in the last couple of decades. Now, more than ever before, many individual manual tasks are being automated to assist designers and fabricators at different stages of product lifecycle. The possibility of automating previously manually carried out tasks is being aided by the ongoing advancement of computational algorithms and ever-increasing computational power. Although the current achievement and progress in automating the design and manufacturing process are commendable, it still has a long way to go. There are a variety of design and manufacturing tasks that are still carried out manually thus leading to tedious repetitions, resources wastage, and unexplored and suboptimal solutions. In this dissertation, I research the innovative usage of geometric reasoning and machine learning techniques (probabilistic models and deep reinforcement learning) to create computational tools that facilitate designers and manufacturers in carrying out several distinct design and manufacturing tasks. Each chapter in the dissertation is targeted at solving a distinct type of problem. Hence, the solution approach for each problem is different but utilizes the common core of geometric reasoning and machine learning. Specifically, I focus on the following main categories of problems: corrective, suggestive, and generative. A brief summary of the problems and the solution approaches is provided below.Corrective:The recent advancement of additive manufacturing technology has allowed for fabrication of highly complex parts that were difficult to produce using traditional manufacturing techniques. However, designers and novice users of additive technologies often lack the awareness of manufacturing considerations leading to lower quality parts and failures. Therefore, it is crucial to develop an automated system that can analyze and correct designs before they are sent for fabrication so that wasteful iterations of build-test-redesign can be minimized. In this work, a variation of shape diameter function and morphological operations are used to identify critical regions in 2D slices of 3D models that can potentially cause errors in printing. Printability of slices has been quantified using an area-based printability index. In addition, a physics-based mesh deformation scheme is adopted to make localized corrections to slices for improving printability. The approach has also been extended to three-dimensions for correcting critically thin regions of 3D models.Suggestive:A novel and intuitive assembly based 3D modeling interface to support conceptual design exploration activities is developed in this work. The modeling interface uses unlabeled segmented components of the objects and allows their assembly to create new 3D models. The development of the interface is motivated by two aspects. First, the focus is on novice users since they stand to gain the most from intuitive interfaces. Second, the intent is on creative reuse of a growing number and variety of 3D models available on vast online repositories like Turbosquid and Trimble 3D warehouse. Specifically, an automated component suggestion algorithm was devised that is based on a probabilistic factor graph. This algorithm helps the user to easily browse and select components from a database that are most compatible with the current state of the 3D model being assembled. The component suggestion algorithm incorporates various aspects such as shape similarity, repetitions of shapes, and adjacency relationships. The suggestive interface overcomes several limitations of traditional CAD interfaces by helping the users to quickly create and explore new conceptual designs.Generative:The advancement of manufacturing techniques, such as additive manufacturing, has allowed fabrication of complex and intricate structural designs that result in materials, called metamaterials, with effective properties significantly different from its constituent materials. Metamaterials can exhibit smart and unusual properties that prove to be useful in a variety of applications, such as photonics, acoustics, seismic protection, heat exchangers, absorbers, and lightweight structures. The metamaterial design is a non-trivial and challenging problem that is usually tackled by restricting to periodic structures and trial-and-error method for parameter selection. The explosion of solution space for aperiodic structures renders the inverse problem, i.e. generating optimal aperiodic metamaterial for satisfying a desired functional requirement, intractable. The use of deep reinforcement learning to obtain desired behavior from 2D and 3D metamaterial structures is proposed in this work. A deep policy gradient actor-critic model is developed and trained by repeated simulation of metamaterial behavior and tweaking its parameters as model actions. The trained model is then used for generating new aperiodic structures for any desired behavior.