Humans exhibit a remarkable ability to design new tools and systems for solving problems. At the heart of this capability lies a rich suite of evolved pattern recognition modules, enabling the reuse and modification of previously existing solutions. Machine learning offers the intriguing possibility of automatically capturing the patterns that arise in various design domains, in particular via generative modeling. By leveraging the general tool of distribution approximation over data, we may hope to computationally encode common design patterns, leading to applications that enable more efficient workflows in engineering, design, and even scientific discovery. But this promising route faces challenges. Integral to the above fields are discrete domains, such as molecules and graphs. In contrast to the Euclidean spaces where generative modeling has recently thrived (e.g., images), discrete domains often exhibit irregular structure, strict validity constraints, and non-canonical representations, presenting unique challenges to modeling. In addition, machine learning as applied to engineering still lacks maturity, and as a result there is a need for curated datasets, benchmarks, and shared pipelines. First, we discuss a novel generative model for discrete domains, known as reversible inductive construction. Building off of generative interpretations of denoising autoencoders, the model employs a Markov chain where transitions are restricted to a set of local operations that both preserve validity and avoid marginalization over a potentially large space of construction histories. Next, we introduce SketchGraphs, a dataset and processing pipeline that contains millions of real-world computer-aided design (CAD) sketches paired with their original geometric constraint graphs. Unlike previously available CAD data, explicit supervision is provided regarding the designer-imposed geometric relationships between primitives, making the typically latent construction operations directly accessible. Lastly, we present Vitruvion, a model that is trained to autoregressively generate CAD sketches as a sequence of primitives and constraints. Generations from the model may be directly imported, solved, and edited in standard CAD software according to downstream design tasks. In addition, we condition the model on various contexts, including partial sketches (primers) and images of hand-drawn sketches. Evaluation demonstrates the potential for this approach to aid the mechanical design workflow.
@phdthesis{seff2021thesis,
year = {2021},
author = {Seff, Ari},
title = {Learning-Aided Design with Structured Generative Modeling},
month = oct,
school = {Princeton University},
address = {Princeton, NJ}
}