From reconstruction and generation to simulation and design, how machines model three-dimensional content is being steadily transformed by deep learning. Implicit Neural Representations (INRs), which describe a shape as the level set of a learned continuous field, are resolution-independent, differentiable end-to-end, and compact enough to act as learned shape parametrizations. Engineering design, however, makes sharp demands: shapes are not only observed but created, analyzed, and optimized; they form multi-part assemblies governed by explicit design parameters. Standard INRs, while expressive, fall short on these fronts, and this thesis develops structured implicit neural representations to address this gap.

We first tackle the lack of geometric regularity in pure neural representations: they fit free-form surfaces accurately but miss features prevalent in manufactured objects, e.g., cylinders or planar faces, and expose no parametric handles for direct edits. We propose a hybrid representation that unifies three primitive types: a fully implicit one for free-form parts, an analytic one with explicit parameters (radius, height, pose), and a geometry-assisted neural mixing the two. A disentangled decoder separates latent from explicit parameters, and the full shape is their union. This reconstructs complex and exactly regular parts together, supports parametric edits without re-optimization, and recovers geometric parameters from images and sketches.

Engineering objects are also inherently part-based, yet most implicit representations treat them as monolithic; part-aware alternatives model parts in isolation, output only a holistic shape, or rely on unsupervised discovery without semantic correspondence. We propose a supervised composite representation where each part has its own latent vector and pose, processed by an auto-decoder that alternates single-part and cross-part layers to share information across components. A region-based supervision scheme derives per-part signals from the global shape, without requiring watertight per-part meshes. With a single frozen decoder, we attain the best reconstruction and generation among part-based baselines, and uniquely support constrained optimization, e.g., drag reduction on a car body with fixed wheels.