Structured 3D Shape Optimization with Part-Based Implicit Neural Representations

Part-based implicit representation. (Left) An auto-decoder network fθ decodes latent vectors and poses into the parts of a shape. (Right) It can be used to generate new shapes given a coarse part configuration (a), to edit individual parts (b), and do part-based optimization (c). (Taken from [1].)

Modern engineering design requires structured, modular, and optimizable 3D shape representations to facilitate automated design and optimization. Traditional methods struggle to represent multi-part objects in a way that ensures design flexibility, structural integrity, and efficient performance optimization. Our research focuses on part-based neural representations, which enables parameterization, generation, and optimization of 3D assemblies. By leveraging deep implicit functions, this approach allows for modular shape optimization, making it particularly relevant for design and engineering applications. 

Objective

This project aims to extend the capabilities of a part-based representation [1] to tackle engineering-driven shape representation and optimization under constraints. Key challenges include:

  1. Structured 3D Representation for Design: Developing methods to represent detailed multi-part objects in a way that maintains the overall shape’s consistency and learning better part-aware priors that enable realistic and meaningful optimizations without breaking the object integrity.
  2. Optimization and Constraint Handling: Enforcing physical and design constraints while optimizing the shape for, e.g., aerodynamics or structural efficiency, and test across different engineering applications.
  3. Scalable Learning and Generalization: Designing more efficient architecture that scale to complex many-part objects and are able to learn larger datasets, while ensuring a high variety of realistic shape generation, without invalid outputs.

Prerequisites

  • Proficiency with PyTorch and Python.
  • Solid background in mathematics or physics.
  • Experience with 3D data (point clouds, implicit representations, CAD) would be a plus.
  • Familiarity with computer assisted engineering would be a plus.

References

[1] Talabot, N., Clerc, O., Demirtas, A. C., Oner, D., & Fua, P. (2025). PartSDF: Part-Based Implicit Neural Representation for Composite 3D Shape Parametrization and Optimization. In arXiv Preprint, 2025.