Real-time design of architectural structures with differentiable mechanics and neural networks

Rafael Pastrana, Eder Medina, Isabel M. de Oliveira, Sigrid Adriaenssens, Ryan P. Adams

TL;DR

This paper introduces a novel approach to designing mechanically efficient architectural structures in real-time. By combining neural networks with a differentiable physics simulator, the authors enable rapid exploration of complex designs that are both structurally sound and aesthetically pleasing, significantly outperforming traditional methods in speed and offering comparable accuracy.

What Problem Does This Paper Address?

Designing mechanically efficient architectural structures like shells, towers, and bridges is a complex and time-consuming process. Traditional methods, such as finite element analysis and form-finding, require iterative optimization that is computationally expensive and slow, limiting design exploration. While neural networks offer potential for faster, data-driven solutions, they often struggle to guarantee mechanical integrity and require extensive fine-tuning. This paper tackles the challenge of accelerating this design process by developing a system that leverages the speed of neural networks while ensuring mechanical constraints are met, bridging the gap between pure data-driven approaches and physics-based simulations.

Methods

The core of this work is a "physics-in-the-loop" surrogate model that integrates a neural network with a differentiable mechanics simulator.

  • Neural Encoder: A neural network takes a target architectural shape as input and maps it to a "stiffness space." This space represents the internal forces and material properties of the structure.
  • Differentiable Mechanics Simulator (Force Density Method - FDM): This component acts as the "physics engine." It takes the stiffness values from the encoder and boundary conditions (like fixed points) to compute a mechanically sound shape. The FDM is chosen for its efficiency in solving for equilibrium shapes in bar systems.
  • End-to-End Training: The entire system is trained end-to-end using automatic differentiation. This allows gradients from the physics simulator to flow back to the neural encoder, enabling it to learn to produce stiffness values that result in both accurate shape approximations and mechanical integrity.
  • Amortized Inference: Unlike traditional optimization that solves for each design individually, this model amortizes the solution process. Once trained, it can generate a mechanically sound design for a new target shape in real-time.

Datasets

The paper does not explicitly mention pre-existing datasets. Instead, it describes data generation processes for two specific tasks:

  • Masonry Shells: Data is generated by sampling from a family of Bezier surfaces, specifically focusing on doubly symmetric shapes and then exploring asymmetric variations.
  • Cable-Net Towers: Data is generated by parametrizing tower geometries using ellipses for the compression ring and varying their radii and in-plane rotation.

Key Insights and Ideas

  • Physics-in-the-Loop is Crucial for Safety: Purely data-driven neural networks (NNs) and even physics-informed neural networks (PINNs) can produce shapes that look good but are mechanically unstable. The authors demonstrate that their physics-in-the-loop approach, by construction, guarantees mechanical integrity, which is paramount for safety-critical applications like structural design.
  • Amortization for Real-Time Design: The ability to amortize the design process through a trained neural network is key to achieving real-time performance. This allows designers to interactively explore a vast design space, a feat impossible with traditional iterative optimization.
  • Differentiable Simulators as Powerful Decoders: Using a differentiable mechanics simulator as the decoder in an autoencoder-like architecture provides a principled way to enforce physical constraints, offering a significant advantage over learnable decoders that might struggle to capture complex physics.
  • Generalization Beyond Training Distribution: The authors show that their model generalizes better to unseen (asymmetric) geometries compared to PINNs, even when trained exclusively on symmetric shapes. This highlights the robustness gained from embedding explicit physics.
  • Neural Initialization for Optimization: The model can provide excellent starting points for traditional optimization methods, significantly reducing the time and iterations required to reach an optimal solution.

Key Results

  • Speed: The proposed model achieves real-time inference (milliseconds), which is orders of magnitude faster than direct optimization (thousands of milliseconds).
  • Accuracy: The model achieves comparable shape accuracy to direct optimization while guaranteeing mechanical integrity. For masonry shells, it outperforms PINNs in generalization.
  • Mechanical Integrity: The model consistently produces mechanically sound designs, with physics losses close to zero, unlike NN and PINN baselines which exhibit significant residual forces.
  • Fabrication: A physical prototype of a masonry shell designed by the model was successfully fabricated and demonstrated stability, validating the real-world applicability of the approach.
  • CAD Integration: The model was integrated into Rhino3D, enabling interactive design exploration within a familiar CAD environment.

Limitations and Future Work

  • Task-Specific Models: The current model is trained for specific tasks and topologies. Generalizing to entirely new structure types or discretizations would require retraining. The authors suggest exploring graph networks to overcome this limitation.
  • Numerical Instabilities: The differentiable simulator can be ill-conditioned for certain structures, affecting training. Gradient stabilization techniques like clipping are used, but further research into handling these instabilities is needed.
  • Parameter Space Design: Task-specific parameter spaces need to be designed for learning.
  • Uniqueness of Solutions: The mapping from stiffness to geometry is not always unique. The authors suggest exploring variational settings to offer designers a diversity of possible solutions.

Summary updated

BibTeX
@inproceedings{pastrana2025real,
  author = {Rafael Pastrana and Eder Medina and Isabel M. de Oliveira and Sigrid Adriaenssens and Ryan P. Adams},
  title = {{Real-time design of architectural structures with differentiable mechanics and neural networks}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=Tpjq66xwTq},
}
Figure