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

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

Abstract

Designing efficient architectural structures like shells and towers requires a delicate balance between geometric form and mechanical stability. Traditional optimization methods are accurate but slow, while pure neural networks are fast but often produce physically unsound designs. This paper introduces a "physics-in-the-loop" hybrid model that integrates a neural network encoder with a differentiable mechanics simulator decoder. This architecture guarantees that every generated design is in mechanical equilibrium. The model achieves inference speeds up to 10,000 times faster than traditional methods, enabling real-time design exploration without compromising structural integrity.

Introduction

The design of long-span structures, such as masonry shells and cable-net towers, is governed by the principle of "form-finding"—the process of determining a geometry where internal forces are perfectly balanced. Historically, engineers have faced a trade-off between the slow, iterative accuracy of traditional optimization and the fast but unreliable predictions of standard deep learning models. While Physics-Informed Neural Networks (PINNs) attempt to bridge this gap by penalizing unphysical outputs, they cannot provide hard guarantees of mechanical safety, particularly for out-of-distribution geometries.

Methods: Physics-in-the-Loop Architecture

The core innovation is a hybrid autoencoder-like architecture where the decoder is replaced by a differentiable mechanics simulator.

Overview of the hybrid architecture: a neural network encoder maps target shapes to stiffness parameters, which a differentiable simulator decoder then transforms into a physics-constrained structural design.
  1. Encoder (Neural Network): A Multi-Layer Perceptron (MLP) processes a designer's target shape $\hat{\mathbf{X}}$ and maps it to a latent space representing force densities or stiffnesses $\mathbf{q}$.
  2. Decoder (Differentiable Simulator): Instead of learning to reconstruct the shape, the decoder uses the Force Density Method (FDM) to solve a linear system of equilibrium equations. This ensures that the output $\mathbf{X}$ is always a physically valid structure.

Because the simulator is differentiable, the entire pipeline can be trained using standard backpropagation. The physics is a hard constraint embedded within the network's forward pass, ensuring that even "bad" predictions from the encoder result in stable, albeit potentially non-matching, structures.

Results and Discussion

The model was tested on masonry shells and complex cable-net towers. A key finding was the model's ability to generalize to asymmetric "out-of-distribution" shapes despite being trained only on symmetric data. This is attributed to the simulator's inherent knowledge of universal physical laws.

Traditional iterative optimization process for form-finding, which the proposed neural network model replaces with a millisecond-scale 'one-shot' inference.

Performance Metrics

  • Inference Speed: The model generates designs in 0.6 to 4.6 milliseconds, compared to over 5.5 seconds for traditional gradient-based optimization—a speedup of up to four orders of magnitude.
  • Mechanical Soundness: The "Physics Loss" (residual force) for the proposed model was effectively zero, whereas pure neural networks and PINNs produced structures with significant residual forces that would lead to collapse.
  • Real-World Validation: The system was integrated into Rhino3D/Grasshopper, allowing architects to manipulate shapes in real-time while receiving immediate, physically sound feedback. The researchers validated the approach by 3D-printing a masonry shell that remains stable without any adhesives.

Conclusion

By embedding a differentiable simulator directly into the neural network architecture, this research provides a robust framework for safety-critical engineering design. Future work will explore Graph Neural Networks (GNNs) to handle varying structural topologies and improve numerical stability for highly complex cable systems.

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 simulators and neural networks}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=Tpjq66xwTq},
}
Figure