Learning to Simulate and Design for Structural Engineering

Kai-Hung Chang, Chin-Yi Cheng

Structural engineering is a critical bottleneck in the modern construction industry. When designing a building, engineers must select specific cross-sections for hundreds of columns and beams. This design process must balance two competing goals: minimizing material mass to save costs and reduce carbon footprints, while strictly adhering to safety codes—such as limiting "drift" (how much a building sways under wind or seismic loads).

The challenge is massive. Even a simple five-story building contains over 500 design variables. Furthermore, traditional Finite Element Analysis (FEA) tools used to simulate these designs are computationally expensive, often taking minutes to evaluate a single configuration. Consequently, engineers are forced to rely on manual trial-and-error, which frequently results in "over-designed" buildings that satisfy safety codes but waste significant amounts of steel and concrete.

To solve this, researchers at Autodesk have developed an end-to-end deep learning pipeline that automates the structural design of buildings. By combining a graph-based "Sizer" with a differentiable "Simulator," the system can propose optimal beam and column sizes in milliseconds.

The Core Method: Differentiable Design

The researchers represent building skeletons as graphs, where each structural bar (column or beam) is a node and the joints connecting them are edges. They introduce two specialized Graph Neural Networks (GNNs) to handle the design and simulation phases:

  1. NeuralSizer (The Designer): This model learns to assign the optimal cross-section sizes to every bar in the building's skeleton.
  2. NeuralSim (The Approximator): Acting as a surrogate for traditional simulation software, this model takes a building's geometry and member sizes and predicts the "drift ratio" (sway).
The end-to-end differentiable pipeline combining NeuralSizer and NeuralSim.

The breakthrough of this approach lies in its end-to-end differentiability. Because NeuralSim is a neural network, the "loss" (errors in safety or excess mass) can flow backward through the simulator as a gradient signal. This tells the NeuralSizer exactly how to adjust its design to improve, creating a fully automated, trainable loop that removes the human bottleneck from the optimization process.

Physics-Informed Neural Simulation

Rather than using a generic GNN, the authors designed NeuralSim with a physics-informed architecture. They introduced a Structured Decoder (SD) that explicitly accounts for the physical reality that lower floors must support the weight and sway of everything above them.

The Structured Decoder mimics the hierarchical nature of physics, calculating sway from the roof down.

By calculating sway hierarchically from the roof down to the ground, this "inductive bias" significantly improves the model's accuracy over standard GNNs. To prevent the model from "cheating" by picking a single massive size for every beam (which is safe but highly wasteful), the researchers also implemented an entropy constraint to encourage a realistic variety of member sizes.

Training and Datasets

Because real-world structural design data is scarce and highly proprietary, the authors synthesized a custom dataset of 4,000 building skeletons with randomly sampled cross-sections at a real-world scale. They used Autodesk Robot Structural Simulator (RSA)—a professional-grade industry tool—to compute the "true" physics for these buildings, creating a robust ground truth to train NeuralSim. The dataset was split into 3,200 buildings for training, 400 for validation, and 400 for testing.

Key Results and Impact

The performance of this neural pipeline represents a massive leap over traditional methods:

  • Blazing Speed: NeuralSim is 1,900x faster than the industry-standard RSA software, completing a simulation in just 6.7 milliseconds compared to 13 seconds, while maintaining an impressive 97.36% accuracy.
  • Optimization Efficiency: Finding an optimal design using a Genetic Algorithm (GA) with traditional tools would typically take roughly two weeks. The NeuralSizer/NeuralSim pipeline generates a comparable design in just 30 minutes.
  • Human-Like Logic: In user studies, professional structural engineers noted that NeuralSizer’s designs looked "natural." The model correctly learned underlying engineering principles without explicit programming—such as placing stronger, heavier columns at the base of the building where loads are highest.

Limitations and Future Work

While groundbreaking, the current system has a few limitations. The model currently only decides the size of the bars; it cannot yet propose the layout (topology) of the skeleton itself. Additionally, the graph representation focuses exclusively on beams and columns, omitting other critical structural elements like shear walls or floor panels. Finally, because building codes vary wildly by country and city, future versions will need to incorporate local regulations as part of the input vector to be truly field-ready.

By proving that Graph Neural Networks can master the complex interplay of physics and optimization, this research opens the door to a future where buildings are designed to be lighter, cheaper, and significantly more sustainable.

Summary updated

Abstract

The structural design process for buildings is time-consuming and laborious. To automate this process, structural engineers combine optimization methods with simulation tools to find an optimal des...
BibTeX
@inproceedings{chang2020learning,
  author = {Kai-Hung Chang and Chin-Yi Cheng},
  title = {{Learning to Simulate and Design for Structural Engineering}},
  booktitle = {International Conference on Machine Learning},
  year = {2020},
  url = {https://openreview.net/forum?id=MNhxSWjBd5},
}
Figure