⚡ electrical engineering

Physics-informed, Generative Adversarial Design of Funicular Shells

This paper proposes a physics-informed Generative Adversarial Network framework that utilizes a membrane factor constraint and an auxiliary discriminator to successfully generate stable, structurally efficient, and previously unseen funicular shell geometries, thereby generalizing the concept of funicular polygons to three-dimensional structures.

Original authors: Ruben Lourenço, Iciar Alfaro, Beatriz Moya, Elias Cueto

Published 2026-09-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruben Lourenço, Iciar Alfaro, Beatriz Moya, Elias Cueto

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

For centuries, architects and engineers have chased a specific kind of structural perfection: a building that stands up using only its own shape to hold weight, without needing heavy internal supports or steel bars to stop it from cracking. This idea dates back to the 17th century, when scientists realized that if you hang a chain loosely between two points, it forms a curve that is under pure tension. If you flip that hanging chain upside down, the resulting arch is under pure compression, meaning every part of it is being squeezed together rather than pulled apart. This is the most efficient way to build with materials like stone or concrete, which are very strong when squeezed but weak when pulled. While this principle works beautifully for simple arches, extending it to complex, three-dimensional shell structures—like the curved roofs of modern buildings—has remained a difficult puzzle. Designing these shapes usually requires complex computer simulations that take a long time to run, and finding the perfect form often involves guessing and checking until the structure is stable.

A team of researchers from Spain and France has developed a new way to solve this puzzle using a type of artificial intelligence that learns by playing a game against itself. Instead of running slow, traditional simulations for every new design, they trained a computer system to invent new shell shapes that are naturally efficient. The researchers created a database of thousands of different shell forms, simulating how they would behave under their own weight. They then taught an artificial intelligence model to recognize the difference between a shape that is structurally sound and one that is not. The system uses a "generator" to create new, unseen shell designs and a "discriminator" to judge them. Crucially, the discriminator is guided by physics; it does not just look for realistic-looking images, but checks if the shape would actually hold up in the real world by calculating how much of the structure is being squeezed versus how much is being bent.

The results show that this method can successfully generate smooth, complex shell geometries that are dominated by compression, which is the ideal state for concrete structures. In their tests, the computer produced shapes where the "membrane factor"—a measure of how much the structure relies on squeezing forces rather than bending forces—was very high, often close to the maximum possible value. This suggests the AI is learning to create forms that would not require steel reinforcement, a significant advantage for new construction techniques like 3D printing concrete, where adding steel is difficult. The researchers found that the system could create a wide variety of shapes, including some with holes in the middle, without collapsing into a single, repetitive pattern. When they took the computer-generated shapes and ran them through standard engineering software to double-check the physics, the results confirmed that the AI was producing viable, efficient structures.

The study also highlights the importance of how the computer learns. The researchers found that simply letting the AI guess randomly did not work well; the system needed a specific guide to understand the rules of physics. By using a pre-trained model that acts as a physics expert, the system learned to avoid designs that would fail. They also discovered that the AI needed to see the "holes" in the design clearly, so they taught it to generate a map of where the material exists and where it does not. This allowed the system to handle complex designs with empty spaces, which are common in real architecture. The final output was a collection of smooth, flowing surfaces that looked like they were grown rather than built, all while adhering to the strict laws of mechanics.

This work does not claim to have solved every problem in structural design, but it demonstrates a powerful new tool for exploring the possibilities of shell architecture. The researchers showed that by combining artificial intelligence with the fundamental laws of physics, it is possible to generate structural forms that are both beautiful and efficient. The system produced designs that were previously unseen, proving that the computer can explore the space of possible shapes more broadly than traditional methods. While the generated shapes still require a final check by human engineers to ensure they are ready for construction, the study confirms that this approach can rapidly produce high-quality candidates for new, sustainable building forms. The ability to generate these shapes quickly could help architects and engineers design buildings that use less material and require less energy to construct, bringing the efficiency of the hanging chain into the complex world of modern 3D structures.

Technical Summary: Physics-informed, Generative Adversarial Design of Funicular Shells

Problem Statement
The design of funicular shell structures—geometries that carry loads primarily through membrane (compressive) forces rather than bending—is a longstanding challenge in structural engineering. While the principle of the funicular form is well-established for two-dimensional arches (dating back to Hooke's 1675 discovery of the catenary), generalizing this concept to complex three-dimensional shell structures remains an open research topic. This is particularly critical for emerging construction methods like 3D concrete printing (3DCP), which cannot easily incorporate steel reinforcement and therefore require structures that avoid tensile stresses. Traditional design approaches often rely on optimization algorithms or Finite Element Analysis (FEA), which can be computationally expensive or prone to converging on local minima. Furthermore, existing AI-driven methods often lack strict physical constraints, potentially generating geometrically plausible but structurally inefficient designs.

Methodology
The authors propose a physics-informed generative framework, termed AD-DCGAN (Auxiliary Discriminator Deep Convolutional Generative Adversarial Network), to automate the design of compression-only shell geometries. The methodology consists of three primary components:

Data Generation and Physics Simulation: A procedural tool generates a dataset of 6,058 funicular shell geometries using a physics-based simulation (inspired by Hooke's hanging chain analogy) in Blender. These geometries are subsequently analyzed using FEA (Abaqus/Standard) under self-weight loading to extract structural response variables, including displacements, strains, stress resultants, and bending moments.
Physics-Based Parallel UNet (PB-PUNet): A surrogate model based on a U-Net architecture is trained to predict the structural response of a given shell geometry. This model takes a three-channel input (height map, vertical load, and a binary mask for holes) and outputs thirteen channels representing the solution fields (displacements, strains, moments, etc.). Crucially, this model is used to calculate the membrane factor ($mf$), a metric defined as the ratio of membrane work to total work ( $W_{memb} / (W_{memb} + W_{flex})$ ). A value close to 1.0 indicates a structure dominated by membrane action, which is the design goal.
AD-DCGAN Architecture: The core generative model is a modified DCGAN featuring:
- Generator (G): Converts a latent vector (guided by Principal Component Analysis of real data plus random noise) into a 64x64 pixel image representing the shell height and a binary mask for holes.
- Discriminators: The system employs three discriminators:
  - $D_{shell}$ : Distinguishes between real and generated shell images.
  - $D_{mask}$ : Evaluates the validity of the generated binary masks (handling complex geometries with holes).
  - $D_{aux}$ (Auxiliary Discriminator): A pre-trained PB-PUNet model that evaluates the physical compliance of the generated shells by predicting their membrane factor distribution. This acts as a physics-informed constraint, penalizing geometries that do not exhibit high membrane factors.

The training process utilizes feature matching and spectral normalization to stabilize the adversarial game and prevent mode collapse. The generator is optimized to fool the discriminators while maximizing the membrane factor predicted by the auxiliary discriminator.

Key Contributions

Physics-Informed GAN Framework: The paper introduces a novel architecture where a pre-trained physics surrogate model (PB-PUNet) serves as an auxiliary discriminator. This enforces physical consistency (specifically, high membrane efficiency) directly within the generative process, rather than relying solely on post-hoc optimization.
Handling Complex Geometries: By incorporating a dedicated mask discriminator and a binary mask output channel, the model successfully generates shells with central holes and complex topologies, addressing a limitation of standard image-based generative models.
PCA-Guided Latent Space: The use of PCA to guide the generator's input latent vector, rather than relying purely on random noise, helps mitigate mode collapse and reduces computational complexity.
Surrogate Modeling: The development and rigorous hyperparameter tuning of the PB-PUNet model, which achieves high predictive accuracy (total RRMSE of ~2.2%) for shell structural responses, serves as a robust foundation for the physics-informed guidance.

Results

Surrogate Model Performance: The optimized PB-PUNet model demonstrated excellent predictive capability on the test set, with a total RRMSE of 2.198%. The membrane factor ($mf$) prediction showed a relative error of 5.566%, while vertical displacement errors were higher (10.344%) but still within acceptable bounds for a surrogate model.
Generative Performance: The AD-DCGAN successfully generated diverse shell geometries, including those with and without central holes. Analysis of the latent manifold using Kernel PCA revealed that the generated samples were continuously distributed and did not suffer from mode collapse.
Physical Compliance: The auxiliary discriminator successfully biased the generation toward high membrane factor structures. Histograms of the generated samples showed a distribution skewed toward higher $mf$ values compared to the training data.
FEA Validation: When selected generated surfaces were re-analyzed using standard FEA, the global behavior was well-reproduced. However, the paper notes discrepancies in local membrane factor distributions, attributing this to the "noisy" nature of raw GAN outputs and the surrogate model's difficulty in extrapolating to completely novel geometries.

Significance and Claims
The authors claim that this work provides a stable and effective tool for structural designers to explore efficient, funicular shell forms without assuming simplifications on structural behavior. The primary significance lies in demonstrating that a GAN can be guided by physical laws (via the membrane factor) to generate physically optimal, previously unseen, funicular shells with smooth forms and high membrane factor distributions. The paper modestly acknowledges limitations, noting that while the global behavior is captured, the raw output requires post-processing (smoothing) for FEA validation, and the surrogate model may struggle with extrapolation to entirely new physical regimes. The work positions itself as a step toward overcoming the limitations of traditional optimization-based approaches and enabling the design of complex, compression-only structures suitable for 3D concrete printing.

More like this