Physics-informed, Generative Adversarial Design of Funicular Shells

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

Problem Statement
Shell structures are highly valued in architecture and engineering for their structural efficiency and aesthetic potential. However, the advent of 3D concrete printing (3DCP) has created a specific design challenge: 3DCP structures cannot incorporate steel reinforcement, necessitating designs that withstand primarily pure compression (funicular shells) to avoid tensile stresses. While the concept of the funicular form is well-established for 2D arches (dating back to Hooke's catenary discovery in 1675), generalizing this concept to optimal three-dimensional shell forms remains an open research topic. Traditional computational approaches, such as Thrust Network Analysis or optimization-based methods using the lower bound theorem, often rely on specific simplifications or are susceptible to converging to local minima. This work addresses the need for a generative tool capable of exploring efficient, compression-only shell geometries without assuming behavioral simplifications, while avoiding the local minima issues inherent in gradient-based optimization.

Methodology
The authors propose a physics-informed generative framework based on a modified Deep Convolutional Generative Adversarial Network (DCGAN), termed AD-DCGAN. The methodology integrates three core components:

1. Data Generation and Representation: A procedural database of 6,058 funicular shell geometries was created using Blender, simulating fabric deformation under self-weight (inspired by Isler's hanging fabric method). These geometries were converted into 2.5D level-set representations (height maps on a 64×64 grid) and processed via Finite Element Analysis (FEA) in Abaqus to extract structural fields, including membrane factor distributions ($m_f$), displacements, strains, and stress resultants.
2. Surrogate Model (PB-PUNet): A Physics-based Parallel UNet (PB-PUNet) was developed as a surrogate model to predict the structural behavior of shell geometries. It takes as input the shell height, vertical load, and a binary mask (to handle empty regions) and outputs 13 solution fields. This model was trained with physics-informed regularization (minimizing the total potential energy balance) and hyperparameter tuning (Z-score normalization, Kaiming initialization, GroupNorm, and dropout) to achieve high predictive accuracy.
3. Generative Framework (AD-DCGAN): The generator creates shell geometries and binary masks from a latent vector. To ensure physical compliance and training stability, the framework employs:
   • Auxiliary Discriminator ($D_{aux}$): The pre-trained PB-PUNet acts as a discriminator that evaluates the generated shells based on their membrane factor distribution, penalizing geometries dominated by bending.
   • Mask Discriminator ($D_{mask}$): A dedicated discriminator ensures the generator correctly depicts valid and invalid regions (e.g., central holes).
   • PCA-guided Latent Space: Instead of purely random noise, the generator input is guided by the top $k$ principal components of real data, augmented with random noise, to mitigate mode collapse.
   • Stabilization Techniques: The architecture incorporates feature matching and spectral normalization to address training instability and mode collapse.

Key Contributions

• Physics-Informed GAN Architecture: The paper introduces a novel AD-DCGAN where a pre-trained PINN (Physics-Informed Neural Network) surrogate model serves as an auxiliary discriminator. This enforces physical consistency (specifically, a high membrane factor) directly within the generative process, rather than relying solely on post-hoc optimization.
• Handling Complex Geometries: The inclusion of a mask discriminator allows the model to generate complex shell forms containing empty areas (e.g., central holes), a capability often limited in standard image-based GANs.
• Mode Collapse Mitigation: By combining a PCA-guided latent space with feature matching and spectral normalization, the framework successfully generates diverse geometries without collapsing to a single mode.
• Surrogate Model Accuracy: The PB-PUNet surrogate model achieves a total Relative Root Mean Square Error (RRMSE) as low as 2.2% on the test set, validating its utility as a reliable physics evaluator for the GAN.

Results

• Surrogate Performance: The optimized PB-PUNet model demonstrated high predictive accuracy, with the membrane factor ($m_f$) showing a 5.566% error and total error at 2.198% on the test set. The model effectively captures the spatial distribution of structural fields.
• Generative Diversity: The AD-DCGAN successfully generated diverse, previously unseen funicular shell geometries. Analysis of the latent manifold using Kernel PCA showed that generated samples are continuously distributed across the feature space, confirming the absence of mode collapse.
• Physical Compliance: The auxiliary discriminator successfully guided the generation toward geometries with high membrane factors. Histograms of generated samples showed a distribution skewed toward higher $m_f$ values compared to the training data, indicating a preference for membrane-dominated behavior.
• FEA Validation: Selected generated geometries were validated via FEA. While the generated surfaces required post-processing (radial basis function interpolation) to remove noise for meshing, the resulting structures exhibited global behaviors consistent with funicular forms. However, the paper notes discrepancies between the PB-PUNet predictions and FEA results for generated samples, attributing this to the generator's ability to create novel geometries that may cause the surrogate model to extrapolate beyond its training distribution.

Significance and Claims
The paper claims to offer a stable, AI-driven methodology for the generative design of compression-only shell structures that does not rely on the simplifications often required by traditional optimization methods. By integrating a physics-informed surrogate model directly into the adversarial training loop, the framework ensures that generated designs are not only geometrically realistic but also structurally efficient (high membrane factor). The authors position this work as a step toward overcoming the challenge of generalizing the funicular concept to 3D structures, particularly for applications in 3D concrete printing where tensile reinforcement is impractical. The study concludes that the developed model is capable of producing physically optimal, diverse shell forms, though it acknowledges that further refinement is needed to handle the extrapolation of surrogate models when generating entirely novel geometries.