This research explores the optimization of modular ribbed concrete floor systems as a strategy for circular construction. Ribbed floors can reduce material use and embodied carbon compared to flat slabs, but conventional customized solutions lack scalability while standardized waffle slabs are less efficient. To address this trade-off, a modular design approach is combined with a Deep Generative Design workflow using a Variational Autoencoder (VAE) and Gradient Descent (GD).

The discrete modular configurations are represented as bitmaps, decoupling geometry from structural performance and enabling efficient training of a simple VAE model. Trained on datasets generated in Grasshopper with Karamba3D, the VAE can predict and generate new designs while scaling to larger problem sizes. Optimization is performed by sampling in latent space and refining results with GD.

The workflow flexibly integrates new objectives and constraints without retraining, such as stock availability or embodied carbon. In benchmark tests, the VAE-based optimization outperformed a generic evolutionary solver, achieving lower elastic energy and better stock compliance. The approach demonstrates the potential of deep generative methods for scalable, constraint-aware optimization of modular ribbed floor systems, while challenges remain for extending to more complex structural models.