Bayesian Inference for Modelling Uncertainty in Non-standard Building Systems

Fabian Kannenberg, Marta Gil Pérez, Tim Schneider, Steffen Staab, Jan Knippers, Achim MengesOpenalex Unknown 2024 · DOI · OpenAlex

Introduction

The construction industry is increasingly adopting "non-standard" systems, such as robotically wound carbon fiber and 3D-printed concrete. Unlike traditional materials like steel, these systems lack decades of standardized building codes. In processes like Coreless Filament Winding (CFW)—where fibers are wound around minimal anchors to create lightweight structures—the final product is highly variable due to factors like fiber tension, slippage, and resin viscosity. This inherent "messiness" makes it difficult to predict structural performance, often leading to material-wasting over-engineering.

Methodology: Probabilistic Graphical Models

To address this, researchers propose a Bayesian Inference framework using Probabilistic Graphical Models (PGMs), specifically Bayesian Networks. This approach allows for the quantification and management of uncertainty even when data is scarce or sensors are limited.

Directed Acyclic Graph (DAG) mapping causal relationships in the CFW system.

The core of the methodology involves Directed Acyclic Graphs (DAGs) that map causal relationships. For instance, winding tension influences the cross-sectional area, which subsequently affects buckling displacement. By using Generalized Linear Models (GLMs), the researchers "train" the nodes to understand how parent variables affect child variables. This structure allows for "active trail" reasoning: if a sensor detects high strain in one area, the model can back-calculate likely material properties and predict risks in unmonitored segments.

Experimental Setup and Data

The study utilized a custom dataset derived from six physical coreless filament wound fiber nets. Each specimen consisted of 12 distinct fiber segments, providing a granular look at the fabrication process.

Physical fiber net specimen (left) and robotic coreless filament winding process (right).

The dataset captured fabrication parameters (tension, time), geometry (length, curvature), material properties (modulus of elasticity), and structural response metrics (strain, buckling displacement). The researchers compared different hierarchical modeling strategies, finding that "partial pooling"—where segments have individual distributions constrained by the group—offered a balanced middle ground between accuracy and generalizability.

Key Insights and Results

A major advantage of PGMs over "black box" machine learning is the ability to incorporate expert engineering knowledge—such as the laws of physics—directly into the graph structure. This significantly reduces the amount of data required to achieve reliable results.

The results demonstrated high precision: the "unpooled" model (treating segments individually) achieved a Root Mean Square Error (RMSE) of 0.019 for predicting buckling behavior. However, the researchers found that generalizing through clustering is difficult; when segments were grouped into five clusters, the error rose to 0.33, highlighting the unique nature of non-standard components.

Conclusion and Future Directions

The Bayesian approach effectively "learns" the physics of complex robotic processes, facilitating safer "co-design" between architects and engineers. Future work aims to improve topological consistency, as robotic errors can change the structural logic of a net. Additionally, the researchers plan to fully integrate Life Cycle Assessment (LCA) data into the model, enabling real-time optimization for both structural integrity and the lowest possible carbon footprint.

Summary updated

BibTeX
@inproceedings{kannenberg2024bayesian,
  author = {Fabian Kannenberg and Marta Gil Pérez and Tim Schneider and Steffen Staab and Jan Knippers and Achim Menges},
  title = {{Bayesian Inference for Modelling Uncertainty in Non-standard Building Systems}},
  booktitle = {Openalex Unknown},
  year = {2024},
  doi = {10.1007/978-3-031-68275-9\_6},
}
Figure