An Alternative Approach to Automated Code Checking – Application of Graph Neural Networks Trained on Synthetic Data for an Accessibility Check Case Study

Automated Code Checking (ACC) can be defined as a classification task aiming to classify building objects as compliant or not compliant to a code provision at hand. While Machine Learning (ML) is a useful tool to perform such classification tasks, it presents several drawbacks and limitations. Buildings are complex compositions of instances that are related to each other by functional and topological relationships. This type of data can be easily supported by property graphs that provide a flexible representation of attributes for every instance as well as the relationships between the instances. This, together with the recent developments in the field of graph-based learning led the authors to explore a novel approach for ACC supported by Graph Neural Networks (GNN). This paper presents a new workflow that implements GNNs for ACC to leverage the advantages of ML but alleviate the limitations. We illustrate the suggested workflow by training a GNN model on a synthetic data set and using the trained classifier to check compliance of a real BIM model to accessibility requirements. The accuracy of the classifier on a test set is 86% and the accuracy of obtained results during the accessibility check is 82%. This suggests that GNNs are applicable to ACC and that classifiers trained on synthetic data can be used to classify building design provided by the industry. While the results are encouraging, they also point to the need for further research to establish the scope and boundary conditions of applying GNNs to ACC.