1 / 20 2 / 26 / 13 Assessing implicit knowledge in BIM models with machine learning Thomas Krijnen Technical University Eindhoven, NL Martin Tamke Centre for Information Technology and Architecture, DK 

2 / 20 2 / 26 / 13 Background Building Information Modelling (BIM): Recent paradigm shift in the construction industry. Buildings structured as semantically rich models, with relational semantics and (parametric) geometry. IFC is an open standard to exchange such building models. Models are being used for design coordination and automatic (but imperative and hard coded) building code and requirement checking. Background 

3 / 20 2 / 26 / 13 DURAARK Durable Architectural Knowledge EU funded project started in 2013 Concerns the long-term digital preservation and semantic enrichment of building models Context 

4 / 20 2 / 26 / 13 Aims Offer insights into implicit information stored in building models As an alternative to explicit querying and rule checking mechanisms. Apply both supervised and unsupervised techniques. Aims 

No content

6 / 20 2 / 26 / 13 Unsupervised learning Outlier detection based on geometrical descriptors of building elements segmented by element type {wall, window, column, …}. Unsupervised learning 

7 / 20 2 / 26 / 13 Example model Unsupervised learning 

8 / 20 2 / 26 / 13 Wall geometries Unsupervised learning 

9 / 20 2 / 26 / 13 Anomalies Unsupervised learning 

10 / 20 2 / 26 / 13 Results The algorithm has been able to identify the geometrical essence of building elements. Elements that deviate from this definition are identified and turned out to be misclassified by the authoring tool. Unsupervised learning 

11 / 20 2 / 26 / 13 Supervised learning Apply neural networks to categorize floor plans according to their use {non-residential, residential}. Supervised learning 

12 / 20 2 / 26 / 13 Features Supervised learning 

13 / 20 2 / 26 / 13 Space – wall - ratio Supervised learning 

14 / 20 2 / 26 / 13 Space – slab ratio Supervised learning 

15 / 20 2 / 26 / 13 Wall – column ratio Supervised learning 

16 / 20 2 / 26 / 13 Number of doors per space Supervised learning 

17 / 20 2 / 26 / 13 Average space gyradius and variance Supervised learning 

18 / 20 2 / 26 / 13 Dataset 35 IFC files of varying quality and complexity amounting up to 109 distinct floor plans manually classified 46 residential 63 as non- residential Supervised learning 

19 / 20 2 / 26 / 13 Results Successfully classified 90% of unseen samples [ training set (70%) cross-validation set (20%) test set (10%) ] Supervised learning 

20 / 20 2 / 26 / 13 Conclusions Data collection difficult impedes learning more interesting things Even well-established models contain flaws The quality of samples varies and exporters adhere to an individual IFC ‘dialect’ learn architectural features rather than picking up exporter idioms 

21 / 20 2 / 26 / 13 Further research ML an alternative perspective to explicit querying mechanisms and model checking. Autonomous quality assurance (typical clearance areas, confluence of specific types) by learning from conforming models. More ambitious classifications 

22 / 20 2 / 26 / 13 Assessing implicit knowledge in BIM models with machine learning Thomas Krijnen Technical University Eindhoven, NL Martin Tamke Centre for Information Technology and Architecture, DK