FORMATION OF BIM MODEL DATASET FOR MACHINE LEARNING TASKS

Abstract

This paper presents a reproducible approach to deriving machine-learning-ready datasets from Autodesk Revit Building
Information Models. We unify four complementary export routes‒Schedules→CSV, Dynamo→CSV/XLSX/JSON,
pyRevit/Revit API→CSV/JSON/XLSX, and IFC→Autodesk Platform Services/Speckle→JSON/CSV‒under a single data
contract: a target schema/ontology with mandatory fields, SI units, a stable key policy (IfcGUID/composite key), and formal

data-quality validation. The methodology covers ingestion, normalization of types and units, construction of geometric
descriptors, checks of completeness and referential integrity, and feature preparation for reliable train/validation/test splits.
Provided a practical protocol and a final Elements.csv dataset used in a baseline task that predicts element mass as a
proxy for cost and logistics; templates for quality checks accompany the flow. Across applied scenarios‒production-rate
estimation, cost approximation, and optimization of layout and load-bearing systems‒we show that disciplined data
management (schema, keys, units, rules) reduces information leakage and improves reproducibility.
The contribution aligns tool-specific export routes with a uniform representation, enabling integration of BIM data with
analytics without dependence on a software stack and establishing a foundation for data-centric design and subsequent
optimization. We also document reproducibility artefacts (JSON/Table Schema definitions, DQI reports, and processing logs)
to support audit and repetition.