All Work
AI / ML

3D MEP Component Predictor

Neural network system for automated MEP placement in BIM models

Equans / Maastricht University·2025

>50%

Resource Savings

3D + Type

Prediction

MLP + CNN

Models

BIM/Revit

Domain

Problem

Placing MEP (Mechanical, Electrical, Plumbing) components in 3D building models is a time-consuming manual process requiring specialized knowledge. Equans engineers spend significant hours on repetitive placement tasks.

Approach

Developed PyTorch neural networks (MLP and 1D CNN with depthwise separable convolutions and residual blocks) that predict both the 3D coordinates (x, y, z) and component type for MEP elements within Revit building models.

Data Extraction

BIM model parsing

  • Revit IFC export
  • 3D coordinate extraction
  • Component type labeling

Feature Engineering

Spatial encoding

  • Room context features
  • Adjacency matrices
  • Normalized coordinates

Model Training

Neural network

  • MLP baseline
  • 1D CNN with residual blocks
  • Depthwise separable convolutions

Deployment

Revit integration

  • Coordinate prediction
  • Component type classification
  • Batch placement

Results

Achieved over 50% reduction in resource requirements for MEP placement tasks. The model accurately predicts component locations and types, significantly accelerating the BIM workflow.

Technologies

PythonPyTorchRevit APIIFCNumPypandasscikit-learnMatplotlib