All projects

Machine Learning

Graph Neural Networks for Overheating Prediction

TM59 and Part O compliance in under 10 seconds, ~100x faster than IES VE

A graph neural network that predicts residential overheating risk in under 10 seconds. I led dataset generation, training, deployment, and the Grasshopper interface that moved the tool from notebook to 10+ live consultancy projects.

Live overheating prediction inside Grasshopper — TM59 / Part O results returned in under 10 seconds per option.

PROBLEM

UK residential overheating (TM59, Part O) is normally assessed in IES VE: expensive licensing, slow setup, hour-scale runs. Early-stage teams skip it because it does not fit their iteration loop. Catching overheating late means facade rework, MEP redesign, missed planning conditions.

I joined as the second engineer on an internal alternative: a fast surrogate that lets architects test dozens of options before committing.

DATASET

I built a parametric Part O workflow on EDSL TAS and SAM, calibrated against IES VE on live projects. Once outputs matched within compliance tolerance, I scaled it to generate 100,000+ synthetic overheating cases.

Parameter sweep: UK CIBSE weather files, dwelling typologies (single/dual aspect, 1-4 bed), glazing ratios (10-60%), shading combinations, and ventilation strategies (natural, mixed-mode, mechanical). Targets are TM59 hours-exceeded and Part O pass/fail. Sister dataset to the daylight CNN; shares the parametric Grasshopper setup but varies independently in target metric and parameter sweep.

MODEL

Graph neural network. Chosen specifically because dwellings vary in room count and adjacency; padding-based architectures (CNNs, MLPs) lose topology information. Nodes carry room metadata (area, orientation, glazing); edges encode adjacency and shared boundaries.

Trained on an Azure ML GPU instance. Multiple architecture iterations on message-passing depth and feature representation.

VALIDATION

  • 97% pass/fail agreement with EDSL TAS on a held-out test set, sampled to match the live project parameter distribution.
  • Sub-10s round trip from Grasshopper, vs 1-4 hours for an equivalent EDSL TAS run on the same dwelling (measured on a local workstation).
  • FastAPI tuning to hit the latency target.

DEPLOYMENT

FastAPI service on Azure, Docker containerisation, Grasshopper component for daily use by the Sustainability and Building Physics team. UI is a single Grasshopper node, no remote-service feel.

ACHIEVEMENTS

  • Sub-10s predictions at ~97% pass/fail agreement vs EDSL TAS.
  • 10+ live consultancy projects shipped using the tool (UK residential schemes).
  • Foundation for the residential-sector ML pipeline now reused across the peak solar and daylight tools.