all research

Real Estate & Development · 7 min read

Controlling a building you have no model of

Every predictive HVAC system so far has needed a calibrated thermal model, which is why almost no existing building has one. ADAPT learns the physics instead of being told it.

7.3%

HVAC energy reduction

30.2%

Discomfort reduction

0

Calibrated thermal parameters needed

Analysis of published research

ADAPT: Physics-Aware Diffusion-based World Models for Adaptive Predictive Transferable HVAC Control

Xu Yang, Kailai Sun, Dianyu Zhong, Qianchuan Zhao · arXiv · 20 August 2026

Predictive control for heating and cooling works. It has worked for years. The catch has always been the prerequisite: you need a calibrated thermal model of the building, and calibrating one is a specialist exercise costing real money on a building that already exists and was never documented properly.

That prerequisite is why this technology has landed almost exclusively on new prestige buildings, and why the enormous existing stock that actually needs the savings has been left alone.

Learn the physics rather than being handed it

ADAPT is a physics-aware conditional diffusion model that predicts indoor environmental conditions. Instead of consuming a thermal model, it enforces building thermodynamics through a learnable multi-zone heat-balance regulariser.

The consequence, stated plainly in the paper, is that it works without known geometry or calibrated thermal parameters. You do not have to survey the building before you can control it.

The prerequisite was never the technology. It was the paperwork about the building that nobody kept.
Reported improvement over state-of-the-art baselines
percent reduction
Occupant discomfort30.2
HVAC energy consumption7.3

Read the two bars in the right order. The energy figure is modest. The comfort figure is large, and comfort complaints are what actually consume a facilities team's week. A system that cuts complaints by nearly a third while saving a little energy is an easier internal sell than the reverse.

Transferability is the commercial claim

The paper reports robust performance across unseen seasons and climate regions. That word transferable in the title is doing the commercial work, because it is the difference between a bespoke engagement per building and something you can roll across a portfolio.

For an Irish or UK asset manager holding thirty mixed-age buildings, that distinction decides whether the business case exists at all. Per-building calibration kills portfolio economics. Transferable models create them.

The honest limits

This is simulation work. Simulated buildings are better behaved than real ones: sensors do not drift, occupancy follows a schedule, and nobody props a fire door open in February. The gap between a simulation platform and a 1970s office block in Dublin is the entire risk of the project.

Pair it with the safety certification work published two days earlier, which I have written up separately. ADAPT answers whether you can control a building you have not modelled. That one answers whether you should be allowed to.

Work with me

Run this model against your own project

I am Kanishk Kapoor, Technical Accounts Manager at AI Institute in Dublin. I build agentic AI systems with built-environment teams across Ireland and the UK. If any figure here looks wrong for your business, that is the useful conversation. Send me your assumptions and I will re-run it.

Continue reading