Deploying AI for a Greener Future: Insights from the Field

‍
by Marco Salinas

‍

When most people hear “machine learning,” they think of advanced algorithms and data science breakthroughs. In sustainability, however, the real challenge isn’t the model itself — it’s whether those models lead to practical, trusted, and impactful outcomes in the built environment.

At Hubble, we’ve spent years applying ML to real-world sustainability projects, from predicting building comfort and thermal loads to testing shading systems and estimating carbon impacts. Here are some lessons we’ve learned along the way.

1. Data quality over data quantity

Big datasets alone don’t guarantee useful models. In sustainability, inconsistencies in climate files, building attributes, or occupant data can erode accuracy faster than model choice. We’ve learned that rigorous preprocessing, cleaning, and feature standardisation are where most of the real work lies.

That said, good representation of data is just as critical. For example, at Hubble we recently generated more than 500,000 datapoints in collaboration with WSAA to analyse shading products across diverse building and climate conditions. These representative datasets ensure our models capture the full range of performance outcomes, not just the average.

2. Representative datasets are critical

At Hubble, much of our dataset is deliberately generated and synthesised through modelling to ensure representativeness across the full performance spectrum — from poorly performing dwellings to highly efficient homes. These extremes are rarely observed in real-world samples but are essential for training models with sufficient variance, stress-testing prediction intervals, and avoiding bias towards the “middle of the pack.”

3. Localisation beats generalisation

A model tuned to Victoria’s housing stock won’t automatically work in Queensland, and an algorithm trained on temperate climates may underperform in arid zones. We’ve found that localising models — aligning them with regional climates, construction practices, and regulatory contexts — delivers far more useful and trusted outputs than forcing a one-size-fits-all approach.

4. Explainability builds trust

Sustainability decisions shape investments that last decades, so stakeholders need to understand the “why” behind recommendations. At Hubble, we design for transparency: prediction intervals, clear assumptions, and human-friendly scoring frameworks like ComfortScore™. When users can see both the outputs and the uncertainty, adoption and confidence rise dramatically.

5. Human behaviour is the wild card

You can model insulation R-values or glazing performance with precision. But occupant behaviour — when windows are opened, thermostat settings, or appliance use — introduces variability that’s harder to capture. Incorporating behavioural uncertainty, even imperfectly, has been critical to making our models reflect lived reality rather than lab conditions.

By contrast, in the NatHERS framework, these behavioural parameters remain largely obscured. Default assumptions about heating and cooling schedules, infiltration rates, internal gains, and thermostat set-points are effectively black boxes. Assessors and end-users have little visibility into how these assumptions shape star ratings — even though they significantly influence the result.

Tools like EnergyPlus take a different approach, offering user-defined schedules and granular control over occupant behaviour. This transparency allows for more flexible scenario testing and builds trust in the outputs. If NatHERS moved toward even partial openness in these parameters, it would benefit the entire industry by fostering both confidence and innovation.

6. Impact matters more than accuracy

In academia, accuracy metrics like R², MAE, or RMSE are the finish line. In the real world, they’re only the start. The real question is: did the model drive action? Did councils accelerate retrofits? Did banks integrate sustainability into lending? Did homeowners choose more efficient upgrades?

Machine learning delivers value only when it translates into tangible change.

Who to Follow

If you want to stay ahead of how machine learning is transforming sustainability, here are a few voices worth following:

• Rafa Felix, PhD — expert in building physics, simulation, and ML applications for sustainable design.
• Dr. Clayton Miller — researcher at NUS focused on building performance, digital twins, and data science.
• Sara Behdad, PhD — specialist in sustainable design and circular economy modelling.
• Bahar Dadkhah, PhD — advancing AI-driven energy modelling for high-performance buildings.

👉 Who else should be on this list? Tag them in the comments — let’s keep building the conversation.

Ready to Act?

If you’re looking to begin your AI transformation journey and embed sustainable practices into your projects, Hubble is here to help. Our expertise in machine learning, building science, and data-driven insights can support councils, banks, builders, and homeowners in accelerating their path to net zero.

💡 Let’s work together to make homes more comfortable, efficient, and future-ready. Contact us to learn more and start your journey today.

‍