How I Built a CIBSE-Inspired Cooling Load Tool With AI — Without First Learning Advanced Mathematics and Thermodynamics
Most air-conditioning engineers I've spoken to seem to agree on one thing: for relatively straightforward domestic and light-commercial applications, cooling loads are often estimated using rules of thumb.
100 W/m² is a familiar one.
Interestingly, after spending time developing a room-by-room cooling-load model, I've gained more respect for that rule of thumb, not less.
For fairly ordinary rooms, 100 W/m² often gets surprisingly close.
The problem is knowing when the room isn't ordinary. What are the edge cases and how do we handle them?
Why build a cooling-load tool?
A 14 m² north-facing bedroom with modest glazing clearly isn't thermally equivalent to a 14 m² south-facing office with large windows, two occupants, super computers running all day, and no shading.
Cooling load can be influenced by:
Solar gain and glazing orientation
Glazing performance and shading
External fabric
Thermal mass
Ventilation and infiltration
Occupancy
Lighting
Equipment and appliances
Operating hours
CIBSE provides established methodology for modelling these effects. Other methods (like the German VDI 2078 exist too).
But there's a problem.
The methodology is considerably more complicated than multiplying floor area by 100 W/m².
I could spend a considerable amount of time learning the advanced mathematics and thermodynamics required to translate the methodology into a working model.
Or, increasingly, there's another option.
What if AI allows us to change the order?
I didn't suddenly acquire the mathematical ability to implement the CIBSE admittance method.
Nor did I become a software developer.
Instead, I started with the parts I could contribute.
I understood the practical engineering problem.
I knew what information an engineer could reasonably collect while surveying a room.
I knew that I didn't want an interface resembling a PhD science experiment.
I wanted engineers to enter wall and window dimensions rather than manually calculate areas. I wanted complicated terminology explained in plain English. I wanted an hourly graph, peak-load contributions, energy and running-cost estimates and an indication of whether a nominal indoor-unit capacity looked sensible.
Most importantly, I wanted to be able to change something and immediately see:
What difference did that actually make?
AI then provided capabilities I didn't have. I understand construction techniques, fabric, u-values, thermodynamics (a little), but each of these is a specialism - you would need a whole team of academics to deliver true, deep expertise - AI provides a sufficient extension of my own skill set to take on challenges that far exceed what I could reasonably achieve in a reasonable time frame.
It helped interpret the published methodology, translate equations into JavaScript, implement solar geometry and thermal-response calculations, build the graphical interface and repeatedly modify the model as I tested it. Some call this vibe coding. Personally I hate using the term as a blanket because it can undermine what expertise the user is able to contribute.
That distinction matters.
I designed the tool and interrogated its behaviour. AI enabled much of the mathematics and software engineering required to make it work.
But AI produced answers that were wrong
This is perhaps the most important part of the story.
At one point the tool's estimated room-temperature-without-cooling curve was producing temperatures that simply didn't look physically credible.
The code worked. The graph worked. The mathematics produced an answer.
I didn't believe the answer was correct.
Investigation showed that the model was effectively allowing heat to accumulate without providing an adequate route for that heat to leave the room.
We changed it.
Later, external feedback highlighted that our treatment of thermal mass was too simplistic. Two walls can have the same U-value while behaving very differently dynamically.
So the model evolved again to include internal walls, floors and ceilings and better represent thermal admittance and room response.
This is exactly why I've deliberately called it the: Prototype Room Cooling Load Model.
AI can generate very convincing mathematics and very convincing code. That doesn't independently demonstrate that the engineering model is correct. Within my own limitations, I can identify glaringly obvious mistakes, but my expertise and experience isn’t sufficiently deep to authoritatively state this model works to ± x accuracy. Do I need to though?
Established rules of thumb 70-100W/m2 are established for a reason - maybe 60% of the time, they work every time!
So who is responsible for what?
I think this distinction is important when discussing AI-assisted engineering.
My contribution: identifying the problem, defining the functionality, designing the workflow and UX, selecting inputs, dictating some defaults, testing outputs against engineering intuition, identifying implausible behaviour, repeatedly challenging assumptions and deciding what the tool should communicate.
AI's contribution: extending my capability into areas of mathematics and software development that I couldn't practically have delivered myself.
CIBSE's contribution: the published engineering methodology on which the cooling-load calculation is based. Let’s be real, without that information, the tool would be far less sophisticated.
And there remains another essential contributor:
Peer review.
Neither I nor AI should be the final authority on whether the implementation is sufficiently accurate for professional use.
That's why I'm making the prototype freely available for engineers to test, criticise and hopefully break.
The most interesting output isn't the cooling load
Initially, I thought the headline number would be the interesting bit.
Design cooling load: X kW.
I'm increasingly convinced that it isn't.
One of my favourite features is now the Peak-hour contribution section.
It tells you why the room needs cooling.
You can see how much comes from solar gain, fabric conduction, occupants, equipment and outside air.
And then you can experiment.
Change the glazing orientation.
Add shading.
Change occupancy.
Increase equipment load.
Change the building construction.
The graph responds.
Suddenly the tool isn't just calculating an answer. It's helping explain the physics behind it.
What if we reduce the problem before installing more cooling?
This led me somewhere else interesting.
Suppose the calculation shows that a room's dominant cooling load is solar gain through the glazing.
The obvious response might be:
Install an air conditioner capable of dealing with it.
But that's not the only option.
What if we reduce the solar gain first?
Yes, proper traditional or modern integrated window shutters like they use in mainland Europe are EXPENSIVE, but are there alternative passive methods before or as well as installing active cooling technology?
The limo effect.
3M publishes technical performance data for its Prestige 70 solar-control window film. I was thinking limo tint, but apparently it’s more sophisticated than that.
Using 3M's reference clear double-pane glazing data, the glazing without film has an SHGC/g-value of approximately 0.70. With Prestige 70 applied, that falls to approximately 0.56.
Those figures give us an interesting experiment.
Run exactly the same room twice and change one input:
g = 0.70 → g = 0.56
The model then shows what happens to solar contribution, peak cooling load, estimated free-running room temperature and cooling energy.
In one of my test rooms, the estimated peak temperature without cooling fell by several degrees.
The estimated daily electricity-cost reduction was actually quite small though, so adding it as a bolt-on to active cooling may not have a meaningful impact or RoI. This would be an intersting edge case to actually test in the real-world.
I think the commercial argument is interesting too.
The value of passive solar control may sometimes be less about saving enormous amounts of electricity and more about improving comfort and reducing peak cooling demand.
Manufacturer-applied solar-control coatings, appropriate glazing specifications, external shading, retrofit films, blinds and curtains all potentially influence the problem before refrigeration gets involved.
The calculator lets us explore that.
And after all that… 100 W/m² isn't terrible
Perhaps my favourite outcome is returning to where I started.
After solar geometry, thermal admittance, decrement factors, ventilation, latent loads, glazing properties and hourly calculations…
100 W/m² often isn't miles away.
That's probably why experienced engineers have successfully used rules of thumb for decades.
But perhaps the better question isn't:
“Is 100 W/m² accurate?”
It's:
“Is this room ordinary enough for 100 W/m² to be a reasonable estimate?”
For a typical bedroom or office, perhaps it is.
For a west-facing glass room occupied all afternoon by someone playing World of Warcraft on a gaming PC with the thermal output of a small jet engine, perhaps we should get the calculator out.
What happens now?
For now, the Prototype Room Cooling Load Model is free to use and available for peer review.
I don't want to present it as validated engineering software until people considerably more experienced in cooling-load modelling have had the opportunity to scrutinise it.
There will almost certainly be things to improve.
That's partly the point.
Because the bigger experiment isn't really about cooling loads.
It's about what happens when someone with practical domain knowledge can use AI to cross traditional capability boundaries.
I didn't first have to become an advanced mathematician, thermodynamics specialist and software developer before I could start building, testing and learning from this idea.
AI allowed me to start with the knowledge I already had, venture into areas beyond my existing capability and learn through the process of building something real.
That doesn't remove the need for expertise.
If anything, this project has demonstrated the opposite.
AI dramatically lowers the barrier to building sophisticated technical tools. It doesn't lower the importance of understanding when the answer might be wrong.
If you want to try the Prototype Room Cooling Load Model, you can access it via the i-heatpump toolkit at: www.i-heatpump.com/toolkit.