Beating the Heat: can AI supercharge school retrofit projects in Tanzania?
Author: David Vigoureux, FT Hub
Image: Economical use of Limestone Finishing in a newly-refurbished school building
It’s the Utafiti Elimu conference in Tanzania, 2024, and Mauricia Nambatya is sweating.
Mauricia is sitting in a large conference room with oppressive heat all around her. There are fans, rhythmically passing to and fro in seemingly futile gestures. Other attendees in the room have taken matters into their own hands by fanning themselves with books, mustering whatever breeze they can.
In short, it’s hot. And the irony isn’t lost on Mauricia, who later told us about the event:
“The conference theme was on climate, education and the environment, and the policy makers who were in this room were facing it squarely. So it wasn’t something far-fetched… We were in a large space, abundant, and we had to try and translate what we were experiencing to the classrooms.”
Learning in the heat
Mauricia is a civil engineer working in sustainable construction for Haileybury Youth Trust Uganda. She’s seen how hot classrooms in Tanzania can get, where temperatures regularly exceed 35 degrees Celsius, making it really hard for children to learn.
“You feel the heat straight on, and the discomfort. These are classrooms which are packed to the extent that the teacher doesn’t even have anywhere to teach from. But the classes push on, they don’t stop.”
Research suggests that people have a thermal comfort zone which is different for every person, and the temperature of a classroom could have significant impacts on their learning.
In other words, it’s one thing for a child to attend a two-hour lesson, and another as to how much is actually learned.
Recently, there have been efforts to explore how classroom conditions in Tanzania can be improved by school retrofits. Contextually appropriate retrofits include painting roofs with reflective colours and using different construction materials to reduce classroom temperatures. Other conditions like light and sound can impact learning, and retrofits can improve these too.
The Frontier Tech Hub, a UK FCDO-funded programme, is exploring if artificial intelligence (AI) can be used to prioritise individual schools most in need of temperature-focused retrofits. This could have real, tangible impacts on learning in Tanzania and across similar environments, because it would help maximise money spent on improving learning outcomes by donors and governments.
But why artificial intelligence, and how did we get here? For that, we’re going to have to step back into the classroom…
From research papers to rapid models
Jamie Proctor works in the Research and Evidence Department of the FCDO and is currently based in Ghana, leading the West Africa Research and Innovation Hub (WARIH). Jamie has a background in school construction, with a passion for building more and better schools in order to improve learning outcomes for children. From 2012, he helped to establish an NGO called The Mlambe Project, which works in Malawi to support sustainable and climate friendly classroom construction. He went on to study a sustainable construction MSc, alongside his role at FCDO. We spoke to Jamie to hear more about it:
“With The Mlambe Project I was very hands-on in terms of trying different building methods, building materials and building styles. One thing I always wanted to look at was the classroom temperature, as it was often the biggest challenge to children learning. Later, as part of my Masters, I did a module on building physics, where I built a heat flow model from 1st principles looking at a classroom block in Malawi. It was built on Google Sheets, and included formulas for heat going in, another for heat going out, through the walls, through the roof, with solar gains, etc. That model showed that the heat flow from the roof is the biggest determinant in classroom temperatures for a basic classroom block”.
Jamie then explored how different retrofit interventions would have an impact on classroom temperatures:
“I tested a particular intervention in Tanzania as part of my dissertation – painting roofs white. Out of that, we had the basis for various projections to make the direct link between classroom environment and temperature.”
Jamie built on the concept in his role at the FCDO, with Senior Education Adviser Colin Bangay, and sourced funds from the East Africa Research Hub to lead two complementary pieces of research. Specifically, the organisations Fab Inc and Laterite studied the classroom experience in Tanzania, and OpenDevEd (ODE) explored retrofit options. Jamie set up a steering group (CLEEAR) for advice and input into the studies, including from donors, to ensure that infrastructure investments could be better informed:
“The really important thing for this work is it's moving away from where infrastructure investments haven't worked so well, which is where it is seen as kind of always a good thing, towards it being seen as something that can be a good thing if it's done in a very, very cost-effective way, with people looking at what the learning outcome is.”
This targeted approach has continued with the Frontier Tech Hub. Jamie has worked with the Hub to bring in a research partner to explore if AI could assess indoor classroom temperatures and understand the data collection needed to create a model of sufficient accuracy.
Björn Haßler, CEO of OpenDevEd, is leading the team working on this. OpenDevEd were tasked to evaluate existing literature, which to now has focused on predicting smart building temperature in predominantly high-income contexts:
“We know there's some kind of basic physical modelling that can be done. But the research literature says there are significant improvements that can be made on that through the use of artificial intelligence. Now, when we talk about AI, most of the time what's in the news is about large language models, which is fairly recent development. The AI we're talking about here is neural networks, which is something that's been around for probably about 30 years or so, but has found a really important application in this sort of building temperature control.”
While reviewing the past work, Bjorn and Jamie realised they needed a practical next step to test emerging ideas. They subsequently invited Omdena to lead a challenge that would allow different AI models to be tested rapidly.
Omdena is a platform helping mission-driven organisations and startups build impactful AI solutions through global collaboration, and empowers AI Engineers from all over the world to become changemakers. Duplex Younkap Nina is project manager for this challenge, and sees Omdena’s success as coming from its openness to people from diverse backgrounds and cultures. The challenge format allows machine learning engineers and domain experts to work hand-in-hand with more experienced counterparts, who in some cases are paid as top-level talent. Most importantly, it offers people the opportunity to work on real-world problems and real-world data, the holy grail for data scientists.
“You have people from an academic background who lack experience. For example, I’m a computer engineer by study, and my career is as an air traffic controller but I also have a passion for software engineering and I am an AI engineer.”
Duplex set up a kick-off meeting and ensured that people were split into different teams based on their interests and expertise.
The teams ranged from people specialised in data collection and analysis, to AI/machine learning development, software engineering, and features like simulations. After the kick-off, the challenge had 16 active collaborators across 10 different time zones.
The team quickly learned that the most promising type of model identified in the literature review - long-short-term memory (LSTMs) - were less successful than imagined. The team saw this as stemming from the varied climatic zones across Tanzania which, combined with the different characteristics of each school (orientation, building material, etc.) complicated the development of one particular model.
During the challenge the focus moved onto hybrid models, combining physical simulations of how heat might change under certain building conditions with multiple AI models working together, like a decision tree acting with an LSTM model.
“We don’t quite know just yet what the right combination of models is, and that is part of the challenge. People are trying different models, and we’ll hope to then be able to say which are the most promising.”
The initial results of the challenge were highly promising, and its implications are highlighted in a report authored by OpenDevEd.
The ultimate aim of the tool is to use meteorological data and school building characteristics to predict the maximum indoor classroom temperature of a building, or the number of days per year above a certain temperature. Whereas many of these measures can be taken remotely from satellite images, one of the core recommendations of the report was that extra data collection will be needed to build the model, especially of different classrooms within the same school, to help build a more accurate model.
The hope is that indoor classroom temperatures across Tanzania can be estimated in a cost-effective way, instead of the infeasible approach of measuring temperature directly using sensors in every location. According to Björn Haßler, the impact of this on international development support is clear:
“We [normally] select a set of schools that are in a priority region. But then, within that priority region, we treat all schools the same. Sometimes that's not best for investment. It's quite good to have very detailed models that say, right, these schools need ceiling boards, these schools need new roofs… And I think those assessments are often difficult to do. But with our data, it could become much easier to do that and therefore to target. That would allow the government to assess the building stock, as it were, and say, right, these schools have more than 20, 50, or 100 days of these very hot temperatures, and those are the schools that we direct donor funds to.”
The collective effort needed
This project is just the beginning of exploring this topic and the next step, according to Bjorn, is the data collection needed to build an accurate model.
“I want to be in a place where we have a temperature prediction for every single school in sub-Saharan Africa, because we know where those schools are. We know what the climate parameters are, [and] if we can gather the right data from space… The question is, how much data do we need before we can make a good prediction for all schools in Tanzania, all schools in East Africa, all schools in tropical Africa? So there's a lot of work to be done.”
There is optimism that this initial work will encourage organisations like the World Bank to conduct a data collection campaign in Tanzania, or even lead an AI challenge looking to explore this topic further. The hope of improving classroom conditions and therefore learning outcomes through retrofits is still an emerging field, but how it can inform donor decision making can’t be more pressing for Jamie:
“If you're a policymaker and you're looking at what to do, my advice is to be ruthless with how you spend your money, and only spend it where you can see a clear thing that will be improved. Have discussions with people that have actually done some of this work work before, and particularly deciding what interventions you should pick, where they should go, and how they should be to be delivered. If your aim is to keep learning outcomes at the same level as they currently are, you need to act now.”
Read the project outputs here:
💡 Results of the Omdena challenge
If you’d like to dig in further…
See the OpenDevEd Evidence Library
Get in touch with the Frontier Tech Hub - frontiertechhub@dt-global.com
