Skip to main content

Air Quality: How AI can enhance air quality

Design and control: how AI can enhance air quality in our indoor environments 

The air conditioning and heating control panel for the apartment and office is located on a white wall

One of many things the Covid-19 pandemic brought into sharp focus was the importance of our indoor air quality. From keeping patients safe in hospitals to reducing the spread of infection in schools, offices, restaurants and places of worship, air quality was a hot topic among the politicians and scientists drawing up regulations and giving us advice on how to keep the virus at bay.  

Those efforts – not just limited to Covid-19 – are becoming easier with the help of artificial intelligence (AI) and machine learning (ML). Trained models coupled with real-time data enable us to monitor air quality and adjust conditions in healthcare settings, where lives are at stake. The same technologies are also being applied to save energy amid climate change and rising costs, and to ensure our comfort: many of us will have worked in an office that always seems either too hot or too cold. 

A multidisciplinary team of engineers, mathematicians, computer scientists and medical scientists at the University of Leeds is spearheading the UK’s academic research into how AI and ML can be used to enhance the quality of our indoor environments. There are two main applications of these techniques: one to improve building design by producing simulations that predict the impact of different layouts on air quality or climate, and another that uses data from sensors within a building to provide information about – or even control – the environment in real time.  

The field of engineering critical to this work is fluid dynamics – specifically, the turbulent flow of air inside buildings. Airflow simulations can now be produced and analysed using the techniques of ML – a branch of AI in which algorithms allow a computer to learn from large amounts of data and draw conclusions without being programmed by a human. The simulations themselves are similar to those used in weather forecasts: if 70 out of 100 simulations predict rain, then we see a 70% chance of rain on the graphic. In this context, they can help architects design more energy-efficient workplaces or more infection-resilient hospitals. 

Crucially, these ML models can also assimilate real-time information from sensors – such as temperature or air quality data. This ensures that simulations remain up-to-date and enables them to make predictions or suggestions about the best course of action to take. In a hospital, that could mean reducing humidity to inhibit the growth of pathogens; in a museum, it could save energy by turning on the heating or air conditioning only in the areas that need it. 

One big concern about ML models and AI more generally relates to transparency and the potential for bias. If we simply feed in data and let the model run its course, how can we be confident in its outputs? How do we know nothing strange is happening during the middle part of the process where humans aren’t involved? In this case, Leeds researchers have sought to mitigate the uncertainty around outcomes generated by traditional ML models by programming in the laws of physics. This constraining strategy, known as physics-informed machine learning, makes sure the Leeds team’s models only produce simulations or solutions that are possible in the real world. 

Moreover, incorporating the governing equations of fluid mechanics at the outset may also help reduce the time and computing power required to train the models, without compromising on accuracy. While trained ML models are quick and cheap to run, the training phase itself – involving vast amounts of information from real-world experiments or existing simulation data – is often slow and expensive, representing a bottleneck in the process. 

The research team is now working with external companies to explore how ML models can help them meet their building design needs, and is also investigating how to make environmental control technology accessible in settings that lack the financial or computing resources required to run such systems in their current form. 

Dr Amirul Khan, Lecturer in Environmental Fluid Mechanics at the University of Leeds, says: “Artificial intelligence and machine learning have huge potential to help us design better indoor environments and control those that already exist. We still have plenty of work to do to make sure these models are as accurate as possible and adaptable across different situations. But there is no doubt that ML, if we apply it carefully and properly, can be a game-changer for air quality control around the world – and both Leeds and the UK are in an excellent position to lead the way on this.”