Richard Wilkinson

How do we learn from complex models?

Simulation models are increasingly being used by scientists, engineers, medics etc to model complex phenomena. The simulations can be used to improve our understanding, or to predict the future if we were to take some hypothetical action. But models are not perfect representations of reality. I'm interested in the question of how we can best learn about the real world, from an imperfect model of it. 

Humans are currently conducting a massive experiment on the climate, pumping ever increasing amounts of CO2 into the atmosphere. Our main tool for predicting what future climate will look like are complex physics-based models of the climate system. But the climate is so complex that we can't build a perfect simulation of it. I was fascinated about how we could make good predictions about the future given the limitations of our climate simulators. 

I work on several projects with cardiologists to build models of the hearts of patients being treated for arrhythmias. This involves talking to cardiac modellers, as well as clinicians who collect data, and then developing methods to combine their models and data.  Our aim is to predict what treatment has the greatest chance of success for a given patient. For this to be effective, we need to be able to do it during a short surgery. Consequently, I spend a lot of time trying to find computational efficiencies allowing us to speed-up our calculations. 

John Tukey once said “The best thing about being a statistician, is that you get to play in everyone's backyard”. I like that fact that the problems I'm interested in solving occur across nearly all areas of science, engineering, public policy etc. I love talking to new colleagues and trying to understand whether statistical methods can help them solve their problems.

Talking to scientists from other disciplines often means there is a language barrier. Each discipline has its own terminology, as well as its own epistemological understanding of the value of models. I like to listen to colleagues describe their problem, and then try to explain back to people what I think I've heard. This can be painful at times as it illustrates how little I know, but it does ensure you learn quickly!

We set up a network of cheap low-cost air pollution sensors in Kampala, with the aim of understanding whether the data quality can be improved if we combine it with a complex simulation model. We were able to show that it can be. We're now looking at whether we can infer the source of air pollution in the city in a similar way. Kampala has all the same air pollution sources we experience in Nottingham, as well as pollution from home generators, unpaved roads, cooking on solid fuel stoves etc, and we'd like to understand the contribution of each source.

I hope the methods we develop become widely used by scientists and engineers who work with complex models. In cardiac applications, our hope is that our methods are used by cardiologists to guide catheter ablation treatments in patients suffering from atrial tachycardia, and that each patient will have their own 'digital twin' tracking the health of their heart built using our methods. In Kampala, we hope to guide the local government about the best actions they could take to lower pollution levels.

I'm fortunate and I teach modules directly relevant to my research. For example, one module is on optimization, which I can relate to trying to optimize the treatment cardiac patients receive. The other module is on multivariate statistics, where we look at methods for dealing with high dimensional data. I'm able to show how we use the methods we teach in complex engineering applications to train complex simulation models. 

Publish more. I was too reluctant to put ideas out there earlier in my career and consequently missed out on things. Exposing your ideas to criticism is scary, but is also what improves them.