As we’ve blogged about previously, the Royal Society’s machine learning policy project has been investigating the potential of machine learning in a range of fields, including manufacturing, pharmaceuticals, and cities. But what effect might machine learning have on research itself?


To find out more about how researchers are using machine learning, we talked to Dr Neale Gibson from Queens University, Belfast, an Astronomer by profession, who has been applying machine learning techniques to his resolve research challenges in astronomy, supported by a Royal Society University Research Fellowship.


What fuelled your interest in science, and why astronomy?


I’ve always been interested in science and mathematics – alongside art, they were my favourite school subjects. When I was young, my Dad bought me a set of encyclopaedias which included books about the planets of our Solar System and space. I was fascinated by the beautiful pictures, and this inspired my interest in astronomy. I studied physics at university, but it wasn’t until late in my degree that I realised a career in astronomy was possible! I had the opportunity to do a PhD in astronomy, and I’ve never since doubted that this is the field I want to work in. I feel privileged that I can make a living by thinking and learning about space.


What is the focus of your research?


My research focuses on the characterisation of planets that are not in our Solar System (called exoplanets).

Astronomy and machine learning

Exoplanets are pretty common, there are at least a few hundred billion of them in our galaxy alone – more than there are stars! Some of the exoplanets we have found are located in the so-called “habitable zone”, which is the region around a star where liquid water can exist on a planet’s surface; this is one of the main requirements for life.


My research concentrates on a class of exoplanets called ‘transiting planets’. These periodically eclipse their parent star, blocking the light.


Little is known about exoplanet systems, apart from their size, mass and orbital properties, but we can figure out their atmospheric composition by analysing patterns in the light we detect from them. A small portion of starlight filters through the upper atmosphere of the planet, and the amount that is absorbed changes depending on the gases present. Understanding these changes allows us to predict the composition of the atmosphere.


How are you using machine learning in your research?


Unfortunately the planet’s atmospheric signals are often corrupted by time-correlated noise from instrumentation. This needs to be removed to reveal the planet’s signal and we use machine learning to identify and remove these signals from the data enabling us to analyse and infer the planet’s atmosphere.


Previously we would have to manually specify models to explain the correlated noise in exoplanet signals. The choice of model would then affect the final shape of the planet’s extracted signal, and the atmospheric composition inferred from this. Machine learning enables us to do this more robustly, leaving a clean signal that we can use to understand the composition of the exoplanet’s atmosphere.


In the future we will be searching for ‘biomarker’ gases (signs of life). Finding oxygen, ozone, and water on a nearby, Earth-like planet, would be strong evidence for the presence of extra-terrestrial life! But what if these detections were based on human interpretation and arbitrary choices during analysis? Machine learning techniques increase the strength of our analysis, and help increase our confidence in our conclusions.


What opportunities could machine learning unlock in your field of research?


At the moment, it takes huge amounts of human expertise and time to sift through the data we have to find and characterise exoplanets. Future space missions will produce more data than ever before, and making sense of this will require new tools to automate some of this analysis. We are entering in to a period where the technical challenges associated with data analysis are beginning to exceed the challenge of building new facilities! This means there are lots of opportunities for machine learning in astronomy.


In my field of exoplanet atmospheres, we analyse observations for a single planetary system in isolation. However, it should be possible to combine information from similar observations taken from different targets, observatories and instruments. A machine learning system could learn how to most reliably extract information from the data, developing sophisticated understanding of instrumental quirks, more than a human analyst. We can already make steps towards this by combining more datasets from specific instruments, however for this system to function it would require vast amount of computing power.


Machine learning could also influence how we interpret exoplanet signals. For example, we could train software with detailed physical models of a variety of exoplanet types, giving it the objective to predict the composition of planetary atmospheres. This could identify targets for more detailed modelling and follow-up observations. However, it will be a while before we have enough detailed observations of exoplanet atmospheres to require such a system.


What are the challenges that machine learning presents in your field of research?


One is simply that many astronomers (myself included) are not formally trained in statistics and machine learning techniques, and this can mean there is a bit of reluctance to make use of machine learning, especially in areas where traditional techniques still appear to work.


Another is that, as well as requiring expertise outside of traditional astronomy, machine learning applied to large data sets requires large computational-resources, which can be difficult to access.


A final problem is to do with the interpretation of the results from machine learning. Complex deep-learning algorithms can outperform humans, which could give fascinating insights. But without understanding the connections made that lead to their results, what new physical insight do we learn in the process? What conclusions could we really draw?


The Royal Society is undertaking a project about machine learning and its impact for the UK economy and society. To find out more please visit our project page.