AGU · Exploring the exoplanet catalogue with neural nets

Over the last months we developped several ways to explore the exoplanet catalogue using neural networks. In the ‘Rise of Machine Learning’ session, I presented part of our work: how low dimensional embedding helped us see structure in planetary distribution, and a new method to predict planetary mass with higher accuracy. This is another output from the Planetary Diversity workshop held at ELSI a year before. Abstract below.

Laneuville, Tasker and Guttenberg.

The launch of Kepler in 2009 brought the number of known exoplanets into the thousands, in a growth explosion that shows no sign of abating. While the data available for individual planets is presently typically restricted to orbital and bulk properties, the quantity of data points allows the potential for meaningful statistical analysis.

It is not clear how planet mass, radius, orbital path, stellar properties and neighbouring planets influence one another, therefore it seems inevitable that patterns will be missed simply due to the difficulty of including so many dimensions. Even simple trends may be overlooked if they fall outside our expectation of planet formation; a strong risk in a field where new discoveries have destroyed theories from the first observations of hot Jupiters.

A possible way forward is to take advantage of the capabilities of neural network autoencoders. The idea of such algorithms is to learn a representation (encoding) of the data in a lower dimension space, without a priori knowledge about links between the elements. This encoding space can then be used to discover the strongest correlations in the original dataset.

The key point is that trends identified by a neural network are independent of any previous analysis and pre-conceived ideas about physical processes. Results can reveal new relationships between planet properties and verify existing trends.

We applied this concept to study data from the NASA Exoplanet Archive and while we have begun to explore the potential use of neural networks for exoplanet data, there are many possible extensions. For example, the network can produce a large number of ‘alternative planets’ whose statistics should match the current distribution. This larger dataset could highlight gaps in the parameter space or indicate observations are missing particular regimes. This could guide instrument proposals towards objects liable to yield the most information.