Finding Patterns in Planets: A neural network approach to the exoplanet dataset
We now know of over 3,500 exoplanets; an explosion in growth since the 1990s that shows no sign of abating. 96% of these new worlds have been discovered by either the radial velocity technique or the transit technique. However, both these methods provide only a single measurement of the planet’s bulk properties; either the planet’s minimum mass or the radius. As we step into a new era of exoplanet observations where instruments such as the JWST aim to probe the atmosphere of these worlds, we are left with using this scant information to select the best candidates from a huge dataset for these time-intensive observations.
However, while the information per planet is small, the number of discoveries allows the potential for meaningful statistical analysis. Such techniques can identify relationships between properties to infer missing information. One pathway is to take advantage of the capabilities of neural networks. The principal behind such algorithms is to present the computer with a dataset but no a priori knowledge about links between the elements, allowing the network to locate its own patterns. This technique is particularly strong at finding connections between multiple elements at once and also at identifying connections that have been missed as they fall outside expectations. This research looks at one example of a neural network for estimating the mass and radius of a planet, based on properties available from both radial velocity and transit observations. This could be an added tool to guide sample selection in the future.