Convolutional Neural Networks analysis of numerical simulations
Results of planetary system formation numerical simulations are lists of numerical values that characterise the physical state of synthetic planetary systems. In order to analyse simulation results, one can transform them to other modalities, for example images, in order to benefit from the work developed in other fields. In this project (part of the Certificate for Advanced Studies 'Advanced Machine Learning'), not published yet, we used Convolution Neural Networks (CNNs) to analyse results of simulations, themselves represented as images.
Our goal in this project was to classify planetary systems in two types: 'With Earth-like planet' or 'Without Earth-like planet', an 'Earth-like planet' being a planet that is similar (in mass and distance to its star) to the Earth. This prediction was done assuming that only the most massive and short period planet in the system were known, therefore the system needs to learn, based on the morphology of images representing planetary systems only partially known, it there should be an Earth-twin or not. We used first an home-made CNN (see left), then the Resnet model.
As shown below, the CNN could learn some links relating the presence of an Earth-like planet to the properties of observed planets in the same system (a random guess would be a curve following the first diagonal).
ROC curve (true positive rate as a function of false positive rate) for our CNN Resnet-like model, trained to classify if a planetary system harbours an Earth-like planet or not.
Convolution Neural Network and semantic segmentation for transit data
Transit surveys can easily detect transiting planets with regular orbits. However, when more than one planet is present in the system, and in particular when planets are in mean motion resonance (the ratio of the two orbital periods is close to the ratio between two small integers), the movement of planets is not strictly periodic, making the detection of planets difficult. We participated to the study led by Prof. Adrien Leleu (University of Geneva), in collaboration with Disaitek, to use CNN and semantic segmentation to discover transiting planets in resonant pairs. This was applied to Kepler data.
Architecture of the neural network used for the semantic segmentation task. The encoder (left part of the figure) is a sub- network that works with an increasingly coarse representation of the data by using transition down blocks that reduce the resolution. The decoder (right part of the model) is a subnetwork that progressively combines low-resolution information coming from transition up blocks and high-resolution information coming from the encoder by skip connections. The bottleneck (dense block at the bottom of the figure) is used to refine the high level of understanding that the model has of the input.
Adapted from Leleu et al. (2021), Astronomy & Astrophysics, Volume 655, id.A66, doi:10.1051/0004-6361/202141471
RIVERS.deep confidence matrix. The colours show the confidence of the model that the timing belongs to the track of a planet.
Adapted from Leleu et al. (2021), Astronomy & Astrophysics, Volume 655, id.A66, doi:10.1051/0004-6361/202141471