Neural network sifts Kepler signals and validates 301 new exoplanets

The work

Technical · from the record

ExoMiner is a deep neural network that classifies Kepler threshold-crossing events as planet candidates or false positives, using a separate convolutional branch for each diagnostic view that appears in the mission's Data Validation reports, plus scalar diagnostic and stellar parameters. It was trained on a labelled set of 30,609 Kepler Q1-Q17 DR25 transit signals and evaluated with 10-fold cross-validation against Robovetter, AstroNet, ExoNet and the Gaussian process and random forest classifiers of Armstrong et al. The paper reports that at a fixed precision of 99% ExoMiner recovers 93.6% of exoplanets in the test set, compared with 76.3% for the best existing classifier it compares against. Applying a 0.99 score threshold, scenario priors and catalogue-flag vetoes to 1,922 unlabelled Kepler Objects of Interest, the authors report 301 newly validated exoplanets, and a branch-occlusion test is used to identify which diagnostic branch most influenced each score.

How AI was used

Flux and centroid time series from the Kepler Q1-Q17 DR25 light curves were detrended, phase-folded and binned into full-orbit views of 301 bins and transit views of 31 bins for flux, centroid motion, the weak secondary event and the odd and even transits, and the associated DV diagnostic and stellar scalars were normalised. Each view was fed to its own one-dimensional convolutional branch, with the odd and even views passed through a shared branch followed by a non-trainable subtraction layer; a fully connected layer after each branch merged the extracted features with that test's scalars, and the concatenated features plus stellar and DV diagnostic scalars entered a fully connected block ending in a logistic sigmoid output. Architecture and training hyper-parameters were selected with the BOHB optimizer, and weights were fitted with Adam in Keras on TensorFlow. Models were trained and scored in 10-fold cross-validation with the split taken over target stars rather than individual signals, alongside retrained AstroNet and ExoNet baselines and an ExoMiner-TCE variant split at signal level. Sensitivity to training set size, to the proportion of astrophysical false positives and to injected label noise was examined by retraining on reduced or corrupted sets. Scores for unlabelled Kepler Objects of Interest were averaged over the ten models, combined with published false positive scenario priors under a Bayes rule, and filtered by a 0.99 rejection threshold together with catalogue disposition flags and a minimum multiple event statistic. A branch-occlusion test, which zeroes each time series branch in turn and re-scores the signal, was used to attribute each disposition to a diagnostic branch, and a reduced model trained only on Kepler time series and stellar parameters was applied to preprocessed TESS SPOC signals without further training.