Machine learning predicts whether a dying massive star explodes, from its density profile

The work

Technical · from the record

A random forest classifier was trained to predict whether a massive star explodes, using labels from 100 two-dimensional axisymmetric Fornax core-collapse supernova simulations, of which 64 exploded and 36 did not. Input features were the compactness parameter, the Ertl condition parameters, a newly defined pair describing the mass coordinate and density jump of the silicon/oxygen interface, and an eight-dimensional embedding produced by a convolutional auto-encoder trained on the 1412 progenitor density profiles that were not simulated. On the authors' cross-validation splits the silicon/oxygen interface features gave an accuracy of 0.89, the auto-encoder embedding 0.84, compactness 0.83 and the Ertl condition 0.70, and the authors report roughly 10% fluctuations in scores between splits. A semi-supervised test that dropped 50% or 75% of the training labels and reassigned them by label propagation changed accuracy by a few percentage points, in both directions depending on the feature set.

How AI was used

Explosion outcomes from a suite of 100 two-dimensional Fornax simulations supplied binary labels. Density profiles of the 1512 one-dimensional progenitor models were truncated to mass coordinates between 1 and 2.3 solar masses, re-binned by interpolation onto a uniform 128-point grid in the logarithm of density, mean-subtracted and normalised. A convolutional auto-encoder built in PyTorch, with three Conv1d encoder layers, three ConvTranspose1d decoder layers, tanh activations and about 10^6 trainable parameters, was trained on the 1412 unsimulated profiles using mean squared reconstruction error, the Adam optimiser, Kaiming-normal weight initialisation, batch size 100, learning rate 10^-2 and 500 epochs, with embedding dimensions varied from 2 to 32 by adjusting the convolutional strides; no hyperparameter search was performed. The resulting embeddings, and separately the hand-computed compactness, Ertl and silicon/oxygen interface parameters and simple profile summary statistics, were used as feature sets for scikit-learn RandomForestClassifier models with five estimators, gini criterion, maximum depth three, minimum samples per leaf two and sqrt max features, trained and evaluated over a fixed-seed stratified five-fold 80/20 split of the labelled set. A semi-supervised variant removed 50% or 75% of the training labels and reassigned them with scikit-learn LabelSpreading using a k-nearest-neighbour kernel with five neighbours and a clamping factor of 0.1 before retraining the same classifier.