ThomasMBury/deep-early-warnings-pnas

Implements an ensemble of 20 convolutional neural networks trained on 500K–200K synthetic time series to classify critical transitions into fold, Hopf, or transcritical bifurcations. The approach uses TensorFlow/Keras with dual padding strategies (symmetric and left-only) to handle variable-length input sequences, then compares deep learning predictions against classical early warning signals (variance, lag-1 autocorrelation) via ROC analysis. Training data generation leverages AUTO-07P bifurcation continuation software to simulate diverse dynamical systems, with the workflow designed for distributed CPU/GPU execution across cluster environments via Slurm job submission.