Cross-Frequency Pulsar Noise Prediction Via LSTM With GAN-Based Data Augmentation

This work proposes a novel solution for the prediction of pulsar noise while addressing the critical challenge of data scarcity in specific spin frequencies. The prediction backbone is a bi-directional LSTM network, trained with mixed-frequency data or in the transfer learning strategy, where high-frequency noises are generated via adversarial training. Evaluated on the IPTA dataset, our solution produces accurate predictions in different high-frequency domains, requiring only 10 % real high-frequency noise in the adversarial training process. Such robust generalization makes our solution suitable for real-world scenarios where pulsar observations are limited.