Title: Leveraging Long Short-Term Memory Networks for Few-Shot Pulsar Noise Prediction

Abstract:

This study introduces an innovative approach to forecasting pulsar timing residuals under conditions of data scarcity, a significant hurdle when dealing with spin-frequency subgroups of millisecond pulsars within Pulsar Timing Array (PTA) datasets. The method employs a Long Short-Term Memory (LSTM) network, which is refined through a model-agnostic meta-learning algorithm. This optimization allows the model to quickly adapt to new frequency domains by adjusting its parameters using only a small number of ground truth timing residuals. Additionally, the automatic tuning of hyperparameters is achieved via a particle swarm optimization algorithm, which enhances the overall prediction accuracy.

When tested against the second data release from the International Pulsar Timing Array (IPTA), the proposed solution exhibited strong generalization capabilities. It delivered precise predictions across three distinct performance metrics in high-frequency test domains, despite relying on merely 10% of the available timing residuals from those specific domains for fine-tuning. The architecture is notably efficient, consuming only 16.86 MB of CPU memory and requiring just 18 milliseconds for a single-step residual prediction. These attributes render the solution particularly well-suited for practical deployment, where real-time and effective forecasting is crucial, especially in settings characterized by limited computational resources, memory constraints, or energy restrictions.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC