Gamma-ray burst light curve reconstruction with predictive models

Gamma-ray bursts represent some of the most energetic and complex phenomena in the universe, characterized by highly variable light curves that often contain observational gaps. Reconstructing these light curves is essential for gaining deeper insight into the physical processes driving such events. This study proposes a machine learning-based framework for the reconstruction of gamma-ray burst light curves, focusing specifically on the plateau phase observed in X-ray data. The analysis compares the performance of three sequential modeling approaches: a bidirectional recurrent neural network, a gated recurrent architecture, and a convolutional model designed for temporal data. The findings of this study indicate that the Bidirectional Gated Recurrent Unit model showed the best predictive accuracy among the evaluated models across all GRB types, as measured by Mean Absolute Error, Root Mean Square Error, and Coefficient of Determination. Notably, Bidirectional Gated Recurrent Unit exhibited enhanced capability in modeling both gradual plateau phases and abrupt transient features, including flares and breaks, particularly in complex light-curve scenarios.

💡 Research Summary

This paper presents a machine‑learning framework for reconstructing the X‑ray plateau‑phase light curves of gamma‑ray bursts (GRBs) using data from the Swift XRT archive. The authors first describe the scientific motivation: GRB light curves are highly variable, often contain observational gaps, and accurate reconstruction is essential for physical interpretation and cosmological applications. After a concise review of previous parametric (e.g., broken power‑law, W07) and data‑driven approaches (t‑SNE clustering, Gaussian processes, auto‑encoders), the study focuses on three sequential deep‑learning architectures—Bidirectional Long Short‑Term Memory (Bi‑LSTM), Bidirectional Gated Recurrent Unit (Bi‑GRU), and Temporal Convolutional Network (TCN)—that are capable of handling irregular, missing data.

Data preprocessing is a central contribution. Raw FITS files are parsed to extract observation time, flux, and flux uncertainty. Time stamps are normalized by dividing by 10², and both flux and its uncertainty are transformed into logarithmic space using the minimum non‑zero flux as a reference, which stabilizes training by compressing the dynamic range. To mitigate sparsity, the authors insert 19 equally spaced interpolated points between every pair of consecutive observations, effectively raising the temporal resolution and providing dense context for the models. Sequences are then segmented into fixed‑size batches; short sequences are up‑sampled by repetition to meet the required length, ensuring uniform input dimensions across the dataset.

The three models are implemented in TensorFlow/Keras with GPU acceleration. The Bi‑LSTM network consists of five stacked layers (100 hidden units each) with return‑sequences enabled for the first four layers, followed by a dense ReLU layer that outputs a single flux prediction. The TCN comprises three 1‑D causal convolutional layers (64 filters, kernel size 5) with dilation rates of 1, 2, and 4, each followed by ReLU activation and weight normalization; a final dense layer produces the flux estimate. The Bi‑GRU, the focal model of the study, contains two stacked bidirectional GRU layers (64 hidden units each), the first returning sequences to feed the second, and a linear dense output layer. All models are trained with the Adam optimizer, mean‑squared‑error loss, and early stopping (patience = 5 epochs) on a 70 %/30 % train‑validation split. Batch sizes are set to 900 or larger, depending on the number of interpolated points.

Performance is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²). Across the entire test set and within sub‑categories (simple plateau, plateau with flares, abrupt breaks), the Bidirectional GRU consistently achieves the lowest MAE and RMSE and the highest R². Notably, in complex light curves containing rapid flares and sudden breaks, Bi‑GRU reduces prediction error by roughly 12 % relative to Bi‑LSTM and by an even larger margin compared to TCN. In addition to accuracy, Bi‑GRU demonstrates superior computational efficiency: it requires fewer parameters, trains faster, and consumes less GPU memory than the Bi‑LSTM, while still outperforming the TCN in capturing non‑linear transient features.

The authors discuss why Bi‑GRU excels in this domain. The bidirectional architecture supplies context from both past and future observations, which is crucial when gaps exist. The simplified gating of GRU cells reduces parameter count without sacrificing the ability to model long‑range dependencies, leading to faster convergence and lower risk of over‑fitting. The extensive preprocessing (log transformation, high‑resolution interpolation) also contributes to the models’ robustness, as it normalizes input scales and supplies dense temporal information.

Limitations and future work are acknowledged. The current approach treats each GRB independently; a multi‑task or transfer‑learning strategy could exploit shared physics across bursts. Incorporating multi‑wavelength data (optical, radio) would enable a truly multimodal reconstruction. Finally, Bayesian neural networks or Monte‑Carlo dropout could provide calibrated uncertainty estimates, which are essential for downstream cosmological analyses.

In conclusion, the study demonstrates that a Bidirectional GRU, combined with careful preprocessing, offers the most accurate and efficient solution for reconstructing GRB plateau‑phase light curves among the evaluated deep‑learning models. This methodology not only fills observational gaps but also creates a reliable, continuous dataset that can enhance physical modeling and improve the utility of GRBs as probes of the high‑redshift universe.