⭐ GRB Light Curve Reconstruction – NAOJ Winter Research Internship (2024–25)

Author: Debanik Debnath
Institute: National Astronomical Observatory of Japan (NAOJ)
Mentor: Prof. Maria Giovanna Dainotti (SOKENDAI / NAOJ)
Duration: December 2024 – February 2025 (Remote)

---

🔭 Overview

This repository documents my research contributions during my Winter Internship at the National Astronomical Observatory of Japan (NAOJ). The project focused on reconstructing Gamma-Ray Burst (GRB) X-ray afterglow light curves using:

• Classical statistical models
• Machine learning regression
• Deep learning architectures (LSTM, Bi-LSTM, GRU, CNN-LSTM, Transformers)
• Hybrid and probabilistic pipelines

The objective was to improve reconstruction fidelity and extract key astrophysical parameters such as:

• Plateau end time (log Tₐ)
• Flux at plateau (log Fₐ)
• Temporal decay slopes
• Light-curve smoothness and discontinuities

This work contributes to a broader cosmological effort to evaluate GRBs as potential standardizable candles for studying the expansion of the high-redshift universe.

---

📚 Primary Scientific References

1. Gamma-Ray Burst Light Curve Reconstruction (arXiv)

https://arxiv.org/abs/2412.20091 A multi-model framework for reconstructing GRB light curves using machine and deep-learning approaches. This includes models closely related to my LSTM / Bi-LSTM pipeline designs.

2. Dainotti et al. (2023) – ApJS 267, 42

https://ui.adsabs.harvard.edu/abs/2023ApJS..267...42D/abstract A foundational study on GRB X-ray afterglows, plateau correlations, and cosmological implications. This paper heavily influenced the modeling direction of my internship.

3. Dainotti et al. – GRB Standardization Work

https://ui.adsabs.harvard.edu/abs/2023ApJS..267...42D/abstract Further methodological inspiration for GRB classification, data preprocessing, and statistical handling.

---

🎯 My Key Contributions

✔ Processed and analyzed 200+ GRB light curves

• Trimming, cleaning, normalization, time-binning, and multi-epoch comparison.

✔ Built multiple deep-learning architectures

• LSTM (baseline + optimized)
• Bi-LSTM (best-performing, bidirectional temporal context)
• GRU
• 1D-CNN
• CNN-LSTM hybrids
• Autoencoders & Denoising AEs
• Transformer-based experiments

✔ Classical & Probabilistic Techniques

• Gaussian Process Regression (GPR)
• Kalman Filtering
• Cubic / B-spline interpolation

✔ Optimization & Evaluation

• Hyperparameter tuning with Optuna
• Cross-validation
• Loss-surface visualization
• MSE / MAE / RMSE comparison across dozens of GRB IDs

✔ Documentation & Manuscript Contribution

• Prepared detailed plots, recorded performance metrics, and contributed text and figures to the collaborative draft.

---

🧠 Model Pipeline Summary

🔹 Classical Stage

• Spline interpolation
• Gaussian Processes
• Kalman filters

🔹 Machine Learning

• k-NN regression
• Polynomial & piecewise regressors

🔹 Deep Learning

• LSTM / GRU sequence models
• Bidirectional LSTM for dual-direction temporal context
• Transformer encoders for long-range dependencies
• Autoencoders for noise reduction

🔹 Hybrid Models

• CNN feature extractor + LSTM decoder
• Kalman residual correction applied to DNN outputs

---

📂 Repository Structure (Beta)

GRB-LightCurve-Reconstruction-NAOJ/
│
├── code/
│   ├── lstm_template.ipynb
│   ├── bilstm_template.ipynb
│   ├── gru_template.ipynb
│   ├── gp_regression_template.ipynb
│   └── placeholder.txt
│
├── datasets/
│   └── placeholder.txt
│
├── plots/
│   └── placeholder.txt
│
├── results/
│   └── placeholder.txt
│
└── report/
    └── Internship_Report_Placeholder.pdf

(Folders include placeholders until original files are recovered.)

---

📝 Internship Report

The complete detailed report, code summaries, and model comparisons are included in the /report/ directory (placeholder added for now).

---

🤝 Acknowledgements

I sincerely thank Prof. Maria Giovanna Dainotti for her guidance, mentorship, and support throughout this project. Her expertise in GRB astrophysics and machine learning made this internship a transformative learning experience for me.

---

⭐ Future Plans

• Re-upload lost GRB result files once recovered
• Add new Transformer-based reconstructions
• Extend GRB forecasting to prompt + afterglow multi-band models
• Publish a cleaned dataset wrapper for reproducible training