Skip to content

Part 7: Paper Writing Guide -- From Research to Publication

Goal: A practical guide on how to write and publish your ML + astronomy research paper.
Audience: First-time paper writers. No prior publishing experience assumed.

Table of Contents

  1. How Academic Publishing Works
  2. Choosing a Venue
  3. Paper Structure
  4. Writing Each Section
  5. Figures and Visualizations
  6. Common Mistakes to Avoid
  7. The Review Process
  8. Tools and Resources

1. How Academic Publishing Works

The Pipeline

Your Research
     │
     ▼
Write Paper (LaTeX + Figures)
     │
     ▼
Post to arXiv (preprint -- establishes priority)
     │
     ▼
Submit to Journal/Conference
     │
     ▼
Peer Review (1-3 anonymous reviewers)
     │
     ├── Accept → Published!
     ├── Minor Revision → Fix small issues → Accept
     ├── Major Revision → Significant changes → Re-review
     └── Reject → Try another venue

Key Terms

Term Meaning
arXiv Free preprint server. Most astronomy/ML papers go here FIRST, before journal submission
Peer review Anonymous experts evaluate your paper for correctness, novelty, and significance
Open access Paper is freely available to everyone (most astronomy journals are OA)
Impact factor A metric of how often papers in a journal are cited. Higher = more prestigious
DOI Digital Object Identifier -- permanent URL for your published paper
BibTeX Citation format used in LaTeX

The arXiv Strategy

In astronomy and ML, posting to arXiv is standard practice: 1. Write your paper 2. Post to arXiv (this establishes your priority -- proves you did it first) 3. Submit to a journal simultaneously 4. The journal reviews it while it's already publicly available on arXiv 5. Most astronomers read arXiv daily and may cite your paper even before it's formally published

2. Choosing a Venue

Astronomy Journals

Journal Abbreviation Impact Factor Review Time Best For
The Astronomical Journal AJ ~5.5 2-4 months Methods papers, catalogs
The Astrophysical Journal ApJ ~4.8 2-4 months Results-heavy papers
Monthly Notices of RAS MNRAS ~4.8 1-3 months Broad astronomy topics
Astronomy & Astrophysics A&A ~5.4 2-4 months European-leaning
Astronomy & Computing A&C ~1.9 1-3 months ML/computational methods
Publications of the ASP PASP ~3.3 2-3 months Instrumentation + methods

ML Conferences/Workshops

Venue Type Deadline Best For
NeurIPS ML4PS Workshop September ML + Physical Sciences
ICML ML4Astro Workshop March ML for Astrophysics
ICLR Conference October If your ML contribution is strong
AAAI Conference August Applied AI

3. Paper Structure

The standard structure for an ML + astronomy paper:

1. Title
2. Abstract (~200 words)
3. Introduction (1.5-2 pages)
4. Related Work (1-1.5 pages)
5. Data (1-1.5 pages)
6. Methods (2-3 pages)
7. Experiments & Results (2-3 pages)
8. Discussion (1-1.5 pages)
9. Conclusion (0.5-1 page)
10. Acknowledgments
11. References
12. Appendix (optional)

Total: 12-18 pages (typical for AJ/MNRAS)

4. Writing Each Section (Example)

4.1 Title

Formula: [Method Name]: [What it does] for [Domain Application]

Examples: - "XAI-Transit: An Explainable Physics-Informed Neural Network for TESS Exoplanet Classification" - "PhysNet: Physics-Constrained Deep Learning for Transit Signal Vetting in TESS Photometry" - "Explaining the Machine: Interpretable Exoplanet Detection with Multi-Head Neural Networks"

Tips: - Include key differentiators (explainable, physics-informed, multi-head) - Mention the mission (TESS) for discoverability - Keep it under 15 words if possible

4.2 Abstract

Structure (one sentence each):

  1. Problem: "Transit-based exoplanet detection suffers from high false positive rates and opaque ML classifiers."
  2. Gap: "Existing deep learning approaches lack physical grounding and interpretability."
  3. Method: "We present [name], a hybrid CNN + Bi-LSTM + physics-prior network with physics-informed loss functions and multi-level explainability."
  4. Results: "Our model achieves X% precision and Y% recall on [dataset], competitive with ExoMiner, while providing Grad-CAM and attention-based explanations."
  5. Impact: "Explainability analysis reveals the model uses physically meaningful transit features, building trust for deployment in large-scale surveys."

Length: 150-250 words.

4.3 Introduction

Structure (4-5 paragraphs):

  1. Context: Exoplanets are everywhere. TESS is producing massive data. We need automated classification.
  2. Problem: Current methods are either (a) traditional algorithms with limited accuracy or (b) black-box deep learning with limited trust.
  3. What exists: AstroNet, ExoMiner, etc. Brief overview of prior work.
  4. The gap: No existing approach combines (a) hybrid architecture, (b) physics-informed training, and (c) comprehensive explainability.
  5. Our contribution: We present [model name]. Our key contributions are: (1) ..., (2) ..., (3) .... We release our code and trained model.

Tips: - End the Introduction with a clear list of contributions (bullet points) - Don't oversell -- be precise about what's new - Cite relevant prior work even in the Introduction

Organize by category:

  1. Traditional transit detection: BLS (Kovacs 2002), TLS (Hippke & Heller 2019)
  2. Classical ML for vetting: Random Forest (McCauliff 2014), feature engineering approaches
  3. Deep learning for vetting: AstroNet (2018), ExoMiner (2021), ExoMiner++ (2024)
  4. Transformers: Salinas et al. (2023, 2025)
  5. Explainable AI in astronomy: (sparse literature -- note the gap!)
  6. Physics-informed neural networks: PINNs in other domains, limited application in exoplanets

End with: "To our knowledge, no prior work has combined a multi-head hybrid architecture with physics-informed training and comprehensive explainability analysis for transit classification."

4.5 Data

Must include: - What dataset you used (Kepler DR25, TESS TOI catalog, etc.) - How you preprocessed it (normalization, phase folding, view generation) - Train/validation/test split (with exact numbers) - Class distribution (N planets, N false positives) - Any data augmentation (synthetic injection, jitter)

Example table:

| Split     | Planets | False Positives | Total |
|-----------|---------|----------------|-------|
| Train     | 2,500   | 25,000         | 27,500|
| Validation| 500     | 5,000          | 5,500 |
| Test      | 500     | 5,000          | 5,500 |

4.6 Methods

Structure by component:

  1. Architecture overview (diagram + description)
  2. ShapeExpert (CNN): Input, layers, output, what it detects
  3. SequenceExpert (Bi-LSTM): Input, layers, attention, output
  4. PhysicsPrior: Input features, normalization, output
  5. Fusion Head: Concatenation, dense layers, output heads
  6. Physics-informed loss: Standard loss + constraint terms, lambda values
  7. Training details: Optimizer, LR schedule, epochs, batch size, augmentation
  8. Explainability methods: Grad-CAM, attention extraction, SHAP

Include an architecture diagram (most important figure in the paper).

4.7 Experiments & Results

Must include:

  1. Comparison with baselines:
| Model            | Precision | Recall | F1    | AUC-ROC |
|-----------------|-----------|--------|-------|---------|
| Random Forest    | 85.2%     | 80.1%  | 82.6% | 0.92    |
| AstroNet (CNN)   | 90.3%     | 88.5%  | 89.4% | 0.96    |
| Ours (no physics)| 91.7%     | 90.2%  | 90.9% | 0.97    |
| Ours (full)      | 93.1%     | 92.4%  | 92.7% | 0.98    |
  1. Ablation study (remove one component at a time):
| Variant                        | AUC-ROC | Delta |
|-------------------------------|---------|-------|
| Full model                     | 0.98    | --    |
| Without LSTM                   | 0.96    | -0.02 |
| Without Physics Prior          | 0.97    | -0.01 |
| Without Physics-informed Loss  | 0.975   | -0.005|
| CNN only (AstroNet-like)       | 0.96    | -0.02 |
  1. XAI analysis:
  2. Grad-CAM figures showing what the CNN focuses on
  3. Attention maps showing which time steps the LSTM prioritizes
  4. SHAP plots showing physics feature importance

  5. Examples for correctly and incorrectly classified cases

  6. Application to real TESS targets:

  7. Run on your cached targets (WASP-18b, TOI-700, Pi Mensae, EB)
  8. Show the pipeline correctly identifies known planets and rejects the EB

4.8 Discussion

Cover:

  1. What the model learns: XAI analysis shows the model focuses on physically meaningful features (ingress/egress shape, periodicity, depth-radius consistency)
  2. Edge cases: Where the model struggles (grazing transits, blended systems, very shallow signals)
  3. Physics-informed benefits: How the constraint loss improves performance on physically ambiguous cases
  4. Limitations:
  5. Single-sector data limitations (27-day baseline)
  6. No centroid analysis (can't detect background EBs from photometry alone)
  7. Small training set for TESS-specific features
  8. Future work: Transfer learning, anomaly detection, atmospheric prediction

4.9 Conclusion

3-4 sentences: 1. We presented [model name], a [brief description] 2. Key findings: [1-2 most important results] 3. The XAI analysis demonstrates [key insight about what the model learns] 4. Our code and model are publicly available at [GitHub URL]

5. Figures and Visualizations

Must-Have Figures

Figure What It Shows Section
Architecture diagram Full model with 3 heads + fusion Methods
Training curves Loss and accuracy vs. epochs Results
ROC curve Your model vs. baselines Results
Confusion matrix TP, FP, FN, TN breakdown Results
Grad-CAM examples CNN attention on transit shapes Results/XAI
Attention maps LSTM attention over time series Results/XAI
SHAP summary Physics feature importance Results/XAI
TESS target examples Pipeline output for known planets/EBs Results

Figure Design Tips

  • Use matplotlib or plotly for consistency
  • Use a consistent color scheme throughout
  • Always include axis labels with units
  • Use colorblind-friendly palettes (viridis, etc.)
  • Keep figures clean -- no gridlines unless necessary
  • Resolution: 300 DPI minimum for print

Architecture Diagram

Create this using draw.io, tikz (LaTeX), or matplotlib:

┌─────────────────────────────────────────────────────────┐
│                   XAI-Transit Architecture                │
│                                                          │
│  Local View          Global View        Physics Vector   │
│  (1×200)             (1×1000)            (5)            │
│     │                    │                  │            │
│     ▼                    ▼                  ▼            │
│ ┌─────────┐      ┌────────────┐     ┌──────────┐       │
│ │  Shape   │      │  Sequence  │     │ Physics  │       │
│ │  Expert  │      │  Expert    │     │  Prior   │       │
│ │  (CNN)   │      │ (Bi-LSTM)  │     │ (Dense)  │       │
│ └────┬────┘      └─────┬──────┘     └────┬─────┘       │
│      │(64)              │(256)            │(32)          │
│      └─────────────┬────┘────────────────┘              │
│                    │                                     │
│              ┌─────▼─────┐                              │
│              │  Fusion    │                              │
│              │   Head     │                              │
│              └─────┬─────┘                              │
│                    │                                     │
│         ┌─────────┼─────────┐                           │
│         ▼         ▼         ▼                           │
│    confidence  shape_s  temporal_s                       │
│    (0.0-1.0)   (0-1)    (0-1)                          │
│                                                          │
│              Physics-Informed Loss:                       │
│    L = FocalLoss + λ₁·duration_constraint               │
│                  + λ₂·radius_constraint                  │
│                  + λ₃·depth_consistency                  │
└─────────────────────────────────────────────────────────┘

6. Common Mistakes to Avoid

Technical Mistakes

Mistake Why It's Bad How to Fix
Data leakage Test set performance is artificially inflated Ensure strict train/test separation. No data augmentation on test set.
No baseline comparison Reviewers can't assess if your method is actually better Always compare with at least 2 published baselines
No ablation study Can't tell which components actually help Remove one component at a time and measure
Reporting only accuracy Meaningless for imbalanced data Report Precision, Recall, F1, AUC-ROC
Small test set Results aren't statistically significant Use 5-fold cross-validation or bootstrap confidence intervals

Writing Mistakes

Mistake How to Fix
Overclaiming Use hedging language: "suggests," "indicates," not "proves"
Not citing enough Cite EVERY related paper you're aware of. Underciting annoys reviewers
Burying the contribution State your contributions explicitly in bullet points in the Introduction
Too much jargon Define every term the first time you use it
No code release Reviewers value reproducibility. Release code on GitHub

Presentation Mistakes

Mistake How to Fix
Low-resolution figures Use vector graphics (PDF) or 300+ DPI
Inconsistent notation Define notation once and use consistently
Wall of text Use tables, figures, and equations to break up text
Missing error bars Report standard deviation or confidence intervals

7. The Review Process

What Reviewers Look For

  1. Novelty: Is this genuinely new? Or just a minor tweak of existing work?
  2. Correctness: Is the methodology sound? Any errors in the math/code?
  3. Completeness: Are baselines included? Ablation study? Error analysis?
  4. Significance: Does this matter to the community?
  5. Clarity: Is it well-written and easy to follow?
  6. Reproducibility: Is there enough detail to replicate the results?

How to Respond to Reviews

When you get reviews back:

  1. Don't take it personally. Reviewers are doing their job.
  2. Create a point-by-point response to each comment
  3. Be respectful even if you disagree
  4. Make the changes or explain clearly why you didn't
  5. Highlight changes in the revised manuscript (use colored text)

Example Review Response

Reviewer 1, Comment 3: "The authors should compare with ExoMiner 
on the same dataset."

RESPONSE: Thank you for this suggestion. We have added a comparison 
with ExoMiner using the Kepler DR25 test set in Table 3 (Section 5.1). 
Our model achieves comparable precision (93.1% vs 95%) with improved 
recall (92.4% vs 93%) while additionally providing explainability 
features that ExoMiner lacks. The relevant text has been updated 
in Section 5.1, paragraph 2 (highlighted in blue).

8. Tools and Resources

Writing Tools

Tool Purpose
Overleaf Online LaTeX editor (collaborative, free tier available)
LaTeX The standard typesetting system for scientific papers
BibTeX Citation management within LaTeX
draw.io Free diagram tool for architecture figures
Zotero Free reference manager (imports from arXiv, ADS)

Scientific Python

Library Purpose
matplotlib Publication-quality figures
seaborn Statistical visualization
scikit-learn Baseline ML models, metrics
shap SHAP explainability analysis
captum PyTorch model interpretability (Grad-CAM, etc.)
pytorch-grad-cam Grad-CAM implementation

Astronomy-Specific

Resource Purpose
NASA ADS Search for astronomy papers
arXiv astro-ph.EP Exoplanet preprints
lightkurve TESS/Kepler data access
exoplanet.eu Confirmed planet database
AASTeX LaTeX template for AAS journals (AJ, ApJ)
MNRAS LaTeX class Template for MNRAS

Key Takeaways

  1. Post to arXiv first -- establishes priority and gets community feedback
  2. Target AJ or MNRAS for the full paper
  3. Write Methods first (you know it best), then Results, then Intro, then Abstract last
  4. Every claim needs evidence: comparison tables, ablation study, statistical significance
  5. Release your code -- reviewers and the community value reproducibility
  6. Start with the evaluation infrastructure -- you can't write results without numbers

Congratulations!

You've completed the entire learning guide. Here's your path forward:

YOU ARE HERE
     │
     ▼
┌─────────────────────────────────────┐
│  1. Read the recommended papers     │
│  2. Set up evaluation infrastructure│
│  3. Run experiments                 │
│  4. Write the paper                 │
│  5. Submit!                         │
└─────────────────────────────────────┘

Go find some exoplanets.