Skip to content

Part 5: Research Landscape -- What's Been Done and Where Are the Gaps

Prerequisites: Parts 1-4.
What you'll learn: Key published papers, the current state of the art, what has NOT been done yet.

Table of Contents

  1. Timeline of Key Papers
  2. Foundational Papers
  3. State of the Art (2024-2025)
  4. What Has Been Done
  5. The Gaps -- What's Missing
  6. Reading List (Priority Order)

1. Timeline of Key Papers

2014  ──  First ML applied to Kepler (McCauliff et al.) - Random Forest
          │
2017  ──  Autovetter (Catanzarite 2017) - Automated transit vetting
          │
2018  ──  AstroNet (Shallue & Vanderburg) - CNN, validated 2 new planets  ◄── LANDMARK
          │
2018  ──  TESS launches
          │
2019  ──  Ansdell et al. - CNN on Kepler with global+local view
          │
2020  ──  Osborn et al. - Random Forest for TESS (NGTS follow-up)
          │
2021  ──  ExoMiner (Valizadegan et al.) - NASA's ML vetter         ◄── LANDMARK
          │    [Validated 301 new Kepler planets]
          │
2022  ──  ExoMiner v2 - Multi-task learning, diagnostic features
          │
2023  ──  Salinas et al. - Transformers for TESS FFIs               ◄── EMERGING
          │
2024  ──  ExoMiner++ - Transfer learning Kepler→TESS                ◄── LATEST
          │    [First large-scale ML catalog for TESS 2-min cadence]
          │
2024  ──  Gorchakova & Creaner - Synthetic augmentation for transits
          │
2025  ──  CNN+RNN hybrids for Kepler DR25
          │
2025  ──  Trehan et al. - Bayesian ML for atmospheric prediction

2. Foundational Papers

Paper 1: AstroNet (Shallue & Vanderburg, 2018)

Title: "Identifying Exoplanets with Deep Learning: A Five Planet Resonant Chain around Kepler-80 and an Eighth Planet around Kepler-90"

Why it matters: This was the paper that proved deep learning could find planets that humans missed. Google Brain + Harvard collaboration.

Architecture: 

Global View (2001 pts) → CNN → Feature Vector ─┐
                                                 ├── Dense Layers → Planet Probability
Local View (201 pts) → CNN → Feature Vector  ───┘

Key contributions: - Dual-view approach (global + local) - Validated Kepler-90i (8th planet in a system!) and Kepler-80g - Showed CNNs outperform hand-crafted features

Limitations: - No physics injection (purely data-driven) - No temporal modeling (CNN only, no LSTM) - No attention mechanism - Trained and tested only on Kepler

Paper 2: ExoMiner (Valizadegan et al., 2021)

Title: "ExoMiner: A Highly Accurate and Explainable Deep Learning Classifier That Validates 301 New Exoplanets"

Why it matters: NASA's official ML-based planet vetter. Used to validate more planets than any other ML system.

Architecture: 

Transit View → CNN ──────────────────────┐
Odd/Even View → CNN ─────────────────────┤
Secondary Eclipse View → CNN ────────────┤
                                          ├── Dense → Planet Prob
Diagnostic Features → Dense ─────────────┤
  (centroid shift, crowding, etc.)       │
Stellar Parameters → Dense ──────────────┘

Key contributions: - Multi-input design with multiple diagnostic views - Used centroid analysis (spatial information) -- detects background EBs - Explainability through per-view scores - Validated 301 new confirmed planets

Limitations: - Kepler-specific (different noise than TESS) - No temporal (LSTM) component - Complex pipeline with many specialized inputs

Paper 3: ExoMiner++ (2024)

Title: Transfer learning from Kepler to TESS

Key contribution: Showed that models trained on Kepler can be adapted to TESS through fine-tuning, producing the first large-scale ML vetting catalog for TESS 2-minute cadence data.

Limitation: Transfer wasn't perfect -- TESS's shorter baselines and different systematics cause performance degradation.

3. State of the Art (2024-2025)

What the Best Systems Look Like

System Architecture Data Precision Recall Unique Feature
ExoMiner++ Multi-CNN Kepler→TESS ~95% ~93% Transfer learning
Salinas 2023 Transformer TESS FFI ~90% ~88% Self-attention
CNN+RNN 2025 CNN + BiLSTM Kepler DR25 ~92% ~95% Hybrid temporal

Current Benchmarks

On Kepler DR25 test set: - ExoMiner: AUC-ROC = 0.98, Precision = 95%, Recall = 93% - AstroNet: AUC-ROC = 0.96, Precision = 90%, Recall = 88% - Random Forest (baseline): AUC-ROC = 0.92, Precision = 85%, Recall = 80%

On TESS data (less standardized benchmarks): - ExoMiner++: Precision ~93%, Recall ~90% (fine-tuned) - Most other models haven't been rigorously benchmarked on TESS

4. What Has Been Done

Well-Explored Areas

Area Status What's been done
CNN for transit classification Mature AstroNet, ExoMiner, dozens of papers
Random Forest for transit vetting Mature Multiple papers, well-understood
BLS for signal detection Very mature Standard algorithm, widely used
Kepler data classification Saturated Little room for improvement
Hot Jupiter detection Saturated Easy targets, all found
Synthetic transit generation Growing Several papers on physics-informed generation
Transfer learning (Kepler→TESS) Active ExoMiner++ leading, still room for improvement

What This Means for us

Basic exoplanet detection with ML is well-trodden ground. A paper that just says "we used a CNN to classify TESS transits" will struggle to get published because dozens of groups have done similar work.

We need a unique angle.

5. The Gaps -- What's Missing

These are the areas where our research could make a genuine contribution:

Gap 1: Explainable AI (XAI) for Transit Classification

The problem: Deep learning models are black boxes. Astronomers don't trust classifications they can't understand. A model says "planet with 95% confidence" -- but WHY?

Current state: ExoMiner has per-view scores (some explainability). But no one has done deep XAI analysis (Grad-CAM, SHAP, attention visualization) specifically for transit classifiers.

Opportunity: Can make model which has component scores (shape_score, temporal_score). and could extend this with: - Grad-CAM on the CNN to visualize which parts of the transit shape the model focuses on - Attention weight visualization from the LSTM to show which time steps matter - SHAP values for the physics features to show how BLS parameters influence the decision

Novelty level: HIGH. No paper has done comprehensive XAI for a hybrid transit classifier.

Gap 2: Physics-Informed Neural Networks (PINNs) for Transit Detection

The problem: Most ML models are purely data-driven -- they learn from labeled examples but don't "know" physics. When they encounter unusual situations (rare planet types, unusual stars), they fail silently.

Opportunity: - Physics-informed loss functions (penalize predictions that violate Kepler's laws) - Embedding transit models (Mandel & Agol 2002) as differentiable layers - Using physics constraints during inference (reject candidates that violate physical laws, even if the CNN says "planet")

Novelty level: HIGH. True PINNs haven't been applied to transit classification.

Gap 3: Multi-Mission Transfer Learning

The problem: Models trained on Kepler don't work well on TESS (different noise, cadence, baseline). Future missions (PLATO, Roman) will have yet different characteristics.

Current state: ExoMiner++ showed basic transfer learning works. But the methodology is ad-hoc.

Opportunity: Systematic study of domain adaptation techniques: - Feature alignment between missions - Adversarial domain adaptation - Multi-task learning across missions simultaneously - Zero-shot transfer to PLATO (before it launches)

Novelty level: MEDIUM-HIGH.

Gap 4: Anomaly Detection for Novel Phenomena

The problem: All current classifiers are trained on KNOWN categories (planet vs. EB). But what about phenomena we haven't categorized yet?

Current state: Almost no work on unsupervised/semi-supervised anomaly detection in TESS light curves.

Opportunity: - Autoencoders trained on "normal" light curves -- flag anything unusual - Potential discoveries: exomoons, exocomets, disintegrating planets, Boyajian's-Star-like dippers

Novelty level: HIGH. Very few papers on this for TESS.

Gap 5: Atmospheric Prediction from Photometry

The problem: Can we predict what a planet's atmosphere looks like BEFORE JWST observes it, using just transit photometric data?

Current state: Trehan et al. (2025) started this with a Bayesian ML framework, but it's very early.

Opportunity: Can use transit light curve features (depth, duration, wavelength-dependent depth if available) to predict atmospheric properties (composition, cloud coverage, scale height).

Novelty level: VERY HIGH. Extremely new area.

Gap 6: Robust Evaluation and Benchmarking

The problem: Every paper uses different datasets, different preprocessing, different train/test splits, and different metrics. It's nearly impossible to fairly compare methods.

Current state: No standardized benchmark exists for TESS transit classification.

Opportunity: Create a standardized benchmark dataset and evaluation protocol for TESS. This would be a valuable community resource paper.

Novelty level: MEDIUM (but HIGH impact).

6. Reading List (Priority Order)

Must-Read (Start Here)

  1. Shallue & Vanderburg (2018) - "Identifying Exoplanets with Deep Learning"
  2. The paper that started it all.

  3. arXiv: 1712.05044

  4. Valizadegan et al. (2022) - "ExoMiner: A Highly Accurate and Explainable Deep Learning Classifier"

  5. NASA's state of the art.

  6. arXiv: 2111.10009

  7. Pearson et al. (2018) - "Searching for Exoplanets using Artificial Intelligence"

  8. Good overview of ML approaches for transit detection.
  9. arXiv: 1706.04319

Important Context

  1. Kovacs et al. (2002) - "A box-fitting algorithm in the search for periodic transits"

  2. The BLS algorithm paper.

  3. Kempton et al. (2018) - "A Framework for Prioritizing the TESS Planetary Candidates Most Amenable to Atmospheric Characterization"

  4. TSM formula.

  5. Kopparapu et al. (2013) - "Habitable Zones Around Main-Sequence Stars"

  6. Habitable zone calculation.

Cutting Edge (For Our Research Angle)

  1. ExoMiner++ (2024) - Transfer learning Kepler→TESS

  2. arXiv: 2601.14877

  3. Salinas et al. (2023/2025) - "Exoplanet Transit Candidate Identification via a Transformer-Based Algorithm"

  4. Transformers for TESS.

  5. Trehan et al. (2025) - Bayesian ML for atmospheric prediction

  6. New frontier.

  7. Gorchakova & Creaner (2024) - Synthetic augmentation for transit detection

For XAI Direction

  1. Selvaraju et al. (2017) - "Grad-CAM: Visual Explanations from Deep Networks"

    • The Grad-CAM technique.

  2. Lundberg & Lee (2017) - "A Unified Approach to Interpreting Model Predictions" (SHAP)

    • SHAP values for feature importance.

How to Find Papers

  • arXiv (arxiv.org): Preprint server. All astronomy and ML papers appear here first. Search for "exoplanet machine learning" or "transit classification deep learning."
  • NASA ADS (ui.adsabs.harvard.edu): The astronomy-specific paper search engine. More focused than Google Scholar.
  • Google Scholar: Good for finding citation counts and related papers.
  • Semantic Scholar: Good for finding the most influential papers in a topic.

Key Takeaways

  1. CNN-based transit classification is mature -- we can't publish "just another CNN" paper
  2. The biggest gaps are in XAI, physics-informed ML, anomaly detection, and multi-mission transfer
  3. We need a clear unique angle and rigorous evaluation to publish
  4. Read the AstroNet and ExoMiner papers first -- they're our direct predecessors

What's Next?

Part 6: Research Directions -- Detailed analysis of each potential research direction, with feasibility assessment and implementation roadmap.