Astronomy is entering a new era where telescopes and detectors are only half the story — the other half is intelligence. Artificial intelligence (AI) is transforming the way we operate telescopes, process images and spectra, and translate petabytes of noisy data into robust discoveries: new galaxies at the edge of detectability, faint dwarf satellites, transient events, and candidate habitable exoplanets. AI accelerates discovery (by automating routine tasks and surfacing anomalies), improves fidelity (by denoising and calibrating data beyond classical methods), enables autonomy (real-time adaptive observing and target prioritization), and expands scientific reach (simulating and inverting complex physical processes).

This long-form article explains how AI is being applied across the entire telescope lifecycle — from instrument control and adaptive optics to image formation, transient detection, galaxy classification, and exoplanet discovery — and outlines the concrete technical methods, limitations, evaluation metrics, operational practices, and a roadmap for researchers, observatories, and funders. It covers core AI architectures (CNNs, transformers, GNNs, probabilistic models, physics-informed networks), real-world pipelines (detection → vetting → follow-up), infrastructure (data lakes, federated learning, edge inference), and critical issues (bias, interpretability, reproducibility, ethical data sharing). Where relevant, we offer practical examples, pseudo-algorithms, and implementation patterns so teams can move from concept to operational systems.

If your goal is to design, deploy, or fund an AI-enabled telescope program — whether for a ground-based survey, a space telescope, or a network of smaller instruments — this article functions as a technical primer, strategic playbook, and policy brief in one.

1. Why AI matters for modern telescopes

1.1 The scale and the opportunity

Modern astronomical facilities produce enormous, heterogeneous data streams:

Wide-field synoptic surveys produce many terabytes per night and revisit the same sky repeatedly (enabling fast transient detection).
High-resolution imagers and spectrographs deliver dense spatial and spectral channels.
Space missions generate deep, precise photometry and spectroscopy across thousands of targets.

These instruments push observational sensitivity to limits where signals are faint, systematics (instrumental effects, scattered light, atmospheric turbulence) are complex, and rare events are buried in noise. Classical pipelines — hand-tuned calibrations, aperture photometry, heuristics for candidate selection — struggle to scale, miss subtle patterns, and are brittle to instrument changes. AI complements physics-based models: it learns patterns in data, speeds up complex inversions, reduces human workload, and offers probabilistic outputs that better quantify uncertainties.

1.2 New science enabled by AI

AI-powered telescopes make possible:

Deeper galaxy detection by denoising low-SNR images and identifying low-surface-brightness features (stellar halos, tidal streams).
Faster and more reliable transient discovery, to catch kilonovae, supernova shock breakouts, and short-lived electromagnetic counterparts to gravitational waves.
Scalable exoplanet discovery and vetting: sift millions of light curves for transit signals, identify false positives, and prioritize follow-up.
Efficient spectroscopic extraction and parameter inference, turning noisy spectra into stellar/galaxy redshifts and atmospheric compositions.
Autonomous telescope operations, maximizing science return by optimizing schedules and adapting to weather, seeing, or incoming alerts.

2. End-to-end architecture of an AI-powered telescope system

A robust AI-enabled observatory is a tightly integrated stack. The following end-to-end architecture shows where AI is most impactful.

2.1 Data acquisition and edge preprocessing

Sensors & detectors: CCDs, CMOS, infrared arrays, integral-field units (IFUs), photon-counting detectors.
Edge AI: Basic denoising, cosmic-ray removal, frame coaddition, and lossless compression are performed close to the instrument to reduce latency and data volume.

2.2 Calibration & image formation

Classical steps: Bias/dark subtraction, flat-fielding, wavelength calibration.
AI enhancements: Super-resolution reconstruction, PSF modeling with neural nets, and data-driven removal of instrumental systematics.

2.3 Detection & cataloging layer

Object detection: CNNs and matched-filtering hybrids detect sources (point-like and extended).
Transient/variability detection: Time-series models identify changes across epochs.
Catalog creation: Probabilistic catalogs (positions, fluxes, classification probabilities) with uncertainty propagation.

2.4 Science inference & interpretation

Morphological classification: CNNs and graph-based models classify galaxies, identify mergers, and measure structural parameters.
Spectral analysis: Deep learning and Bayesian inference convert spectra into redshifts, chemical abundances, and physical parameters.
Exoplanet pipelines: Transit and radial velocity detection models, vetting classifiers, and atmospheric retrieval engines.

2.5 Operations & decision layer

Autonomous scheduler: Reinforcement-learning or optimization-based agents schedule observations, react to weather/seeing, and prioritize follow-ups of candidate events.
Alert broker: Real-time event brokers (with ML ranking) distribute high-value alerts to follow-up facilities.
Human-in-the-loop dashboard: Prioritized candidate lists with explanations for astronomers to inspect and confirm.

2.6 Data management & model lifecycle

Data lake & metadata: Well-structured storage with provenance and experiment metadata.
Model registry & CI/CD: Versioned models with reproducible training pipelines, continuous integration, and evaluation on held-out benchmarks.
Federated learning & privacy controls: Across consortiums and distributed sensors when raw data cannot be centralized.

3. Core AI techniques and why they are chosen

This section maps common astronomy tasks to AI architectures and explains trade-offs.

3.1 Convolutional Neural Networks (CNNs)

Strengths: Exceptional for image tasks — source detection, morphological classification, deblending, PSF estimation.

Use cases:

Detecting faint galaxies and low-surface-brightness features.
Classifying galaxy morphology (spiral vs elliptical, merger signatures).
Deblending overlapping objects via U-Net-style segmentation.

Caveats: Require well-labeled images for supervised training; domain shift (e.g., instrument changes) can hurt performance.

3.2 Recurrent Neural Networks and Temporal Models (LSTMs, GRUs, Transformers)

Strengths: Modeling sequences and time series.

Use cases:

Transit detection in unevenly sampled light curves.
Modeling AGN variability and periodic signals.
Forecasting systematics or instrument drift.

Modern preference: Transformers and attention-based models often outperform older RNNs for long-range dependencies.

3.3 Autoencoders, VAEs, and Flow Models

Strengths: Unsupervised representation learning, anomaly detection, generative simulation.

Use cases:

Learning latent representations of galaxy or stellar spectra.
Detecting anomalies (novel transients, unexpected artifacts) via reconstruction error.
Simulating realistic instrument noise and synthetic datasets for training.

3.4 Graph Neural Networks (GNNs)

Strengths: Modeling relational structures.

Use cases:

Modeling galaxy clustering and environmental relationships.
Deblending where object relationships matter.
Cross-matching catalogs across wavelengths with relational constraints.

3.5 Probabilistic Models & Bayesian Inference

Strengths: Provide principled uncertainty estimates and priors.

Use cases:

Spectral parameter estimation and atmospheric retrieval for exoplanets.
Probabilistic catalog generation where source existence is uncertain.
Combining multiple uncertain observations into a posterior.

Hybrid approach: Physics-informed neural networks combine differentiable physical models with learned corrections.

3.6 Reinforcement Learning (RL)

Strengths: Learning policies for sequential decision-making.

Use cases:

Autonomous scheduling and adaptive exposure control.
Active learning to select optimal follow-up observations.

Caveats: Requires careful reward design and simulated environments for training.

4. Detecting new galaxies: signals, challenges, and AI solutions

Discovering faint galaxies, dwarfs, and diffuse structures is a major scientific frontier. AI improves sensitivity and fidelity in several ways.

4.1 Scientific goals and observational signatures

High-redshift galaxies: Extremely faint, redshifted light, often visible only in stacked or deep exposures.
Low-surface-brightness (LSB) features: Stellar halos, tidal tails, and dwarf satellites have surface brightness near or below classical detection thresholds.
Diffuse intra-cluster light: Extended emission requiring careful background modeling.

4.2 Key challenges

Low SNR and background subtraction: Sky background, scattered light, and detector systematics dominate.
Blended sources: High source density leads to overlap of point sources and extended galaxies.
Instrumental signatures: Stray light, ghost images, and spatially varying PSFs produce spurious features.
Selection bias and completeness estimation: Estimating what fraction of a population is missed is nontrivial.

4.3 AI techniques and pipelines

4.3.1 Data-driven background and PSF modeling

Neural networks model spatially varying backgrounds and PSFs better than global polynomial fits. Architectures often use multi-scale U-Nets that take raw frames plus instrument telemetry and output per-pixel background/PSF maps. The AI-corrected maps are used to subtract background and perform matched-filter detections.

4.3.2 Super-resolution and coaddition

AI-based image coaddition (learned stacking) improves depth by reconstructing high-fidelity images from multiple dithered frames, accounting for sub-pixel shifts and variable seeing. Super-resolution networks can recover morphological detail from undersampled detectors.

4.3.3 Deblending with segmentation networks

U-Net variants and instance segmentation models (Mask R-CNN style) separate overlapping sources probabilistically. Modern deblenders output multiple hypotheses with uncertainty on flux partitioning.

4.3.4 Probabilistic catalogs and completeness estimation

Rather than a binary catalog, probabilistic catalogs assign posterior existence probabilities and flux distributions. Inject-and-recover simulations powered by generative models quantify completeness and selection functions in the presence of learned processing.

4.3.5 Near real-time discovery of faint transient galaxies

Transient surveys may detect rapid changes that indicate starbursts or transient phenomena in low-mass hosts. Time-aware models combining image differencing with recurrent networks increase sensitivity to sudden low-level flux changes.

4.4 Measuring and mitigating biases

AI can inadvertently learn instrument artifacts as features of galaxies. Strategies to mitigate:

Train with realistic simulations that span instrument conditions.
Use domain randomization during training (vary PSF, background, noise).
Evaluate on held-out real datasets and perform injection-recovery tests.
Calibrate catalogs using forward modeling where synthetic galaxies are injected into raw frames and processed end-to-end.

5. Hunting habitable planets: AI in exoplanet detection and characterization

Exoplanet science benefits immensely from AI across detection, vetting, and atmospheric retrieval.

5.1 Data modalities and signals

Transit photometry: Periodic dips in brightness when a planet crosses its star.
Radial velocity (RV): Doppler wobble of the host star due to planetary mass.
Direct imaging: High-contrast imaging suppressed stellar light to reveal planets.
Astrometry: Tiny position shifts of a star due to planets.
Transit spectroscopy: Wavelength-dependent transit depth gives atmospheric signatures.

Each modality has its own noise sources and detection challenges.

5.2 Transit detection: from raw light curves to candidates

5.2.1 Traditional approach

Matched filters (Box Least Squares, Lomb-Scargle), manual vetting, and rule-based filters. Powerful but labor-intensive and often fails on noisy or irregularly sampled data.

5.2.2 AI-augmented pipelines

Denoising and systematics removal: CNNs and recurrent networks model stellar variability and instrument systematics (e.g., spacecraft pointing jitter), producing cleaned light curves.
Transit search: Deep-learning classifiers (1D CNNs, Transformers) detect candidate signals at low SNR and are robust to nonstationary noise.
Vetting & false-positive reduction: ML classifiers trained on labeled examples distinguish planetary transits from eclipsing binaries, instrumental artifacts, and stellar activity. Feature sets include transit shape metrics and centroid movement; deep models can learn these directly.
Period-finding & folding: AI speeds up identification of the true period in the presence of aliasing and data gaps.

5.2.3 Example approach (high-level pseudo-workflow)

Preprocess light curve: remove outliers, detrend using learned basis functions.
Run transit search via learned matched filter or CNN scanning window.
For each candidate, compute classification score and posterior probability.
Apply odd/even transit checks and centroid motion tests.
Prioritize high-probability candidates for follow-up spectroscopy or higher-cadence photometry.

5.3 Radial velocity and stellar noise modeling

RV signals are small and contaminated by stellar activity (spots, granulation). Gaussian Processes (GPs), combined with physical spot models and deep learning, are used to separate activity-induced signals from planetary signals. Recent approaches use multi-output GPs and neural networks trained on spectroscopic indicators to jointly model activity and RV variations.

5.4 Direct imaging: speckle suppression and planet detection

Direct imaging at high contrast requires removing quasi-static speckles due to imperfect optics. AI-based post-processing:

PSF subtraction with deep learning: Networks learn the mapping from raw coronagraphic images to residuals, reducing speckle noise beyond classical angular differential imaging (ADI).
Anomaly detection: Autoencoders trained on speckle patterns flag deviations consistent with point sources.
Debiasing: Careful injection-recovery is needed: deep models can overfit and remove real planet signals if not trained properly.

5.5 Atmospheric retrieval & habitability assessment

Once a transit or direct spectrum is obtained, atmospheric retrieval inverts spectra to molecular abundances and temperature profiles. Traditional Bayesian retrievals (nested sampling, MCMC) are computationally heavy. AI options:

Emulators: Neural networks emulate radiative-transfer forward models, enabling orders-of-magnitude speedups in posterior sampling.
Conditional density estimators: Normalizing flows and neural posterior estimators directly learn the posterior distribution of atmospheric parameters conditioned on spectra.
Physics-informed constraints: Enforce chemical equilibrium or photochemistry via hybrid models to prevent unphysical posteriors.

AI accelerates the exploration of parameter space, enabling near-real-time interpretation of candidate habitable signatures (water vapor, oxygen, methane in disequilibrium) and better prioritization for expensive follow-up.

6. Transients and multi-messenger astronomy: ranking what to follow

Detecting the electromagnetic counterparts of gravitational waves, neutrinos, or unknown transients demands immediate prioritization: telescopes cannot follow every candidate.

6.1 The real-time challenge

Surveys produce thousands of transient alerts per night. The community needs systems that:

Rapidly rank alerts by scientific value (probability of being a kilonova, early supernova, or exotic event).
Recommend optimal follow-up exposures, instruments, and scheduling across distributed facilities.

6.2 AI ranking and broker architectures

Feature extraction: For each alert, compute light-curve features, colors, host-galaxy associations, and contextual metadata (galactic latitude, proximity to known variable sources).
Classifier ensembles: Trained on labeled historical events, ensembles estimate probabilities for classes (supernova Ia, CC, GRB afterglow, AGN flare).
Active learning & uncertainty-aware ranking: Rank by expected information gain for follow-up (not just by class probability). RL or acquisition functions pick observations that maximally reduce overall uncertainty across the event population.

6.3 Coordinated follow-up

Alert brokers push prioritized lists to telescopes. Autonomous schedulers (Section 8) accept high-priority targets and execute observations, adjusting priorities as new data arrives. AI balances usage across the network: variety of filters, depth vs. cadence trades, and multi-wavelength coverage.

7. Instrument-level AI: adaptive optics, noise reduction, and instrument calibration

AI improves raw image quality and effective angular resolution.

7.1 Adaptive optics (AO)

AO corrects atmospheric turbulence in real time. AI contributes by:

Wavefront prediction: Predict future wavefront distortions from past measurements using RNNs or transformers, reducing latency and improving correction performance.
Sensor fusion: Combining multiple wavefront sensor modalities with ML increases robustness under low SNR guide-star conditions.
Control law learning: Data-driven controllers (imitation learning or RL) can optimize deformable mirror commands to maximize Strehl ratio.

7.2 Denoising and deconvolution

Plug-and-play priors and deep denoisers: Use learned denoisers as priors within optimization-based deconvolution frameworks, providing state-of-the-art restoration while controlling overfitting.
Blind deconvolution with neural nets: Estimate PSF and true scene jointly using supervised or self-supervised strategies.

7.3 Detector calibration and cosmic-ray removal

Deep models trained on detector telemetry can predict pixel-level systematics (non-linearity, persistence) and remove them more accurately than traditional look-up tables. They also excel at classifying and removing cosmic-ray hits in single exposures.

8. Telescope operations: scheduling, autonomy, and cost efficiency

Telescopes are resources — AI improves their yield.

8.1 Autonomous scheduling

Reinforcement learning and constrained optimization-based schedulers ingest scientific priorities, weather forecasts, instrument constraints, and transient alerts to produce a schedule maximizing science utility. Key features:

Robustness to uncertainty: Schedulers plan under uncertainty (weather, seeing) and re-optimize continuously as conditions change.
Multi-tenant optimization: For shared facilities, AI balances competing programs fairly and optimizes total science return.
Adaptive exposure control: Adjust exposure times and filter selection in real time based on instantaneous SNR estimation and science needs.

8.2 Fault detection and predictive maintenance

AI analyzes telemetry to detect anomalies in actuators, tracking systems, or cooling loops, predicting failures before they occur and scheduling maintenance opportunistically to minimize downtime.

8.3 Cost-aware operations

For large survey facilities, marginal cost per observation is nonzero (power, storage, compute). AI optimizes trade-offs: depth vs. area, cadence vs. sensitivity, and uses dynamic reallocation to respond to high-value events.

9. Data infrastructure, model ops, and reproducibility

AI is only as good as the data and the engineering around it.

9.1 Data provenance and metadata

Store processing provenance (raw frames, calibration steps, model versions) so scientific results are reproducible. Data schemas should include timestamps, instrument configuration, calibration files, and environmental telemetry.

9.2 Model governance and lifecycle

Model registry: Versioned models with evaluation metrics and data lineage.
Continuous evaluation: Monitor model performance on holdout and newly acquired data to detect drift.
Reproducible training recipes: Containerized training with pinned dependencies and random seeds.

9.3 Benchmarks and open challenges

Public benchmarks (light-curve datasets, simulated images, spectrum libraries) accelerate progress and improve comparability of methods. Organize blind challenges for detection, deblending, and retrieval tasks.

9.4 Federated and cross-facility collaboration

When raw data sharing is limited, federated learning lets observatories collaboratively improve models while preserving data locality. Carefully designed privacy and security measures are required.

10. Human-AI interaction: interpretability, interfaces, and community workflows

Scientific discovery is social; AI must integrate with human expertise.

10.1 Explainability and trust

Astronomers must understand why an AI flagged a candidate. Techniques:

Saliency maps showing image regions contributing to a detection.
Counterfactual explanations (how small input changes would alter the prediction).
Model cards documenting training data, intended use, limitations, and performance.

10.2 Visualization and decision support

Interactive dashboards with ranked candidates, uncertainty bands, and suggested follow-up steps enable efficient human-in-the-loop validation.

10.3 Citizen science and augmentation

Platforms that surface AI-suggested candidates to citizen scientists (with curated UI) accelerate vetting and provide labeled training data. AI can learn from human corrections in active learning loops.

11. Failure modes, biases, and how to guard against them

AI introduces new risks that can compromise science if unaddressed.

11.1 Dataset bias and selection effects

Training data often reflects instrument-specific biases, astrophysical selection effects, and historical discovery patterns. Mitigations:

Use simulated data to cover rare conditions.
Inject synthetic signals to measure sensitivity.
Calibrate selection functions through forward modeling.

11.2 Overfitting to systematics

Deep models can pick up instrument artifacts as features. Guard with:

Cross-instrument validation.
Domain randomization at training time.
Regular audits using known physical priors.

11.3 Catastrophic forgetting and model drift

Continuous learning can cause models to forget old behaviors. Use replay buffers, ensemble techniques, and conservative update policies.

11.4 Adversarial vulnerabilities

Deliberately crafted inputs could fool AI — important for autonomous scheduling and alert distribution. Hardening involves adversarial training and human oversight for high-impact decisions.

12. Evaluation metrics and scientific validation

Selecting appropriate metrics ensures models do what scientists need.

12.1 Detection tasks

Precision / recall / F1 for candidate detection.
ROC curves and area under curve (AUC) for classification.
Completeness (sensitivity) and purity (1 − contamination) as a function of magnitude / surface brightness.
Injection-recovery curves to measure detection thresholds and selection functions.

12.2 Time series and periodic signals

False alarm rates (FAR) per unit time.
Detection efficiency as a function of transit depth / period.
Missed-opportunity metrics (fraction of high-value events that were not flagged in time for follow-up).

12.3 Parameter inference

Calibration of credible intervals (fractional coverage of nominal intervals).
Bias and RMSE for point estimates.
Posterior predictive checks for model adequacy.

12.4 End-to-end scientific value

Ultimately evaluate how many robust, reproducible discoveries the system produces (new validated exoplanets, measured galaxy populations, peer-reviewed publications), adjusted for resource cost.

13. Ethical, legal, and social considerations

AI in astronomy raises governance questions.

13.1 Data access and equity

Large facilities and well-funded groups dominate data and compute. Encourage open data policies, shared benchmarks, and cloud credits for under-resourced institutions to democratize discovery.

13.2 Responsible automated discovery

Automated alerts and press-worthy claims should follow a staged disclosure policy: internal vetting → multi-instrument confirmation → preprint → press release. Avoid sensationalism from unvetted AI outputs.

13.3 Intellectual property and credit

When AI or citizen scientists contribute to discovery, develop clear attribution norms and authorship practices.

13.4 Environmental cost

Large AI models have carbon footprints. Optimize models, use efficient architectures, and use cloud resources powered by renewables where possible.

14. Roadmap: short-, mid-, and long-term milestones

A community roadmap guides funding and operations.

Short-term (1–3 years)

Deploy robust AI denoisers, PSF models, and transient brokers for existing surveys.
Establish public benchmarks and injection-recovery toolkits.
Implement model registries and reproducibility workflows at major observatories.

Medium-term (3–7 years)

Operationalize probabilistic catalogs with uncertainty propagation to downstream science.
Integrate AI-driven schedulers at multi-facility networks and robotic telescopes.
Deploy federated learning across consortia to improve models without raw data sharing.

Long-term (7–15 years)

Fully autonomous chains from detection to follow-up for many transient classes.
AI-native space telescopes with on-board inference for exoplanet spectroscopy prioritization.
Community-agreed standards for AI verification, discovery thresholds, and attribution.

15. Case studies and illustrative projects

(Here are generalized case-study sketches illustrating typical successes and pitfalls.)

15.1 Large synoptic surveys

In surveys that revisit sky patches nightly, AI brokers filter millions of nightly detections, rank them for follow-up, and reduce human vetting burdens by orders of magnitude. Injection-recovery tests reveal population completeness curves used for cosmological inference.

15.2 Space photometry missions

Onboard denoising and transit detection allow rapid triage of candidate small planets, enabling near-real-time follow-up with ground-based RV resources — a major efficiency gain.

15.3 High-contrast imaging

AI speckle suppression yields contrast improvements near the inner working angle, enabling the detection of smaller, closer-in exoplanets than previously feasible.

16. Practical guidance & implementation patterns

For teams starting an AI-telescope project:

16.1 Start with clear scientific metrics

Design models to optimize for scientifically meaningful metrics (completeness at target depth, correct redshift fraction, low false-alarm rates), not just raw ML loss numbers.

16.2 Build realistic simulators

High-fidelity forward simulators that generate raw detector outputs (with noise, PSF, jitter) are indispensable for training robust models and measuring selection functions.

16.3 Use injection-recovery as a gold standard

Always validate detection pipelines with synthetic injections into real raw data processed end-to-end.

16.4 Invest in provenance and model governance early

Without versioned models and provenance, scientific reproducibility is fragile and expensive to reconstruct.

16.5 Combine physics and data

Where possible, embed physical constraints (radiative transfer, conservation laws, orbital dynamics) in architectures to improve generalization and interpretability.

17. The frontier: speculative directions and blue-sky ideas

Onboard AI telescopes in space that perform initial candidate vetting and automatically trigger higher-precision spectroscopy or maneuvering instruments to study transients in seconds.
Cross-modal retrieval systems that jointly learn from images, spectra, and time-series to detect subtle signatures of habitability (e.g., joint photometry-spectroscopy anomalies).
Generative discovery: Use generative models to propose novel physical models that explain anomalies, guiding new theory and instrument designs.
Citizen/AI hybrid discovery platforms where humans curate model outputs in real time and correct AI biases through structured feedback loops.
Global federations of telescopes with privacy-preserving joint decision-making for coordinated follow-up of rare events.

18. Conclusion — the promise and the prescription

AI is a revolutionary amplifier for astronomical telescopes. It opens routes to deeper discovery, faster reaction, and more complete science from data that were previously intractable or uninformative. But AI is not a plug-and-play magic: it requires rigorous engineering, careful evaluation, principled uncertainty quantification, and community norms for validation and disclosure.

Actionable prescription:

Fund integrated instrument + AI teams so models are co-designed with hardware.
Invest in realistic simulators and injection-recovery frameworks as foundational infrastructure.
Build public benchmarks and model registries to accelerate progress and ensure reproducibility.
Prioritize probabilistic outputs and uncertainty propagation into downstream science.
Democratize access to data and compute to widen participation and scientific creativity.
Require staged validation and multi-instrument confirmation for high-profile claims.

If the community follows these steps, AI-powered telescopes will reliably expand our view of the universe — revealing faint galaxies that trace cosmic history and identifying the most promising candidate worlds where life might exist. The next decades promise not only a flood of data, but the intelligent systems that make sense of it and turn raw photons into new understanding.

Appendix A — Example AI pipeline for transit detection (practical pseudo-algorithm)

Ingest raw photometry with timestamps and quality flags.
Preprocess: Mask outliers; apply learned detrending network conditioned on spacecraft/instrument telemetry.
Search: Slide a learned CNN matched filter to score candidate transit windows across periods. Use multi-scale filters for durations.
Vet: For top candidates, compute centroid motion, odd-even transit depth checks, secondary eclipse search. Input features into a classifier ensemble (CNN + gradient-boosted trees) to estimate planet probability.
Prioritize: Rank candidates by probability × detection SNR × scientific utility (planet radius, equilibrium temp).
Schedule: Request spectroscopic follow-up if probability exceeds threshold; if ambiguous, schedule higher-cadence photometry.
Update: After follow-up, add labels to training set; retrain periodically with conservative replay.

Appendix B — Recommended reading and resources (start here)

Introductory texts on machine learning for scientists (foundational ML).
Reviews on astroinformatics and AI in astronomy (survey literature).
Public datasets and benchmarks (survey archives, light-curve repositories).
Toolkits: ML libraries, radiative-transfer codes, and simulation toolkits for forward modeling.