Exoplanet
Classificatio
n using
Machine
Learning
Harnessing AI for the Next Generation
of Discovery
Presented by Team LUGIA | Project: Automated Exoplanet Disposition
NASA Space Apps Challenge 2025
The Astronomical Challenge: Classifying
Worlds Beyond
The volume of data generated by modern space observatories has outpaced the human capacity for analysis, particularly in the
critical task of exoplanet validation. We address the core challenge of precisely classifying planetary candidates found within
NASA's extensive archive.
The Core Classification Problem Why Automated Classification is Imperative
How do we efficiently and accurately categorize observed Accurate and rapid classification is fundamental to the
transits into one of three critical dispositions? mission of searching for life and habitable worlds.
• Confirmed Exoplanet Accelerated Discovery: Filters out noise so scientists can
• Candidate Planet (Requires further study) focus on genuine, Earth-like prospects.
• False Positive (Instrumental or stellar noise) Resource Optimization: Reduces the costly and time-
consuming manual validation process.
Filtering Noise: Eliminates thousands of spurious
detections caused by instrumental artifacts or background
stars.
Identifying Critical Flaws in Current Exoplanet Validation
Existing methodologies struggle to keep pace with the exponential increase in observational data. Our project is driven by the need to overcome several
systemic limitations inherent in traditional and partially-automated approaches.
Data Ingestion and Integrity Issues Heterogeneous Data Sources
Raw datasets from missions like Kepler, K2, and TESS are inherently Data is dispersed across multiple archives and observatories, each with
messy. They contain inconsistent formats, numerous missing values its own observational biases and feature sets. Combining data from
(NaNs), and require extensive pre-processing before any meaningful different telescopes (e.g., K2 vs. TESS) is complex, leading to
analysis can begin. This lack of standardization is a significant compatibility and scaling challenges.
bottleneck.
Manual and Time-Intensive Validation Lack of a Unified, Scalable Pipeline
The final validation of candidates often relies on painstaking human The scientific community needs a single, end-to-end processing pipeline
review of light curves and complex statistical modeling. This process is that can ingest raw data, unify diverse feature sets, train state-of-the-art
slow, subjective, and creates a significant backlog, limiting the rate of models, and provide reliable, actionable predictions across all major
discovery. exoplanet missions.
The shift from manual analysis to machine learning is not just an optimization; it is a necessity for handling petabyte-scale astrophysical datasets.
LUGIA's Solution: The Automated Exoplanet ML Pipeline
We propose and have implemented a robust Machine Learning pipeline designed to ingest heterogeneous exoplanet data and output highly
accurate disposition classifications, creating a new standard for automated planetary validation.
Data Cleaning & Unification Feature Engineering
Proprietary techniques for handling large-scale missing values and Automated selection and calculation of the top 20 most scientifically
merging disparate formats from Kepler, K2, and TESS archives into a relevant transit and stellar parameters (e.g., Orbital Period, Stellar
singular, clean dataset. Effective Temperature, Transit Depth).
Model Training & Optimization High-Confidence Prediction
Implementation of a high-performance Random Forest Classifier, Deployment of the trained model to rapidly and reliably predict the
optimized using cross-validation and hyperparameter tuning for disposition (Confirmed, Candidate, or False Positive) of new or
maximum generalization. archived exoplanet transit signals.
This automated approach dramatically reduces the analytical latency from months to mere seconds, providing a critical tool for follow-up
observation planning.
The Data Foundation: NASA Exoplanet Archive
Our model's performance is grounded in its training on the most comprehensive and authoritative source of exoplanet data: the NASA Exoplanet Archive. We carefully curated and combined
data from multiple missions to build a robust training set.
Cumulative Exoplanet Catalog
1
Core dataset providing confirmed planets and robust false positives from the primary Kepler mission,
offering high-fidelity labels for initial training.
K2 Candidate Data
2
Data from the K2 "second light" mission, introducing diversity in observational noise and stellar
targets, crucial for generalized model performance.
TOI (TESS Objects of Interest)
3
Integration of data from the ongoing TESS mission, enabling our model to classify signals from the
next generation of space-based transit surveys.
Merged Dataset Snapshot
Total Entries: ~21,000 unique planetary signals
Feature Count: 221 initial parameters
Detailed Methodology: From Raw Data to Accurate
Classification
Our pipeline utilizes a sophisticated sequence of steps to ensure data quality, extract maximum information, and achieve high predictive performance
1 2
Data Preprocessing Feature Engineering
Imputation of null values using K-Nearest Neighbors (KNN) to Employed feature importance ranking (based on Random Forest
maintain statistical distribution, and rigorous standardization of scores) to isolate the 20 most discriminative parameters,
numeric features. focusing on planetary radius, orbital period, and stellar
properties.
3 4
Model Selection Evaluation Metrics
The Random Forest Classifier was chosen for its robustness Performance was assessed using overall accuracy, precision,
against overfitting, ability to handle high-dimensional data, and recall, F1-scores for each class, and detailed visualization of the
intrinsic feature importance calculation. Confusion Matrix.
Quantifiable Results: High-Confidence Exoplanet Disposition
The Random Forest model achieved excellent overall predictive accuracy, demonstrating its capability to distinguish genuine planetary signals from instrumental noise, a critical objective for resource allocation.
90%
Overall Prediction Accuracy
This high level of accuracy ensures that the vast majority of classified candidates
are reliably categorized, significantly reducing the burden on human validators.
The model excels at identifying False Positives (99%), which is crucial for pruning the largest portion of the initial signal
pool. While performance on Confirmed and Candidate planets is strong (84% and 77% respectively), it highlights areas for
future optimization, particularly in distinguishing borderline candidates.
Confusion Matrix Visualization
The full confusion matrix reveals excellent separation between False Positives and the planetary classes, confirming the model's effectiveness as a high-pass filter for exoplanet validation.
Transformative Impact on Astrophysical Research
The implementation of this ML pipeline provides immediate, tangible benefits to researchers, mission planners, and the overall
pace of exoplanetary discovery.
Rapid Validation Optimized Resource Allocation
The model can classify thousands of candidates per second, By reliably filtering out 99% of false positives, the system
allowing researchers to quickly prioritize the most promising ensures that scarce observational time on major telescopes is
signals for follow-up observation using ground-based telescopes directed only towards high-probability targets, maximizing
or JWST. scientific return.
Future-Proof Scalability Enhanced Data Integrity
The pipeline’s modular design means it can be readily adapted The model provides a standardized, objective classification
to ingest data streams from future high-throughput missions, metric, reducing the subjectivity and inter-observer variability
such as the upcoming PLATO mission, ensuring continuous, inherent in purely manual classification methods.
high-speed analysis.
Trajectory Beyond Today: Future Scope and
Development
While our Random Forest model delivers high performance, the pathway to even greater accuracy and utility lies in exploring more
complex architectures and creating user-friendly interfaces for the scientific community.
Advanced Model Architecture 1
Transition from traditional ML to Deep Learning,
specifically implementing Convolutional Neural
Networks (CNNs) to analyze raw light curve data 2 Interactive Visualization Dashboard
directly, potentially yielding higher accuracy and
discovering subtle, complex features that evade Develop a web-based interface allowing scientists to
engineered features. upload new candidate data, receive real-time
classification, and interactively explore the model's
Live API Integration 3
feature importance and output confidence levels.
Integrate the classification pipeline with the live NASA
Exoplanet Archive API and TESS/JWST data streams to
enable near real-time, instantaneous disposition of
newly detected transit signals.
Conclusion: Machine Learning as the Engine of
Discovery
Team LUGIA successfully developed and validated a Machine Learning pipeline that effectively addresses the data deluge challenge in
exoplanetary science. We have demonstrated that AI can serve as a powerful force multiplier for space exploration.
Noise Reduction Automated Classification High Performance
Our solution effectively reduces The pipeline provides an objective, Achieving a 90% overall accuracy, the
observational noise, reliably isolating scalable method for dispositioning model provides high-confidence
False Positives with 99% accuracy. candidates from Kepler, K2, and predictions that are immediately
TESS datasets. actionable for research teams.
By automating the tedious, high-volume task of initial classification, our ML pipeline directly supports NASA’s core mission:
accelerating the discovery and validation of new worlds, bringing us closer to answering the question: Are we alone?