TinyML Models

LARUN uses a federation of specialized TinyML models, each optimized for a specific astronomical detection or analysis task. All models are under 100KB, enabling edge deployment and browser-based inference.

Specialized Models

<100KB

Per Model

81.8%

Detection Accuracy

<50ms

Inference Time

Model Architecture

All LARUN models share a common architecture optimized for TinyML deployment:

┌─────────────────────────────────────────────────────────────┐
│                    INPUT LIGHT CURVE                         │
│              (normalized flux vs. time)                      │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                  PREPROCESSING LAYER                         │
│     Normalization │ Detrending │ Outlier Removal            │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│              FEATURE EXTRACTION (1D-CNN)                     │
│     Conv1D(32) → ReLU → MaxPool → Conv1D(64) → ReLU         │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                 CLASSIFICATION HEAD                          │
│        Flatten → Dense(128) → Dropout → Dense(N)            │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                      OUTPUT                                  │
│           Class probabilities or regression values           │
└─────────────────────────────────────────────────────────────┘

Optimization Techniques

INT8 Quantization: Reduces model size 4x with minimal accuracy loss
Pruning: Removes redundant weights
Knowledge Distillation: Trained from larger teacher models
TensorFlow Lite: Optimized for edge inference

Model Federation

LARUN models work together as a federation, sharing information and combining predictions for robust analysis.

                    ┌─────────────────┐
                    │   ORCHESTRATOR  │
                    │   (Coordinator) │
                    └────────┬────────┘
                             │
        ┌────────────────────┼────────────────────┐
        │                    │                    │
        ▼                    ▼                    ▼
┌───────────────┐   ┌───────────────┐   ┌───────────────┐
│ EXOPLANET-001 │   │   VSTAR-001   │   │  FLARE-001    │
│ Transit Det.  │   │ Variable Star │   │ Flare Det.    │
└───────────────┘   └───────────────┘   └───────────────┘
        │                    │                    │
        └────────────────────┼────────────────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │ ENSEMBLE VOTE   │
                    │ Final Decision  │
                    └─────────────────┘

How Federation Works

Task Routing: Orchestrator routes input to relevant models
Parallel Inference: Multiple models process simultaneously
Ensemble Voting: Results combined with confidence weighting
Conflict Resolution: Disagreements resolved by specialized arbiter

EXOPLANET-001

Transit Detection

Core Model

The flagship model for detecting exoplanet transits. Identifies periodic dips in stellar brightness caused by planets passing in front of their host stars.

Size 48 KB

Accuracy 81.8%

Inference 23 ms

Training Data 5,000+ transits

Input

Light curve: 1000-5000 flux measurements
Cadence: 2-minute to 30-minute
Normalized to median flux = 1.0

Output

is_transit: Binary classification (true/false)
confidence: Probability score (0-1)
period_estimate: Rough period if detected

Use Cases

Initial transit candidate detection
Vetting potential false positives
Bulk screening of target lists

VSTAR-001

Variable Star Classification

High Accuracy

Classifies variable stars into categories including eclipsing binaries, pulsators, and rotational variables. Critical for distinguishing real transits from stellar variability.

Size 68 KB

Accuracy 87.2%

Inference 31 ms

Classes 12

Classification Categories

Class	Description	Period Range
EA	Detached eclipsing binary	0.5 - 100 days
EB	Beta Lyrae type binary	0.5 - 10 days
EW	W UMa contact binary	0.2 - 1 day
RRAB	RR Lyrae fundamental mode	0.3 - 1 day
RRC	RR Lyrae first overtone	0.2 - 0.5 day
DSCT	Delta Scuti pulsator	0.02 - 0.3 day
GDOR	Gamma Doradus pulsator	0.3 - 3 days
CEPH	Classical Cepheid	1 - 100 days
ROT	Rotational variable	0.1 - 50 days
LPV	Long period variable	50 - 1000 days

FLARE-001

Stellar Flare Detection

Real-time

Detects stellar flares - sudden brightening events caused by magnetic reconnection on the stellar surface. Important for assessing habitability around active stars.

Size 42 KB

Accuracy 89.5%

Inference 18 ms

Training Data 10,000+ flares

Output Parameters

flare_detected: Boolean flag
flare_start: Start time (BJD)
flare_peak: Peak brightness time
flare_energy: Estimated energy (ergs)
flare_amplitude: Peak amplitude (relative flux)

ASTERO-001

Asteroseismology

Stellar Physics

Analyzes stellar oscillations to determine fundamental stellar parameters like mass, radius, and age. Essential for accurate planet characterization.

Size 72 KB

Accuracy 92.1%

Parameters nu_max, delta_nu

Derived Parameters

nu_max: Frequency of maximum power (microHz)
delta_nu: Large frequency separation (microHz)
stellar_mass: Derived mass (solar masses)
stellar_radius: Derived radius (solar radii)
stellar_age: Estimated age (Gyr)

SPECTYPE-001

Spectral Classification

Fast

Classifies stellar spectral types (O, B, A, F, G, K, M) from photometric colors or low-resolution spectra. Used for habitable zone calculations.

Size 38 KB

Accuracy 94.3%

Classes OBAFGKM

GALAXY-001

Galaxy Morphology

Image-based

Classifies galaxy morphologies (spiral, elliptical, irregular, merger). Uses 2D image input rather than 1D light curves.

Size 95 KB

Accuracy 88.7%

Input 64x64 image

SUPERNOVA-001

Supernova Classification

Alert System

Classifies supernova types from early light curve data. Enables rapid classification for follow-up observations.

Size 58 KB

Accuracy 85.4%

Classes Ia, II, Ibc

MICROLENS-001

Microlensing Detection

Rare Events

Detects gravitational microlensing events - temporary brightening caused by a massive object passing between us and a background star. Can reveal planets even around distant stars.

Size 52 KB

Accuracy 78.9%

Events/year ~3000

TRAINER-001 (ASTRA)

ASTRA - Educational AI Tutor

RAG-Powered

ASTRA (Astronomical Science Tutor and Research Assistant) is LARUN's AI-powered space science educator. Unlike other models, ASTRA uses Retrieval-Augmented Generation (RAG) to provide accurate, sourced answers about space science.

ASTRA's Personality

ASTRA is enthusiastic, patient, and curious. It loves sharing "did you know?" facts and uses analogies to explain complex concepts. ASTRA always cites sources and encourages further exploration.

Knowledge Sources

NASA: TESS, Kepler, JWST mission data
ESA: Gaia, CHEOPS, ExoMars
ISRO: AstroSat, Chandrayaan
JAXA: Japanese space exploration
SpaceX: Launch and mission data
CNSA: Chinese space program
arXiv: Research papers

Features

Interactive Lessons: Structured learning paths
Quizzes: Test your knowledge
Model Explainer: Understand why models make predictions
Memory: Remembers your learning progress
Hot-swap Models: Always uses latest model versions

Example Interactions

# Learning
"/learn topic exoplanets"
"/learn quiz transits"

# Questions
"What is the transit method?"
"How does TESS find planets?"
"Explain the habitable zone"

# Explanations
"/learn explain"  # Why did the model predict this?

Explore the Models in Action

See how our TinyML federation works together to find exoplanets.

Try LARUN Now