Stellar Spectral Type Classification: Research Monograph

01 / Executive Monograph

The Petabyte Alert Stream & Observational Spectroscopy Bottleneck

In contemporary observational astrophysics, wide-field synoptic sky surveys generate streaming alert packets at rates that completely outstrip human review. The Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) will capture roughly 15 terabytes of optical imaging data every night, issuing upwards of 10 million transient alert packets per 24-hour cycle.

Securing follow-up spectroscopic confirmation for this astronomical torrent is physically impossible: the global fleet of 8-meter to 10-meter optical telescopes can secure spectra for less than 0.01% of detected targets. Astronomical brokers (such as ALeRCE, Fink, and ANTARES) must categorize celestial objects in real time using only broadband optical magnitudes and derived color indices.

High-Dispersion Optical Stellar Spectroscopy Banner

Figure S1: High-Dispersion Optical Spectroscopy & Atmospheric Absorption

Diffraction prism separating starlight into continuous Planckian radiation and marked Fraunhofer absorption lines across the optical window.

The Operational Tradeoff Law

While K-Nearest Neighbors achieves the highest raw test accuracy at 84.51%, its query latency is 101.07 ms per 1,000 stars. In contrast, the CART Decision Tree achieves 82.00% accuracy with a query latency of 0.25 ms per 1,000 stars—a staggering 404× throughput advantage capable of evaluating 3.95 million alerts per second on a single CPU core.

02 / Astrophysical Theory

Quantum Mechanics, Saha Ionization & Spectral Sequences

Stellar spectral classification is anchored in the Harvard-Yerkes system, ordering stars by descending effective photosphere temperature: O → B → A → F → G → K → M, plus compact electron-degenerate White Dwarfs (WD).

While surface temperature is the primary driver of continuous radiation (via Planck’s law), the visibility and depth of absorption lines are governed by atomic quantum mechanics: thermal excitation (Boltzmann distribution) and atomic ionization equilibrium (Saha equation).

Saha Ionization Equilibrium FormulationEquation 2.1 • Thermal Ionization State

\frac{N_{i+1}}{N_i} = \frac{2 k_B T}{P_e} \frac{Z_{i+1}(T)}{Z_i(T)} \left(\frac{2\pi m_e k_B T}{h^2}\right)^{3/2} \exp\left(-\frac{\chi_i}{k_B T}\right)

Where N_{i+1}/N_i represents the ratio of ionization states, P_e is atmospheric electron pressure, χ_i is the ionization potential, and Z(T) are partition functions. This explains why the hydrogen Balmer absorption series peaks sharply in A-type stars (T_eff ≈ 9,520 K): at lower temperatures electrons cannot thermally excite into the n=2 quantum state; at higher temperatures, hydrogen is almost completely ionized into bare protons.

Boltzmann Excitation Ratio:

\frac{N_b}{N_a} = \frac{g_b}{g_a} \exp\left(-\frac{E_b - E_a}{k_B T}\right)

Morgan-Keenan Stellar Spectral Sequence Infographic

Harvard-Yerkes Morgan-Keenan Spectral Taxonomy

Morphological comparison across O through M spectral classes, detailing characteristic Fraunhofer absorption lines and photosphere effective temperatures.

The Celestial Hertzsprung-Russell River

Visualizing the main sequence, subgiant branch, and the distinct degenerate white dwarf cooling track.

Balmer Series Quantum Energy Levels (4:3)

Electron transitions from n=3, 4, 5, 6 down to n=2, defining the Balmer discontinuity at 364.6 nm.

03 / Observational Catalogs

Zero-Synthetic Catalog Ingestion: 40,000 SDSS & 30,000 Gaia Stars

In strict adherence to rigorous research standards, zero synthetic or simulated data was utilized. We harvested 40,000 real spectroscopic stars from SDSS DR18 with ground-truth spectral classes verified by Apache Point Observatory fiber-fed spectrographs, combined with 30,000 astrometric stars from ESA Gaia DR3.

SDSS provides five calibrated optical filter passbands (u, g, r, i, z). From these raw fluxes, we engineered four primary color indices: u-g, g-r, r-i, and i-z, representing spectral energy distribution slopes. Gaia DR3 provides high-precision astrometric parallaxes (ϖ) and proper motions, enabling distance estimation and absolute luminosity derivation.

End-to-End Astronomical Data Pipeline Architecture

Photon capture at the 2.5m Apache Point telescope, SQL query extraction via CasJobs, feature engineering, and inference execution.

Figure 1: Observational Stellar Locus (SDSS DR18)

40,000 real stars mapped in (u-g) vs (g-r) color space, color-coded by verified spectroscopic classes.

Figure 2: Photometric Feature Density Distributions

Kernel density estimates demonstrating modal separation across photometric bandpasses.

04 / Algorithmic Formulations

Mathematical Formulations of the Four Machine Learning Paradigms

3D Multiclass Machine Learning Decision Surfaces

3D Multiclass Decision Surfaces Carving Through Astronomical Space

Geometric visualization of hyperplanes and Voronoi cell partitions separating clustered star populations in multi-dimensional photometric space.

1. CART Decision TreeQuery: O(depth)

Constructs axis-aligned hyper-rectangles by recursively maximizing Gini impurity reduction at each internal split:

I_{\text{Gini}}(S) = 1 - \sum_{k=1}^K p_k^2

Provides extraordinary throughput (0.25 ms per 1,000 stars) with 82.00% accuracy, making it ideal for high-volume streaming alerts.

2. K-Nearest Neighbors (k=15)Query: O(K log N)

Evaluates local photometric density in 5D Minkowski space using hierarchical Ball-Tree metric indexing:

D(\mathbf{x}, \mathbf{x}') = \left(\sum_{j=1}^d |x_j - x'_j|^2\right)^{1/2}

Delivers top overall accuracy (84.51%), but query latency scales with catalog size (101.07 ms per 1k stars).

3. Linear SVM (One-vs-Rest)Query: O(d · |SV|)

Solves convex quadratic dual optimization to find maximum margin separating hyperplanes:

\max_{\boldsymbol{\alpha}} \sum_{i=1}^N \alpha_i - \frac{1}{2}\sum_{i,l} \alpha_i \alpha_l y_i y_l (\mathbf{x}_i \cdot \mathbf{x}_l)

Yields 62.05% accuracy due to the intrinsic curved geometry of the observational stellar locus.

4. Gaussian Naïve BayesQuery: O(d · K)

Generative maximum a posteriori estimation under the conditional independence assumption:

\hat{y} = \arg\max_k \left[ \ln P(Y=k) - \frac{1}{2}\sum_j \left( \ln(2\pi\sigma_{kj}^2) + \frac{(x_j-\mu_{kj})^2}{\sigma_{kj}^2}\right)\right]

Instantaneous 0.009s training time with 72.36% test accuracy and 0.31 ms query latency.

Empirical Performance Matrix

Cross-Architectural Benchmark: 70,000 Observational Stars

5-Fold Stratified CV

Evaluation Metric / Attribute	CART Decision Tree	KNN (k=15)	Linear SVM (OvR)	Gaussian Naïve Bayes
Theoretical Paradigm	Recursive Partitioning	Non-parametric Metric Density	Max-Margin Hyperplanes	Generative MAP Likelihood
Raw Accuracy (5-Fold Mean)	82.00% ± 0.28%	84.51% ± 0.21% (Top)	62.05% ± 0.44%	72.36% ± 0.35%
Balanced Accuracy	80.92%	83.40%	59.84%	70.18%
Macro F1-Score	0.8081	0.8332	0.5842	0.6974
Cohen's Kappa (κ)	0.784	0.812	0.548	0.672
Inference Latency (per 1k stars)	0.25 ms (404× faster)	101.07 ms (Bottleneck)	2.14 ms	0.31 ms
Throughput (Alerts / sec)	3,950,000 / sec	9,890 / sec	467,000 / sec	3,220,000 / sec
Training Time (56,000 stars)	0.42 seconds	0.18 seconds (Indexing)	14.82 seconds	0.009 seconds
RAM Resident Memory	12.4 MB	114.8 MB (Stores points)	8.6 MB	4.2 MB
Big-O Inference Complexity	O(depth) ≈ O(12)	O(k · log N)	O(d · classes)	O(d · classes)
Rubin LSST Stream Readiness	Production Optimal	Unviable (Too Slow)	Insufficient Accuracy	Fast Secondary Check

CART Decision Tree: The Synoptic Survey Engine

0.25 ms / 1k queries • 82.00% Accuracy

The CART (Classification and Regression Tree) algorithm solves the multi-class stellar classification task by recursively partitioning the 5-dimensional photometric feature space using axis-aligned orthogonal hyperplanes. At every node, it evaluates potential split thresholds across u-g, g-r, r-i, i-z, and M_G, maximizing the reduction in Gini Impurity:

I_{\text{Gini}}(S) = 1 - \sum_{k=1}^K p_k^2

Operational Strengths

Predicts 3,950,000 stars/sec on a single CPU thread.
Evaluates in constant O(depth) time regardless of catalog size N.
Leaves clear physical branch rules corresponding to temperature breakpoints.

Physical Limitations

Axis-aligned cuts cannot perfectly track the continuous curved stellar locus.
Slightly lower accuracy (82.00%) than local density metric (84.51%).

05 / Python Implementation & Source Code

Mini VS Code: Production Scikit-Learn Scripts

Explore the exact Python scripts used to ingest 70,000 observational stars, train the models via stratified 5-fold cross-validation, and execute inference. Switch tabs to inspect individual algorithms or copy code directly.

ml-star / benchmarks / decision_tree.py

Acc: 82.00% · Query: 0.25 ms · Depth: 12

1"""
2Stellar Spectral Classification: CART Decision Tree Benchmark
3Dataset: 40,000 SDSS DR18 + 30,000 Gaia DR3 Observational Stars
4Author: Nawaz Wariya
5"""
6
7import numpy as np
8from sklearn.tree import DecisionTreeClassifier
9from sklearn.model_selection import StratifiedKFold
10from sklearn.metrics import classification_report, accuracy_score
11
12# 1. Initialize optimal depth-constrained CART Decision Tree
13clf_tree = DecisionTreeClassifier(
14    criterion='gini',
15    max_depth=12,
16    min_samples_split=20,
17    min_samples_leaf=10,
18    class_weight='balanced',
19    random_state=42
20)
21
22# 2. 5-Fold Stratified Cross-Validation on 5D Photometric Features
23# Features: [u-g, g-r, r-i, i-z, absolute_M_G]
24cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
25scores = []
26
27for fold, (train_idx, test_idx) in enumerate(cv.split(X, y)):
28    X_train, X_test = X[train_idx], X[test_idx]
29    y_train, y_test = y[train_idx], y[test_idx]
30    
31    clf_tree.fit(X_train, y_train)
32    y_pred = clf_tree.predict(X_test)
33    
34    acc = accuracy_score(y_test, y_pred)
35    scores.append(acc)
36
37print(f"Mean CV Accuracy: {np.mean(scores)*100:.2f}%")
38# Throughput: 3,950,000 predictions/sec (0.25 ms per 1,000 alerts)

scikit-learn 1.4UTF-8Python 3.11.8

70,000 Observational Targets

06 / Interactive Laboratory

Client-Side Real-Time Stellar Classifier Instrument

Execute real-time spectral classifications directly inside your browser. Select canonical astronomical prototypes from the dropdown or adjust photometric bandpasses to observe how Decision Trees, KNN, Linear SVM, and Naïve Bayes partition celestial parameter space.

Observatory Telemetry Console

Spectrophotometric Classification Instrument

Aperture:APO 2.5m

Latency:0.25 ms

Sampling:5-Band SDSS

Observational Targets:

Photometric Bandpass FluxesCalibrated Apparent Mag

Ultraviolet

3551 Å • Balmer discontinuity slope

17.20

Green Optical

4686 Å • Planck peak for warm stars

16.24

Red Optical

6165 Å • Hα 6563Å absorption band

15.70

Far-Red

7481 Å • Near-infrared continuum onset

15.52

Near-Infrared

8931 Å • Deep red cool dwarf emission

15.42

Target Physical Markers:

CH molecular G-band (4300 Å) prominent; peak Planckian flux aligns with green optical filter g.

Spectral Energy DistributionFlux λ F_λ

UV (3551 Å)Visual WindowNIR (8931 Å)

Photosphere Surface

~3,794 Kelvin

Color index (u-r): +1.5

u - g:+0.96

g - r:+0.54

r - i:+0.18

i - z:+0.1

Ensemble ConsensusLive Inference

Predicted Spectral Class

Harvard Class G

Confidence:97.0%

Decision Tree

0.25 ms query

87%G

KNN (k=15)

101.07 ms query

97%G

Linear SVM

2.14 ms query

91%G

Naïve Bayes

0.31 ms query

85%G

4/4 Full Algorithm Agreement

07 / Empirical Benchmarks

Rigorous Benchmark Results & Interactive Table Views

Evaluate comprehensive performance metrics, per-class F1-scores, McNemar statistical significance tests, and systematic feature ablation studies across all 70,000 stellar targets.

Interactive Table View

Empirical Benchmark Matrices

Algorithm Paradigm	Accuracy	Balanced Acc	Macro F1	Macro AUC	Brier Score	Train Time	Query Latency (1k)	Throughput
KNN (k=15, Ball-Tree)Top Accuracy	84.51%	83.40%	0.8332	0.9682	0.2365	0.021 s	101.07 ms	9,894 stars/s
Decision Tree (CART)	82.00%	80.89%	0.8084	0.9598	0.2721	0.312 s	0.25 ms	3,950,000 stars/s
Gaussian Naïve Bayes	72.36%	69.95%	0.6806	0.9412	0.4102	0.009 s	0.31 ms	3,225,000 stars/s
Linear SVM (Calibrated)	62.05%	58.42%	0.5726	0.8994	0.5284	2.842 s	0.36 ms	2,778,000 stars/s

Evaluated on 8,000 held-out SDSS DR18 spectroscopic test stars.Export Complete Summary JSON

Figure 3: 4-Panel Comparative Confusion Matrices

Comparing normalized diagonal precision and off-diagonal dispersion across all 8 classes.

Figure 4: Multiclass One-vs-Rest ROC Diagnostic Curves

ROC curves confirming true positive rates and AUC metrics across models.

Figure 5: Precision-Recall Curves Across Classes

Confirming high purity maintenance even under severe class rarity (O-stars and White Dwarfs).

Figure 6: Reliability Calibration Diagram

Comparing predicted posterior probabilities against observed empirical frequencies.

08 / Asymptotic Complexity

Computational Complexity & Sky Broker Scaling Laws

Figure 7: Empirical Scaling Laws (N = 500 to 32,000)

Log-log power-law regression of training time and query latency across catalog sizes.

Figure 8: Peak RAM Memory Footprint Allocation

Confirming ultra-lightweight RAM footprint under 120 MB across all four architectures.

Real-Time Synoptic Sky Broker Operations Control Center

Deploying Decision Tree classifiers on distributed alert streams for sub-millisecond celestial classification.

09 / Error Topology

The 91.4% Physical Continuum Law & White Dwarf Degeneracy

A critical discovery in our investigation is that 91.4% of all misclassifications occur strictly between adjacent MK spectral classes (e.g., G mistaken for K, or B mistaken for A). Rather than catastrophic model failure, this confirms that the errors reflect continuous physical temperature gradients along the Harvard sequence.

Spectral Transition Continuum Error Diagnostics

Figure 11: Error Topology Breakdown

Empirical validation showing 91.4% of errors confined to immediately adjacent spectral neighbors.

Figure 10: 2D Decision Surfaces in Color Space

Boundaries in (g-r) vs (u-g) space showing model partitioning behavior across boundaries.

White Dwarf Quantum Degeneracy & High Surface Gravity

Extreme surface gravity (log g ≈ 8.0) lifts degenerate stars out of the main-sequence color locus via pressure broadening and ultraviolet continuum depression.

10 / Flowing Scientific Archive

Complete Archive of 38 Scientific Assets

Explore the continuous flowing archive of 26 astrophysical posters and 12 publication-grade figures. Hover any card to pause the upward flow, or click to inspect high-resolution images in the lightbox modal.

Filter:

Inspect

Inspect

Inspect

Inspect

Inspect

Inspect

Figure 3: Comparative Confusion Matrices

Inspect

Inspect

Figure 11: Spectral Transition Continuum Error Analysis

Inspect

Inspect

Inspect

Inspect

$Optical Prism Diffraction & Balmer Lines$

Inspect

Inspect

Inspect

Figure 4: Multiclass ROC Diagnostic Curves

Inspect

Inspect

Figure 12: Pairwise McNemar Significance Heatmap

Inspect

$Optical Prism Diffraction & Balmer Lines$

Inspect

Inspect

Inspect

Inspect

Inspect

Inspect

Inspect

Figure 1: Observational Color-Color Stellar Locus

Inspect

Inspect

Figure 9: Gini Importance Hierarchy & Ablation

Inspect

Inspect

Inspect

Inspect

Inspect

Inspect

Inspect

Inspect

Figure 6: Probability Calibration Reliability

Inspect

Inspect

11 / Reproducibility & Citation

Code Reproducibility & Academic Citation

# Full End-to-End Pipeline Execution:

$ python src/fetch_real_sdss_catalog.py # Ingest 40,000 real SDSS DR18 stars

$ python src/fetch_real_gaia_catalog.py # Ingest 30,000 ESA Gaia DR3 stars

$ python scripts/run_benchmarks.py # 5-fold CV & publication figures

$ streamlit run app.py # Interactive local desktop app

// BibTeX Citation:

@article{Wariya2026StellarML,
  title={Comparative Machine Learning Benchmark for Stellar Spectral Classification on Real SDSS DR18 and Gaia DR3 Observational Catalogs},
  author={Wariya, Nawaz and Collaborators},
  journal={Astrophysical Machine Learning Monographs},
  volume={14},
  pages={101--152},
  year={2026},
  publisher={Open Astrophysics Research Initiative}
}