SM-Net: Learning a Continuous Spectral Manifold from Multiple Stellar Libraries

1. Problem Statement

Stellar spectral libraries (e.g., PHOENIX, C3K, OB-PoWR, TMAP) are fundamental to astrophysics for stellar atmosphere studies and population synthesis. However, existing libraries suffer from several critical limitations:

• Fragmented Coverage: No single library covers the entire stellar parameter space ($T_{\text{eff}}$, $\log g$, $\log Z$). Libraries often have disjoint or only partially overlapping regions.
• Inconsistencies: Different libraries use varying physical assumptions, numerical methods, and line lists, leading to discontinuities and non-physical spectral features at library boundaries.
• Data Artifacts: Many spectra contain masked or zero-valued flux regions due to incomplete line lists or numerical instabilities.
• Interpolation Challenges: Traditional interpolation methods struggle with irregular sampling, sparse regions, and the need to merge heterogeneous grids without propagating artifacts.
• Inverse vs. Forward: Existing machine learning (ML) emulators are typically trained on single libraries, inheriting their specific limitations, and focus on inverse problems (label inference) rather than forward spectral synthesis across multiple grids.

2. Methodology

The authors propose SM-Net, a deep learning framework designed to learn a continuous, unified spectral manifold from multiple heterogeneous stellar libraries.

A. Dataset Construction

• Source Libraries: The model combines four high-resolution libraries: PHOENIX–Husser, C3K–Conroy, OB–PoWR, and TMAP–Werner.
• Unified Grid: The combined dataset spans $T_{\text{eff}} = 2,000–190,000$ K, $\log g = -1–9$, and $\log Z = -4–1$, covering wavelengths from 3,000 to 100,000 Å.
• Preprocessing:
  • Spectra were resampled onto a common, non-uniform wavelength grid (11,655 points) optimized for feature-rich regions.
  • Handling Missing Data: Zero or masked flux values are treated as "unknowns" rather than physical zeros, allowing the network to infer missing regions.
  • Metallicity Replication: Since OB-PoWR and TMAP do not parametrize metallicity, their spectra were replicated across discrete $\log Z$ values to embed them in the 3D parameter space.
  • Overlap Resolution: In regions of exact parameter overlap, a deterministic priority order (TMAP > PHOENIX > OB-PoWR > C3K) was applied to select a single reference spectrum.
• Data Splitting: To ensure strict generalization testing, the dataset was split based on parameter space rather than random sampling. Training and validation sets share no common parameter values (except for boundary conditions), forcing the model to interpolate across unseen combinations.
• Augmentation: To address class imbalance (dense regions vs. sparse regions), spectra were augmented based on pairwise cosine similarity. Distinctive spectra were up-sampled (up to 10x) to ensure rare spectral morphologies influence training.

B. Model Architecture (SM-Net)

SM-Net employs a two-stage architecture:

1. Lookup Block (Global Predictor):
   • Takes input parameters ($T_{\text{eff}}, \log g, \log Z$) and maps them to a coarse, low-frequency spectral estimate.
   • Uses a Parameter Embedding Layer followed by residual Fully Connected (FC) layers.
   • Includes a learnable baseline vector and a Softplus activation to ensure non-negative flux.
2. Refinement Block (Local Corrector):
   • Takes the coarse spectrum, physical parameters, and positional encodings.
   • Uses a Multi-scale Feature Extractor with three parallel convolutional branches (kernel sizes 7, 15, 23) to capture both local spectral lines and broad continuum features.
   • Outputs an additive correction ($\Delta s$) to the coarse spectrum via a residual connection: $s_{\text{final}} = s_0 + \Delta s$.

C. Training Strategy

• Loss Function: A weighted Mean Squared Error (MSE) is used. Crucially, a binary mask ($m_{i,\lambda}$) excludes masked/zero flux regions from the loss calculation, allowing the model to learn to reconstruct missing data based on correlations with neighboring wavelengths and parameters.
• Optimization: Adam optimizer with a dynamic learning rate and batch size schedule. Large batches are used early for global trends, while smaller batches later introduce gradient noise to escape local minima and refine fine structures.
• Scaling: Flux values are transformed using a log1p scaling to handle the large dynamic range of stellar spectra while preserving monotonicity.

3. Key Contributions

• Unified Continuous Manifold: SM-Net is the first model to successfully learn a continuous spectral manifold spanning four distinct, high-resolution stellar libraries with heterogeneous parameter spaces.
• Masked Flux Reconstruction: The model treats masked flux as unknowns, enabling it to infer missing spectral structure (e.g., in TMAP libraries) using learned correlations, a capability beyond standard interpolation.
• High-Throughput Inference: The model achieves an inference speed of >14,000 spectra per second on a single NVIDIA RTX 4090 GPU, enabling rapid exploration of parameter space.
• Open-Source Ecosystem: The authors released the model weights, code, and an interactive web dashboard (Dash/Plotly) allowing users to generate and visualize spectra in real-time without command-line expertise.

4. Results and Performance

• Accuracy:
  • Training Set MSE: $1.47 \times 10^{-5}$ (RMS flux deviation $\approx 0.38\%$).
  • Test Set MSE: $2.34 \times 10^{-5}$ (RMS flux deviation $\approx 0.48\%$).
  • The model achieves sub-1% RMS error for the vast majority of spectra, with the best 10% reaching $\sim 0.1\%$ error.
• Generalization: Performance on the validation set (containing unseen parameter combinations) is comparable to the test set, confirming robust interpolation capabilities.
• Error Analysis:
  • Low Error: Regions dominated by a single library (e.g., hot TMAP stars) show minimal error.
  • High Error: The largest errors occur at low temperatures ($T_{\text{eff}} < 3,000$ K) and at boundaries where library physics mismatches are significant (e.g., PHOENIX vs. C3K).
  • Metallicity Limitation: In the high-temperature regime ($T_{\text{eff}} > 15,000$ K) dominated by OB-PoWR and TMAP, the model produces spectra invariant to $\log Z$ because the source grids lack metallicity parametrization.
• Qualitative Examples: The model successfully reconstructs Sun-like stars, metal-poor systems, and hot stars, accurately capturing fine absorption features. It also demonstrates the ability to "fill in" missing high-wavelength flux in TMAP spectra where the original library had gaps.

5. Significance and Future Outlook

• Paradigm Shift: SM-Net moves stellar spectral modeling from discrete, fragmented grids to a smooth, differentiable, and continuous manifold.
• Applications: The high throughput and continuous nature of the model make it ideal for:
  • Dense parameter sweeps in stellar evolution.
  • Bayesian spectral fitting (e.g., PPXF).
  • Rapid construction of Simple Stellar Population (SSP) models.
  • Interactive exploration of stellar physics.
• Limitations & Future Work:
  • The model currently predicts static spectra (no velocity broadening or instrumental convolution).
  • Extrapolation beyond the training domain is possible but unvalidated physically.
  • Future work envisions hybrid approaches combining SM-Net's speed with physical radiative-transfer solvers to enforce physical constraints (e.g., ionization equilibrium) while leveraging the ML model for efficiency.

In conclusion, SM-Net provides a robust, fast, and accessible tool that bridges the gap between detailed stellar atmosphere calculations and large-scale population synthesis, offering a pathway for next-generation astrophysical modeling.