Characterizing the Instrumental Profile of LAMOST

This paper presents a multi-layer perceptron model based on The Payne to characterize the instrumental profile of the LAMOST telescope, demonstrating that applying this derived profile to stellar radial velocity measurements reduces dispersion by approximately 3 km/s and thereby enhancing the search for long-period binary stars.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Problem: The "Blurry" Telescope Glasses

Imagine you are looking through a pair of glasses to read a very fine print. Even if the text is perfectly sharp, your glasses might be slightly smudged, warped, or dirty. This causes the letters to look a little blurry, stretched, or tilted.

In astronomy, the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) is like a giant, super-powered camera that takes pictures of the light from thousands of stars at once. But just like your glasses, the telescope's "lenses" and detectors aren't perfect. They distort the starlight slightly.

This distortion is called the Instrumental Profile (IP). It's the "fingerprint" of how the telescope messes up a perfect, sharp line of light. If you don't know exactly what that fingerprint looks like, you can't accurately measure things like how fast a star is moving toward or away from us (its Radial Velocity).

The Old Way: Trying to Guess the Shape

For a long time, scientists tried to fix this by assuming the distortion was a simple, symmetrical shape—like a perfect bell curve (a Gaussian). They thought, "Okay, the light is just a little bit spread out evenly."

But the reality is messier. The telescope's distortion changes based on:

• Where the light enters: Different fibers (cables carrying light) act differently.
• What color the light is: Blue light distorts differently than red light.
• When the photo was taken: Temperature changes in the observatory can warp the metal mirrors slightly, changing the distortion over time.

Trying to describe this complex, shifting shape with a simple "bell curve" is like trying to describe a cloud using only a square. It just doesn't fit well enough for high-precision science.

The New Solution: The "AI Chef"

The authors of this paper decided to stop guessing and start learning. They used a Neural Network (a type of Artificial Intelligence) to act like a master chef who has tasted every possible variation of a dish.

1. The Ingredients: They fed the AI thousands of images of "arc lamps." These are special calibration lights the telescope uses, which have very sharp, known lines. By looking at how the telescope distorted these known lines, the AI learned the "fingerprint" of the instrument.
2. The Training: The AI learned that the fingerprint changes depending on the fiber number, the wavelength (color), and the time of day/year.
3. The Result: After training, the AI became a "Time Machine" for light. You can ask it: "What does the distortion look like for Fiber #154, at Red Wavelength X, on a Tuesday in 2019?" And it instantly generates the perfect, mathematical shape of that distortion.

The Test: Catching a Moving Star

To prove their new method worked, they tested it on a specific star.

• The Mystery: When they measured the star's speed using the old "bell curve" method, the results were all over the place. Sometimes the star seemed to be moving fast, sometimes slow, with a sudden jump of 10 km/s in the middle of the data. It looked like the star was jumping around, which didn't make sense.
• The Fix: They used their new AI model to correct the data. They realized the telescope's "glasses" had actually changed shape between 2017 and 2021 (likely due to temperature changes).
• The Outcome: Once they corrected for this change using the AI's precise fingerprint, the "jump" disappeared. The star's speed measurements became smooth and consistent. The "noise" in the data dropped by about 3 km/s.

Why This Matters: Finding Hidden Twins

Why does saving 3 km/s matter?

Imagine you are trying to hear a whisper in a noisy room. If the room is too loud (high noise), you can't hear the whisper. But if you quiet the room down (reduce the noise), you can hear it.

In astronomy, finding binary stars (two stars orbiting each other) is like listening for that whisper. If the stars orbit very slowly (long periods), their movement is tiny. If your telescope's "noise" is too high, you miss them.

By using this AI to perfectly clean up the telescope's distortion, the scientists have lowered the noise floor. This means they can now detect long-period binary stars that were previously invisible. It's like upgrading from a radio with static to a high-fidelity speaker system.

Summary

• The Issue: Telescopes distort light in complex, changing ways that simple math can't fix.
• The Fix: An AI model was trained on calibration lights to learn the exact shape of this distortion for any fiber, color, or time.
• The Win: This AI correction removed errors in speed measurements, making the data much cleaner and allowing astronomers to find faint, slow-moving star systems that were previously hidden in the noise.

Technical Summary

Here is a detailed technical summary of the paper "Characterizing the Instrumental Profile of LAMOST":

1. Problem Statement

The Instrumental Profile (IP) of a spectrograph describes how the instrument broadens and distorts intrinsically narrow spectral lines. Accurate IP characterization is critical for:

• Wavelength Calibration: Determining the precise center of spectral lines.
• Stellar Parameter Measurement: Convolving template spectra with the IP to match instrumental resolution.
• Radial Velocity (RV) Precision: Small deviations in IP shape can introduce biases in derived quantities like RV.

Specific Challenges with LAMOST:

• Complexity: The IP is influenced by numerous factors, including optical system conditions, detector response, and environmental variables (temperature, humidity), leading to asymmetric and time-varying profiles.
• Limitations of Traditional Methods: LAMOST currently approximates the IP with a symmetric Gaussian function. This fails to capture the actual asymmetry and complex variations observed in the data.
• Data Sparsity: Unlike high-resolution spectrographs equipped with Laser Frequency Combs (LFCs) or Iodine cells, LAMOST relies on sparse, unevenly distributed arc lamp emission lines (Hg-Cd for blue, Ne-Ar for red), making traditional parametric modeling difficult.
• Scale: As a multi-fiber system (4,000 fibers), LAMOST requires distinct IP models for hundreds of fibers simultaneously, varying by wavelength and time.

2. Methodology

The authors propose a data-driven, machine learning approach using a Multi-Layer Perceptron (MLP) based on the The Payne neural network framework.

Data Processing:

• Dataset: Utilized low-resolution calibration lamp spectra from LAMOST (Sept 2011 – June 2024).
• Line Selection: Selected 7 emission lines in the blue arm (3700–5900 Å) and 15 in the red arm (5700–9000 Å) based on flux intensity and isolation.
• Preprocessing:
  • Defined wavelength windows around each line to capture full profiles.
  • Performed background subtraction using linear fits between local minima.
  • Resampled profiles to 600 points via spline interpolation and normalized peak flux to 1.
  • Filtered out broken fibers and cosmic ray-contaminated spectra.
• Feature Engineering: The model inputs include three features to capture dependencies:
  1. Observation Time: Minutes since Nov 17, 1858.
  2. Fiber ID: Spatial position (1–250).
  3. Central Wavelength: The specific emission line wavelength.
• Dataset Construction: Data was split into 416 independent datasets based on observing season, spectrograph (1–16), and spectral arm (blue/red) to account for hardware changes.

Model Architecture:

• Network: A feed-forward MLP implemented in PyTorch.
• Structure: Input layer $\rightarrow$ Fully connected hidden layers (LeakyReLU activation) $\rightarrow$ Linear output layer (predicting 600 flux points).
• Training:
  • Used a flux-dependent weighted loss (dividing residuals by flux) to prevent the model from ignoring profile wings.
  • Optimized using the Rectified Adam (RAdam) algorithm.
  • 80/20 train/validation split with early stopping based on validation loss.

3. Key Contributions

• First Comprehensive IP Model for LAMOST: Developed a neural network capable of retrieving the IP for any fiber, at any wavelength, and at any time within the observation period.
• Handling Asymmetry: Successfully modeled the non-Gaussian, asymmetric, and time-varying nature of the IP, which traditional Gaussian approximations miss.
• Open Source Release: The trained models and code are released as a Python package (available on GitHub and Zenodo).
• Methodological Validation: Demonstrated that machine learning can effectively interpolate complex, non-linear dependencies in spectrograph calibration data without explicit physical modeling of every environmental factor.

4. Results

• Model Accuracy:
  • Training Lines: The model achieves high precision, with mean residuals generally below 0.005 (0.5% of peak intensity) and a standard deviation of residuals around 0.006–0.007.
  • Generalization: When tested on an untrained emission line ([O I] 5577 Å), the model recovered the profile with residuals constrained below 0.03 (3% of peak intensity), proving its ability to interpolate between training wavelengths.
  • Outliers: The model identified specific fibers (e.g., Fiber 155) with abrupt profile changes or defects, where residuals were higher, confirming the model's sensitivity to physical anomalies.
• Radial Velocity (RV) Application:
  • Centering Methods: The authors tested six methods for defining the IP center (Max, Centroid, Gaussian fit, Sérsic fit, Half-flux, Fourier). The "Max" (peak) method was found to be the least stable.
  • RV Dispersion Reduction: Applying the derived IP (instead of a standard Gaussian kernel) and performing consistent wavelength recalibration reduced the dispersion of measured RVs by approximately 1–3 km/s.
  • Case Study: Analysis of a specific F7-type star revealed a ~10 km/s discontinuity in RV measurements between 2017 and 2021 data. This was traced to changes in the arc lamp profile shape over time. Recalibrating with the new IP model eliminated this jump and aligned the measurements with Gaia DR3 values.

5. Significance

• Enhanced Scientific Potential: By reducing RV dispersion by ~3 km/s, this work significantly improves the capability of LAMOST to detect long-period binary stars and other subtle RV variations that were previously obscured by instrumental noise.
• Foundation for Future Analysis: The accurate IP characterization lays the groundwork for more precise measurements of stellar parameters (temperature, gravity, metallicity) and chemical abundances in the vast LAMOST dataset.
• Scalability: The methodology is designed to be extended to LAMOST's medium-resolution spectra, offering a unified, data-driven framework for IP reconstruction across different spectral resolutions.

In summary, this paper moves beyond the limitations of symmetric Gaussian approximations for LAMOST, utilizing neural networks to create a high-precision, time-dependent IP model that directly enhances the telescope's radial velocity measurement capabilities.