Short-Term Turbulence Prediction for Seeing Using Machine Learning

This study addresses short-term atmospheric turbulence forecasting by comparing statistical and deep learning models, demonstrating that a novel normalizing flow approach (FloTS) achieves the optimal balance between predictive accuracy and well-calibrated uncertainty for robust decision-making.

The Gist

Imagine you are trying to take a perfect photograph of a distant star. But there's a problem: the air between you and the star is wobbly. Hot and cold pockets of air are constantly shifting, making the starlight dance and blur. Astronomers call this "seeing," and it's the enemy of clear vision.

To fix this, telescopes use "Adaptive Optics"—basically, a super-fast mirror that bends itself to cancel out the wobble. But here's the catch: these mirrors are reactive. They wait for the wobble to happen, measure it, and then fix it. If the air changes too fast, the mirror is always a split-second behind.

This paper is about teaching computers to be proactive. Instead of waiting for the air to wobble, the authors built a system that can predict how the air will behave up to two hours in advance.

The Problem: The Wobbly Air

Think of the atmosphere like a giant, invisible ocean of air. Sometimes it's calm; sometimes it's stormy.

• The Goal: Predict the "smoothness" of this air (the seeing) so telescopes can adjust their mirrors before the image gets blurry.
• The Data: They used 15 years of data from a weather station on top of Mauna Kea (a mountain in Hawaii famous for its clear skies). They looked at how the air behaved every 10 minutes.

The Contestants: Four Different "Weather Forecasters"

The authors tested four different types of AI models to see which one could predict the air's mood best. Think of them as four different types of meteorologists:

1. The RNN (The Short-Term Memory):

   • Analogy: A student who can only remember what happened in the last few seconds. It looks at the immediate past to guess the immediate future.
   • Result: It was okay, but it forgot the bigger picture too quickly.

2. The LSTM (The Long-Term Memory):

   • Analogy: A student with a great memory who can remember events from hours ago. It's very good at spotting patterns and trends.
   • Result: This was the best at guessing the exact number. If the air was going to get blurry, this model said, "It will be 1.5 seconds of blur." It was the most accurate at the specific point prediction.

3. The GP (The Statistician):

   • Analogy: A cautious actuary. It doesn't just guess a number; it draws a "safe zone" around its guess. It says, "I think it will be 1.5, but I'm 95% sure it's between 1.2 and 1.8."
   • Result: It was very good at giving a "confidence interval," but it assumed the air behaves in a very simple, bell-curve way (like a perfect coin toss). Real air is messy and doesn't always follow a perfect bell curve.

4. The FloTS (The Shape-Shifter):

   • Analogy: This is the paper's new invention. Imagine a piece of clay. The GP tries to mold the clay into a perfect sphere. The FloTS, however, can mold the clay into any shape it needs to match the messy reality of the air. It learns the complex, weird shapes of the data.
   • Result: This was the winner. It was almost as accurate as the LSTM at guessing the number, but unlike the LSTM, it could also say, "Here is the range of possibilities, and here is how likely each one is."

The Big Discovery: Why "Uncertainty" Matters

The paper argues that just knowing the exact answer isn't enough. You need to know how confident the computer is.

• The "Safe Zone" Metaphor:
  Imagine you are driving a car in fog.
  • Model A (LSTM) says: "Turn left in exactly 50 meters." (Very precise, but if it's wrong, you crash).
  • Model B (FloTS) says: "Turn left somewhere between 40 and 60 meters, and I'm 90% sure it's safe."
  • Why Model B wins: If the computer says, "I'm not sure, the air is very chaotic right now," the telescope operator can decide to stop taking pictures and wait for better conditions, rather than wasting time on blurry images.

The "Calibration" Fix

The authors found that their models were sometimes too confident (over-confident) or not confident enough (under-confident).

• The Analogy: Imagine a weather app that says "100% chance of rain" every day, but it only rains 50% of the time. The app is "un-calibrated."
• The Fix: They developed a way to "tune" the models (like adjusting the temperature on a thermostat) so that when the model says "90% chance," it actually happens 90% of the time. This makes the predictions trustworthy for real-world decisions.

The Conclusion

The paper concludes that while the LSTM is great at guessing the exact number, the FloTS model is the best overall tool. It combines the accuracy of the LSTM with the ability to understand and predict the "messiness" of the atmosphere.

In simple terms: They built a crystal ball that doesn't just tell you what the weather will be, but also tells you how much you should trust that prediction. This helps astronomers and satellite operators make smarter, safer decisions about when to look at the stars or send data through the air.

Technical Summary

1. Problem Statement

Optical turbulence (OT), caused by atmospheric refractive index fluctuations, degrades the performance of ground-based optical systems, including astronomical telescopes and free-space optical (FSO) communication links. While Adaptive Optics (AO) systems correct turbulence in real-time, they are inherently reactive and struggle with rapidly changing conditions.

• Goal: Develop a predictive solution to forecast the atmospheric seeing parameter ($\epsilon$) up to two hours in advance.
• Challenge: Traditional numerical simulations are computationally too expensive for rapid short-term forecasting. Furthermore, deterministic machine learning models often fail to quantify predictive uncertainty, which is critical for robust decision-making in dynamic environments (e.g., telescope scheduling or link power control).

2. Methodology

The study utilizes historical seeing data from the Maunakea Weather Center (MKWC), collected via a Differential Image Motion Monitor (DIMM) between 2009 and 2024.

Data Preparation

• Sampling: Data was resampled to a uniform 10-minute interval via local linear regression within sliding bins to handle gaps and irregular sampling.
• Input/Output: The models were trained on input sequences ($\epsilon^-$) to predict future windows ($\epsilon^+$).
• Optimization: A 2-hour input window was determined to be optimal for all models, balancing temporal correlation with model complexity.
• Dataset: Approximately 71,906 input-output pairs were split into 81.5% training, 8.5% validation, and 10% testing.

Models Evaluated

The study compares four distinct approaches, all trained exclusively on historical seeing data (no external meteorological variables like wind or temperature were used):

1. Recurrent Neural Network (RNN): A deterministic encoder-decoder architecture using standard RNN layers. It serves as a baseline but suffers from the vanishing gradient problem.
2. Long Short-Term Memory (LSTM): A deterministic encoder-decoder architecture using LSTM cells. It addresses the vanishing gradient issue via gating mechanisms to retain long-term dependencies.
3. Gaussian Process (GP): A probabilistic statistical model. It uses an exponential kernel to model temporal correlations and provides predictions with confidence intervals. It assumes a Gaussian distribution.
4. FloTS (Flow for Time Series): A novel probabilistic deep learning model based on Normalizing Flows (specifically Masked Autoregressive Flow).
   • Mechanism: It learns a bijective mapping from a simple latent Gaussian distribution to a complex target distribution conditioned on past observations.
   • Advantage: Unlike GPs, it does not assume Gaussianity and can model complex, non-Gaussian distributions and multi-modal uncertainties.

Calibration

To ensure the reliability of probabilistic outputs, the authors applied a temperature scaling calibration technique. This adjusts the predictive distributions (for GP and FloTS) to align empirical coverage with nominal credibility levels, correcting for over- or under-confidence.

3. Key Contributions

• Probabilistic Framework: The study moves beyond deterministic point forecasts by integrating probabilistic models (GP and FloTS) that quantify uncertainty, a crucial feature for operational risk assessment.
• Novel Architecture (FloTS): The implementation of Normalizing Flows for time series (FloTS) in the context of atmospheric seeing prediction, demonstrating its ability to capture non-Gaussian characteristics of turbulence.
• Fair Comparison: All models were trained on identical data conditions (historical seeing only) with optimized hyperparameters, isolating the predictive capability of the architectures themselves.
• Calibration Analysis: The paper provides a rigorous analysis of calibration, showing that while GPs are principled, their Gaussian assumption limits their ability to model the true distribution of seeing, whereas FloTS offers superior flexibility.

4. Results

• Predictive Accuracy (Point Forecasting):
  • LSTM achieved the lowest Root Mean Square Error (RMSE) and highest Pearson correlation, slightly outperforming other models in point prediction accuracy.
  • FloTS performed comparably to LSTM, with only a marginal increase in RMSE.
  • RNN lagged behind in both metrics, confirming its limitation in modeling complex temporal dependencies.
  • GP showed strong short-term accuracy but slightly underperformed compared to deep learning models, likely due to the Gaussian assumption not fully capturing the seeing distribution.
• Uncertainty Quantification:
  • FloTS provided the most robust uncertainty estimates. Its predictive density successfully captured non-Gaussian features (e.g., heavy tails, skewness) and local variations in the data.
  • GP provided well-calibrated intervals but was constrained by its Gaussian shape, leading to over-confidence in certain regimes (under-dispersion) that required significant calibration.
  • Case Studies: Visualizations of specific nights showed that while RNN/LSTM provided reasonable point forecasts, FloTS accurately tracked the ground truth trajectory within high-probability density regions, even during rapid fluctuations.
• Input Window Sensitivity: A 2-hour input window was found to be optimal. Longer windows did not significantly improve performance, suggesting that seeing conditions under these specific site conditions do not exhibit long-term dependencies beyond this horizon.

5. Significance and Conclusion

This study demonstrates that Machine Learning is a viable and efficient alternative to physical simulations for short-term turbulence forecasting.

• Operational Impact: The ability to forecast seeing with quantified uncertainty allows for proactive adjustments in telescope scheduling (e.g., prioritizing high-seeing targets) and FSO communication (e.g., adjusting power or beam tracking).
• Best Overall Model: While LSTM is superior for pure point accuracy, FloTS is identified as the most informative and robust model overall. It offers the best trade-off between predictive accuracy and the ability to provide realistic, non-Gaussian uncertainty quantification.
• Future Work: The authors suggest exploring the inclusion of meteorological features (wind, temperature) and advanced architectures like Transformers or Conditional Variational Autoencoders (VAEs) to further enhance predictive capabilities.

In summary, the paper establishes FloTS as a state-of-the-art tool for atmospheric seeing prediction, bridging the gap between high-accuracy deep learning and rigorous probabilistic uncertainty estimation.