Machine learning assisted analysis of visible spectroscopy in pulsed-power-driven plasmas

Rishabh Datta, Faez Ahmed, Jack HareIEEE Transactions on Plasma Science 2023

This paper, recently published by researchers at MIT, explores how machine learning (ML) can revolutionize the way we diagnose high-energy-density (HED) plasmas. By training models on synthetic data, the authors demonstrate that we can predict critical plasma parameters—like density and temperature—almost instantaneously, a task that previously required hours of supercomputing time or complex manual analysis.

TL;DR

Researchers developed a machine learning pipeline to analyze visible light spectra from pulsed-power-driven plasmas with ~98% accuracy. By bypassing slow radiation transport simulations, these models enable rapid, potentially real-time diagnostics of plasma density and temperature during complex experiments.

What Problem Does This Paper Address?

In high-energy-density physics—think exploding wires or Z-pinch experiments—scientists need to know the ion density ($n_i$) and electron temperature ($T_e$) of the plasma they’ve created. They usually find this by looking at the "fingerprint" of light the plasma emits (spectroscopy).

However, there’s a catch: these plasmas are often "optically thick," meaning the light they emit is partially re-absorbed by the plasma itself before it reaches the spectrometer. To account for this, researchers must solve complex radiation transport equations. Traditionally, this involves "curve-fitting" using Genetic Algorithms that can take over 10 minutes per spectrum—far too slow for real-time analysis or processing the massive datasets generated by modern experiments.

Methods

The authors took a "synthetic-to-real" approach:

  1. Data Generation: They used a collisional-radiative code called PrismSPECT to calculate how aluminum ions emit and absorb light at various temperatures and densities.
  2. Radiation Transport: They fed those values into an in-house 1D solver to simulate how light travels through a spatially inhomogeneous plasma (specifically, one with a Gaussian density profile).
  3. ML Training: They compared several models, including Linear Regression, K-Nearest Neighbors (KNN), Random Forests, Deep Neural Networks (DNN), and AutoGluon (an automated ML framework).
  4. Sensitivity Analysis: They used gradient-based sensitivity maps to identify which specific spectral lines the models were using to make their "decisions."

Datasets

The study utilized a newly generated synthetic dataset:

  • Size: 2,500 intensity spectra, each 400 dimensions (representing different wavelengths).
  • Parameters: Electron temperatures ($T_e$) from 0.5 to 25 eV and ion densities ($n_i$) from $10^{16}$ to $10^{19}$ cm⁻³.
  • Target Variables: Peak density ($n_0$), temperature ($T_0$), and the spatial spread of the plasma ($\sigma$).

Key Insights and Ideas

  • The "Aha" on Optical Thickness: The paper demonstrates that ignoring radiation transport (assuming the plasma is "optically thin") leads to massive errors in line ratios. The ML models successfully "learned" the non-linear relationship between density and re-absorption.
  • Shape Over Intensity: The models were trained on normalized spectra. This means they rely on the relative heights and widths of lines rather than absolute brightness, making the tool much more robust for experimentalists who may not have perfectly calibrated equipment.
  • Spectral "Thermometers": Through sensitivity mapping, the authors found that the Al-III 570.7 nm line is the primary feature the model uses to predict temperature, while the Al-III 448.1 nm and 452.5 nm lines (which merge at high density) are the "go-to" features for density.

Key Results

  • Top Performer: AutoGluon was the clear winner, achieving an $R^2$ score of ~98% for density and temperature predictions.
  • Simpler Models Hold Up: Even "simpler" models like Random Forest and KNN achieved $R^2$ scores >90%, suggesting the underlying physics is well-captured by the spectral features.
  • Speedup: While traditional optimization (Genetic Algorithms) took ~10 minutes per prediction, the trained ML models provided answers in milliseconds.
  • Robustness: The models remained effective even when researchers added artificial noise and "stray lines" (simulating impurities) to the test data.

Limitations and Future Work

  • The Simulation Gap: The models are currently trained on "perfect" synthetic data. In real-world experiments, factors like the instrument's response function and optical attenuation need to be baked into the training set to maintain accuracy.
  • Spatial Assumptions: The study assumes a constant temperature along the line of sight. Real plasmas often have temperature gradients, which would require a more complex multi-variable regression approach.
  • Next Steps: The authors plan to apply this framework to experimental data from pulsed-power facilities, moving from "lab-bench" validation to real-world deployment.

Summary updated

Abstract

We use machine learning models to predict ion density and electron temperature from visible emission spectra, in a high energy density pulsed-power-driven aluminum plasma, generated by an exploding wire array. Radiation transport simulations, which use spectral emissivity and opacity values generated using the collisional-radiative code PrismSPECT, are used to determine the spectral intensity generated by the plasma along the spectrometer's line of sight. The spectra exhibit Al-II and Al-III lines, whose line ratios and line widths vary with the density and temperature of the plasma. These calculations provide a 2500-size synthetic dataset of 400-dimensional intensity spectra, which is used to train and compare the performance of multiple machine learning models on a 3-variable regression task. The AutoGluon model performs best, with an R2-score of roughly 98% for density and temperature predictions. Simpler models (random forest, k-nearest neighbor, and deep neural network) also exhibit high R2-scores (>90%) for density and temperature predictions. These results demonstrate the potential of machine learning in providing rapid or real-time analysis of emission spectroscopy data in pulsed-power-driven plasmas.
BibTeX
@article{datta2023machine,
  author = {Rishabh Datta and Faez Ahmed and Jack Hare},
  title = {{Machine learning assisted analysis of visible spectroscopy in pulsed-power-driven plasmas}},
  journal = {IEEE Transactions on Plasma Science},
  year = {2023},
}
Figure