An Interpretable Operator-Learning Model for Electric Field Profile Reconstruction in Discharges Based on the EFISH Method

This paper introduces Decoder-DeepONet (DDON), a novel interpretable operator-learning model that significantly outperforms previous neural network and classical methods in reconstructing electric field profiles from EFISH signals by offering superior accuracy, generalizability, and robustness to incomplete data while identifying optimal sampling windows.

The Gist

Here is an explanation of the paper using simple language and creative analogies.

The Big Picture: Seeing the Invisible

Imagine you are trying to figure out the shape of a hidden object inside a foggy room. You can't see the object directly, but you can shine a flashlight through the fog. As the light passes through, it gets distorted in a specific way. By looking at how the light comes out the other side, you want to guess what the hidden object looks like.

In the world of physics, scientists use a laser to measure electric fields inside a plasma (a super-hot, glowing gas used in things like neon signs or fusion reactors). The laser creates a signal called EFISH.

• The Problem: The laser beam is wide and fuzzy (like a flashlight beam). The signal it picks up is a "blurry average" of the electric field along the whole path. It's like trying to guess the shape of a mountain just by looking at the shadow it casts on a wall from far away. The shadow tells you something is there, but it's hard to know the exact shape of the mountain.
• The Old Solution: Scientists used a "Convolutional Neural Network" (CNN), which is like a smart student who memorized thousands of examples of mountains and their shadows. It works great if the mountain looks exactly like the ones it studied. But if the mountain is a weird, new shape, the student gets confused and makes mistakes.

The New Hero: The "Decoder-DeepONet" (DDON)

This paper introduces a new, super-smart AI model called DDON. Think of it not as a student who memorized pictures, but as a master architect who understands the laws of physics behind how shadows are cast.

Here is why DDON is special:

1. It Learns "Shapes," Not Just Pictures

• The Old Way (CNN): Imagine a chef who only knows how to bake a perfect round cake. If you ask for a square cake, they fail.
• The New Way (DDON): This chef understands the concept of baking. They know how to turn flour, eggs, and heat into any shape. DDON learns the mathematical relationship between the "shadow" (the laser signal) and the "object" (the electric field). Because it understands the rules rather than just memorizing examples, it can reconstruct electric fields it has never seen before, even if they look totally different from its training data.

2. It's a "Function-to-Function" Translator

Most AI models take a list of numbers and spit out a single number (like predicting tomorrow's temperature).
DDON is different. It takes a whole curve (the blurry laser signal) and outputs a whole new curve (the sharp electric field).

• Analogy: Imagine a translator who doesn't just translate word-for-word but understands the entire story in one language and writes a completely new, coherent story in another language. DDON translates the "blurry story" of the laser into the "sharp story" of the electric field.

3. It Works Even with "Bad" Data

In real experiments, data is often messy. You might have gaps in your measurements, or the signal might be noisy (like static on a radio).

• The Old Way: If you miss a few pages of a book, the old AI might get lost.
• The New Way: DDON is like a detective who can solve a mystery even if half the clues are missing. Because it understands the underlying structure of the data, it can fill in the gaps and still give you a very accurate picture of the electric field, even if you only measured a small slice of the signal.

The "Magic Window" (Integrated Gradients)

One of the coolest parts of this paper is how the authors made the AI "explain itself." They used a tool called Integrated Gradients.

• The Analogy: Imagine you are taking a test. The teacher (the AI) gives you a grade. You ask, "Why did I get an A?" The teacher uses Integrated Gradients to highlight exactly which answers on your test were the most important for that grade.
• The Discovery: The AI told the scientists: "You don't need to measure the whole laser beam to get a good answer. You only need to measure the middle part, specifically within a certain distance from the center."
• The Result: They found a "Golden Window." If you measure the electric field within 4.2 times the width of the signal's peak, you get a perfect reconstruction. This saves scientists time and money because they don't need to scan the entire area; they just need to focus on the "key" area.

Why Does This Matter?

1. Better Plasma Control: Electric fields control how plasma behaves. If we can measure them accurately, we can build better fusion reactors, more efficient industrial lasers, and better medical devices.
2. Solving the "Inverse Problem": This is a classic math headache: "I have the result, what was the cause?" This paper solves that headache for electric fields with a tool that is robust, accurate, and doesn't break when the data is messy.
3. Trustworthy AI: By using the "Magic Window" explanation, the scientists aren't just trusting a "black box." They know why the AI is making its predictions, which makes the results reliable for real-world science.

Summary

The authors built a new AI (DDON) that acts like a master architect for electric fields. Unlike previous models that just memorized examples, this one understands the rules of the game. It can reconstruct invisible electric fields from blurry laser signals, even when the data is noisy or incomplete. Best of all, it told the scientists exactly where to look to get the best results, turning a complex physics problem into a manageable, efficient experiment.

Technical Summary

Here is a detailed technical summary of the paper "An Interpretable Operator-Learning Model for Electric Field Profile Reconstruction in Discharges Based on the EFISH Method."

1. Problem Statement

The paper addresses the "inverse EFISH problem": reconstructing the spatial distribution of an electric field ($E_{ext}(z)$) in a plasma discharge from a measured Electric Field Induced Second Harmonic (EFISH) signal profile ($P(2\omega)(z)$).

• Physical Challenge: EFISH is a non-intrusive optical diagnostic where a focused laser beam interacts with a plasma. The resulting signal is not a local measurement but an integral along the beam path (line-of-sight). This integral is complicated by the Gouy phase shift and phase-mismatch effects, meaning the signal depends on the entire axial shape of the electric field rather than just the local value.
• Limitations of Existing Methods:
  • Classical Mathematical Approaches: Methods like Fourier deconvolution struggle because the EFISH forward mapping involves a modulus operation that erases phase information, making the inverse problem non-linear and ill-posed.
  • Previous Machine Learning (CNN): The authors' prior work used Convolutional Neural Networks (CNNs). While effective for specific training data, CNNs suffered from limited generalizability. They were trained on a specific "Fuzzy" function family and failed to accurately reconstruct profiles from different mathematical families (e.g., Lorentzian, Voigt, Bump) or handle incomplete/noisy data effectively.
  • Experimental Constraints: There is no clear guideline on the optimal spatial sampling window required to reconstruct the field, leading to inefficient data acquisition.

2. Methodology

The authors propose a novel Decoder-DeepONet (DDON) architecture, an operator-learning framework designed to learn mappings between function spaces rather than fixed vector-to-vector mappings.

A. Architecture: Decoder-DeepONet (DDON)

Unlike standard DeepONets that use a simple dot product to combine branch and trunk networks, DDON introduces an encoder-decoder structure to enhance feature extraction and stability:

• Branch Network: Encodes the input EFISH signal profile ($P_{norm}^{(2\omega)}$). It utilizes a ResNet (3 residual blocks) followed by a Vanilla Autoencoder (AE). The AE compresses the signal into a 32-dimensional bottleneck (capturing global context and local variations) and expands it to a 512-dimensional latent vector ($\beta$).
• Trunk Network: Encodes the observation locations ($z'$) and optical parameters (wave-vector mismatch $u'$). It uses a similar ResNet backbone to produce a 512-dimensional latent vector ($\tau$).
• Operator Head & Decoder: The branch and trunk latents are combined via a dot product to form a 1D latent response. A Decoder (fully connected layers) then "unembeds" this response to reconstruct the full electric field profile ($E_{ext}'(z')$).
• Key Advantage: This architecture explicitly separates the learning of the signal shape (branch) from the spatial/optical parameters (trunk), allowing for robust function-to-function mapping across diverse profile shapes.

B. Data Generation and Augmentation

• Training Data: The model was trained on 703,682 profile pairs generated via the forward EFISH equation. The dataset included two primary function families: Fuzzy (from previous work) and Voigt (a convolution of Gaussian and Lorentzian), covering single-peak and double-peak symmetric profiles.
• Augmentation: To improve robustness, the training set included:
  • Random Cropping: 20% of inputs were cropped to simulate incomplete experimental data.
  • Noise Injection: Gaussian noise (5% amplitude) was added to 20% of inputs to mimic sensor noise.

C. Interpretability via Integrated Gradients (IG)

To address the "black box" nature of ML, the authors applied Integrated Gradients (IG), an Explainable AI (XAI) technique.

• IG calculates the contribution of each input point in the EFISH signal to the final prediction.
• This was used to identify a "key-feature window"—the specific region of the signal that most critically influences the reconstruction accuracy.

3. Key Contributions

1. Novel Architecture (DDON): Introduction of a ResNet-AE-DeepONet hybrid that outperforms standard CNNs in learning complex operator mappings for EFISH inversion.
2. Cross-Family Generalizability: The model demonstrates superior performance when tested on unseen function families (Lorentzian, Quartic, Laplacian, Raised Cosine, Bump) that were not part of the training set.
3. Noise and Data Sparsity Robustness: DDON maintains high accuracy even with significant noise (SNR = 15) and extremely sparse data (as few as 8 spatial points), whereas classical methods and the previous CNN model fail under these conditions.
4. Actionable Sampling Guidelines: Using IG, the authors quantified a universal sampling window ($\mathcal{K}_{HWHM} \approx 4.2$). This provides experimentalists with a concrete rule: sampling within $\pm 4.2$ times the Half-Width at Half-Maximum (HWHM) of the signal is sufficient for accurate reconstruction, optimizing experimental time and resources.
5. Validation Framework: The study validates predictions by performing a "forward transform" (re-calculating the EFISH signal from the predicted field) to compare against experimental data, even when the "true" ground truth is unknown.

4. Results

• Generalization: On six distinct function families (four unseen during training), DDON achieved significantly lower Mean Squared Error (MSE) compared to the retrained CNN. For example, on the "Bump" function, DDON achieved ~86% accuracy (MSE $\le 4 \times 10^{-3}$) compared to the CNN's poor performance.
• Noise Robustness: Under SNR = 15, DDON maintained high accuracy across most families, while the CNN's performance degraded drastically (e.g., dropping to ~9.6% accuracy on the Bump function). DDON also showed resilience to Perlin and Poisson noise.
• Simulation & Experimental Validation:
  • Simulations: Successfully reconstructed fields from a pin-plane fluid model, matching ground truth even when using only the IG-identified key-feature window.
  • Experiments: Applied to nanosecond pulsed corona discharges. DDON accurately reconstructed electric fields from experimental data with sparse sampling (8 points), outperforming the CNN which required extrapolation and showed significant deviations.
• Interpretability: The IG analysis confirmed that the model relies heavily on the central region of the signal (near the focus) and the peaks, validating the derived sampling window.

5. Significance

This work represents a significant advancement in plasma diagnostics by:

• Solving a Non-Linear Inverse Problem: Providing a robust, data-driven solution to the EFISH inversion problem where classical mathematical methods fail due to phase loss and non-linearity.
• Bridging Simulation and Experiment: Demonstrating that a model trained on synthetic data can generalize to real-world, noisy experimental conditions with sparse data.
• Optimizing Experimental Design: Moving beyond "black box" prediction to provide interpretable, physics-informed guidelines for data acquisition. The identification of the $\mathcal{K}_{HWHM} = 4.2$ window allows researchers to design more efficient experiments without sacrificing accuracy.
• Future Applicability: While currently limited to symmetric, bell-shaped profiles, the framework establishes a pathway for reconstructing complex, non-equilibrium plasma electric fields, potentially extendable to asymmetric profiles with further training.

In summary, the DDON model offers a robust, generalizable, and interpretable tool for reconstructing electric field profiles in plasmas, overcoming the limitations of previous neural network approaches and providing practical guidance for experimental implementation.