EMBER: Machine-Learning Detection of Modulated Ion Acoustic Waves and Associated Core-Electron Heating in the Solar Wind with Parker Solar Probe

This paper introduces EMBER, an open-source machine-learning pipeline that successfully automates the detection of modulated ion acoustic waves in Parker Solar Probe data, achieving high accuracy while confirming their role in driving core-electron heating without relying on temperature data for identification.

The Gist

Imagine the space around our Sun is like a vast, chaotic ocean. This isn't water, but a super-hot gas called plasma, constantly rushing outward as the "solar wind." In this ocean, tiny ripples and waves are constantly crashing into particles, heating them up. Scientists have long suspected that specific types of these waves—called Modulated Ion Acoustic Waves—are the secret chefs cooking up extra heat for electrons in this solar soup.

However, finding these specific waves in the data is like trying to find a single, specific type of seashell in a massive, noisy beach that stretches for miles.

The Problem: Too Much Data, Too Few Eyes

The Parker Solar Probe (PSP) is a spacecraft that flies closer to the Sun than anything ever has. It's equipped with a super-sensitive microphone (the FIELDS instrument) that records the "sound" of the solar wind. But it records so much data that if scientists tried to listen to every second of it by eye, they would never finish.

Previously, experts had to manually look at the data charts (spectrograms) to spot these special waves. It was slow, tedious, and couldn't scale to the entire mission.

The Solution: EMBER (The Smart Wave Hunter)

The authors of this paper created a new tool called EMBER. Think of EMBER as a highly trained, open-source robot detective. Its job is to scan the massive library of solar wind recordings and flag the moments where these special waves appear.

Here is how EMBER works, using a few simple analogies:

1. Turning Sound into a Picture
First, EMBER takes the raw voltage data (the "sound") and turns it into a colorful picture called a spectrogram. Imagine a piano roll where the horizontal axis is time and the vertical axis is pitch.

• The Trick: EMBER doesn't just look at the picture normally. It zooms in and out simultaneously (using "log-log" scaling). This is like having a pair of glasses that can see both the tiny, high-pitched squeaks and the deep, low-pitched rumbles clearly at the same time. This makes the special waves look like a distinct "ladder" pattern or a rapid "chirp" that stands out from the background noise.

2. The Detective Squad (The Ensemble)
Instead of relying on just one detective, EMBER uses a squad of 16 different detectors.

• The Physics Detectives: These look for specific patterns based on how waves should behave according to the laws of physics.
• The "Odd-One-Out" Detectives: These are classic math tools that ask, "Does this picture look weird compared to the millions of normal, boring pictures we've seen before?"
• The AI Detectives: These are deep-learning models (like the ones that recognize cats in photos) that have been trained to recognize the "texture" of these waves, even if they've never seen a solar wave before.

3. The "Background-Only" Training
Here is the clever part: EMBER was never shown the special waves during its training. It only studied millions of "normal" solar wind moments. It learned what "boring" looks like.

• Analogy: Imagine a security guard who has memorized the face of every normal visitor to a building. If a stranger walks in, the guard doesn't need to know who the stranger is; they just know, "This person doesn't look like anyone I've seen before."
• This prevents the AI from getting confused or memorizing the wrong things. It simply flags anything that looks "anomalously different" from the background.

4. The Teamwork (Ensembling)
Each of the 16 detectives votes. Some are very strict (they only flag things they are 100% sure of), while others are more sensitive. EMBER combines all these votes into a final decision.

• The Result: The system found 93% of the known special waves that human experts had previously identified.
• The Cost: It only made a mistake (a "false alarm") about 1 time in every 100 checks. This is a very low error rate for such a difficult task.

The Proof: Does it Actually Heat Things Up?

The authors didn't just stop at finding the waves. They wanted to prove that finding these waves actually meant the electrons were getting hotter.

They checked the data from the spacecraft's other instruments (SWEAP/SPAN), which measure the temperature of the electrons. Crucially, the temperature data was never used to teach EMBER how to find the waves. It was a completely independent check.

• The Finding: Every time EMBER flagged a wave event, the electrons in that spot were indeed hotter than expected. They were warmer than they should be if they were just cooling down naturally as they moved away from the Sun.
• The Metaphor: It's like a smoke detector that beeps whenever it smells smoke. The authors checked the kitchen and confirmed that yes, there was indeed a fire burning. The detector didn't need to know about the fire to do its job; it just needed to know what "normal air" smells like.

Summary

The paper introduces EMBER, a smart, open-source tool that automatically finds specific, heat-generating waves in the solar wind. By using a team of 16 different AI and math-based detectors that only learned what "normal" looks like, it successfully found 93% of these rare events with very few mistakes. Most importantly, it confirmed that whenever these waves are found, the solar wind electrons are getting a significant heat boost, solving a puzzle about how the Sun's atmosphere stays so hot.

Technical Summary

1. Problem Statement

Context: Modulated ion acoustic waves (IAWs), including triggered ion acoustic waves (TIAWs) and frequency-dispersed ion acoustic waves (FDIAWs), are recognized as efficient drivers of electron heating in the solar wind via nonlinear wave-particle interactions.
Challenge: The Parker Solar Probe (PSP) generates massive amounts of high-cadence burst-mode data (FIELDS Digital Burst Memory or DBM). Identifying these specific wave events currently relies on expert visual inspection of spectrograms. This manual approach is:

• Non-scalable: It cannot process the full mission archive.
• Non-reproducible: Subjective to human interpretation.
• Data-Scarce: Anomalous events are rare compared to background turbulence, making standard supervised classification prone to overfitting and label contamination.

Goal: Develop a scalable, reproducible, open-source pipeline to automatically detect modulated IAWs in PSP data without relying on electron temperature data during the detection phase, while verifying that detected events correlate with physical heating signatures.

2. Methodology: The EMBER Pipeline

EMBER (Electron heating from Modulated Burst-mode Event Recognition) is an open-source Python pipeline designed as a background-only anomaly detection system. It operates in three main stages:

A. Data Preprocessing & Representation

• Input: Level-2 FIELDS DBM voltage bursts (DVAC/VAC).
• Transformation: Raw time-series data is converted into log-scaled Fourier spectrograms (log-log axes).
  • Rationale: Modulated IAWs exhibit a low-frequency envelope modulating a high-frequency carrier. Log-log scaling renders the characteristic "ladder-of-sidebands" morphology visible, which linear axes compress into unresolved bands.
• Feature Engineering: A "Physics-motivated feature bank" is extracted, including:
  • Physics features: Band-integrated power, spectral flatness, centroid, and bandwidth.
  • Coupling features: Cross-band power ratios, amplitude modulation depth, and coherence.
  • Augmentation: A PhysicsAugmenter applies frequency jitter and noise to these features during training to improve robustness.

B. The Detector Suite (16 Detectors)

The system employs an ensemble of 16 detectors trained exclusively on background (non-anomalous) data. No anomaly labels are used during the training of individual detectors. The suite is grouped into four families:

1. Physics-motivated:
   • BandDeviation: Models background power distribution as a heavy-tailed density to detect narrowband intensification.
   • LocalPatch: Uses overlapping time-frequency patches to detect localized coupled-wave structures that don't alter global power.
2. Classical Density-based (on Feature Bank):
   • Includes Isolation Forest, Local Outlier Factor (LOF), k-NN distance, Robust Mahalanobis distance, and PCA reconstruction error.
3. Deep Generative/Reconstruction (on Spectrograms):
   • Convolutional Autoencoder (AE), $\beta$-Variational Autoencoder (VAE), and Masked Autoregressive Flow (MAF/Zuko).
4. Pretrained Backbone (Zero-shot):
   • Leverages frozen vision models (PatchCore/WideResNet-50, ResNet-50, DINOv2) trained on ImageNet or self-supervised data. These transfer well to spectrogram features without fine-tuning on PSP data.

C. Ensembling & Calibration

• Rank Normalization: Scores from all detectors are normalized to $[0, 1]$ based on their rank within a held-out background calibration set ($N=100$).
• Optimization: A weighted ensemble (EnsEmber) is optimized via differential evolution to maximize the True Positive Rate (TPR) at a fixed False Alarm Rate (FAR) of 1%.
• Union Strategy: A "zero-extra-FP union" is constructed by adding secondary detectors at thresholds that contribute no additional false positives, maximizing recall without sacrificing the FAR budget.

3. Key Contributions

1. First Scalable Detection Pipeline: Introduced EMBER, the first automated, open-source pipeline capable of processing the full PSP FIELDS DBM archive for modulated IAWs.
2. Background-Only Learning: Demonstrated that high-performance detection of rare plasma events is achievable without labeled anomaly data for training, mitigating overfitting and label bias.
3. Hybrid Architecture: Successfully combined physics-based heuristics, classical statistical outlier detection, deep generative models, and pretrained computer vision backbones into a single robust ensemble.
4. Independent Physical Validation: Proved that the ML-detected events correlate with independent physical measurements (electron heating) without those measurements being used as inputs to the detector.

4. Results

The pipeline was benchmarked on a curated catalog of 538 spectrograms (496 background, 42 anomalies) from PSP Encounters 6–9 and 15.

• Detection Performance:
  • Recall: The full ensemble (EnsEmber + 15 zero-FP secondary detectors) recovered 93% of all anomalous events (39 out of 42).
  • False Alarm Rate: Maintained at 1% (1 false positive per 100 background samples).
  • Class-Specific Recovery:
    • Label 1 (Strongly modulated TIAWs): 84.2% recovery.
    • Label 2 (FDIAWs with weak heating): 100% recovery.
  • Missed Events: Three events remained undetected; these were characterized by low-amplitude modulation or embedded in high-turbulence backgrounds.
• Physical Validation (Coincident Heating):
  • Flagged intervals were cross-referenced with SWEAP/SPAN electron data.
  • Result: Events flagged by EMBER showed core perpendicular electron temperatures ($T_\perp$) significantly above adiabatic cooling expectations and elevated electron-to-ion temperature ratios ($T_e/T_i \gtrsim 2$).
  • This confirms that the ML-detected spectral anomalies correspond to the physical phenomenon of preferential electron heating, validating the pipeline's scientific utility.

5. Significance

• Scientific Impact: Enables statistical studies of wave-driven heating across the entire PSP mission, moving beyond case studies to global understanding of solar wind thermodynamics.
• Methodological Advance: Establishes a blueprint for using "background-only" anomaly detection in space physics, where labeled rare events are scarce.
• Future Integration: The modular design allows for the integration of Foundation Models (e.g., NASA's Surya Helio Foundation Model). Future work will concatenate global coronal magnetic embeddings (from SDO) with local in-situ features to predict heating outcomes from remote sensing data alone, bridging the gap between global coronal context and local kinetic signatures.
• Open Science: The code, data, and trained weights are publicly available, fostering reproducibility and community extension.