peijin94/PhySynthTrainer

Physics-based Radio Burst Training-set Event Detection

This dataset contains 50,000 radio dynamic spectrum images produced using a physics-informed solar radio burst simulator, designed for training modern computer vision models (e.g., YOLO, CNNs, ViTs, diffusion models) to detect fine-scale radio burst features at low frequencies.

It includes 1,092,982 labeled events, spanning Type III / Type IIIb / spike-like and noise storm structures.
The simulations follow realistic flux, duration, bandwidth, drift rate, and turbulence-informed morphology distributions derived from observations at LOFAR and OVRO-LWA.

---

📊 Flux Density Distribution

The simulated event amplitudes follow a power-law distribution consistent with solar radio burst statistics:

This ensures realistic class imbalance and naturally high occurrence of weaker bursts.

---

Physics Background

The simulator uses:

• Plasma emission physics for fundamental/harmonic radiation
• Density models for frequency–height mapping
• Turbulence-based fine-structure generation
• Drift-rate distributions from empirical solar burst measurements
• Stochastic modeling for multi-lane and fragmented bursts

The goal is to bridge physics-based generative models with machine learning detection tasks.

---

Intended Use Cases

• Solar radio burst detection (YOLO, CNNs, transformers)
• Real-time space weather event classifiers
• Benchmarking ML models on fine-structure detection
• Data augmentation for low-frequency radio solar instruments
• Cross-instrument generalization (e.g., LOFAR → OVRO-LWA → SKA-Low)

---

Loading the Dataset in Python

from datasets import load_dataset

ds = load_dataset("peijin94/PhySynthTrainer")

sample = ds["train"][0]
img = sample["image"]
labels = sample["yolo_label"]