Neural Network Compression of Spatio-Temporal Data

This work investigates coordinate-based neural networks for concurrent, real-time compression of streaming scientific simulation data. Multi-layer perceptrons learn continuous mappings from spatial coordinates and time directly to field variables, compressing megabytes of raw simulation output into kilobytes of network weights. The framework is evaluated under two training paradigms: offline batch training on complete datasets achieving PSNR up to 35.72 dB, and online incremental training with temporal windows that simulates real-time in-situ compression. To combat catastrophic forgetting in the streaming setting, continual learning strategies including Experience Replay are investigated. The approach is validated on a vortex shedding CFD dataset comprising approximately 7.9 million spatio-temporal samples across 300 timesteps (30 representative timesteps visualized here).