Residual Modeling for High-Fidelity Learned Compression of Scientific Data

In three linesNew lossless compression approach for massive scientific data. Authors propose LBRC and NGLR, two residual coders improving existing GAE methods by 30-60% (LBRC) and additional 10-40% (NGLR) on E3SM, JHTDB, ERA5 datasets at 10^-6 to 10^-4 block-level NRMSE targets. NGLR adds causal neural predictor to reduce entropy of residual code.

Summary generated by Claude — human-verified