A new approach to learned compression for scientific data, focusing on residual modeling, has been developed, significantly improving compression ratios. The proposed methods, LBRC and NGLR, outperform existing techniques, indicating a potential shift in data compression strategies for high-fidelity applications.