Workshops Archive

PRISM: Lightweight Surrogation for Lossy Compression Quality Prediction


Workshop: PDSW’26: The 11th International Parallel Data Systems Workshop

Authors: Adelina Zheng (Brown University), Dong Dai (University of Delaware), Sheng Di and Franck Cappello (Argonne National Laboratory), and Jinzhen Wang (University of North Carolina at Charlotte)

Abstract: Error-bounded lossy compression is indispensable for managing the data deluge in exascale scientific workflows, yet selecting compressor configurations requires evaluating quality metrics (CR, PSNR, SSIM) that are expensive to compute directly. Deep surrogate models address this bottleneck but must be retrained from scratch whenever new data fields, compressors, or metrics are introduced, creating a combinatorial scaling problem. We present PRISM, a lightweight framework that trains a single foundation backbone and adapts to new tasks via Low-Rank Adaptation (LoRA), removing the need for full retraining. Evaluated on the Hurricane Isabel application across eight fields, four compressors, and three metrics, PRISM achieves MAPE below 10% in 81 of 96 tasks using rank-4 adaptation with under 1% trainable parameters (107x reduction), and is more accurate than from-scratch training when samples are scarce. By amortizing one reusable backbone across the task space, PRISM offers scalable compression quality prediction for large-scale HPC workflows.


Back to PDSW’26: The 11th International Parallel Data Systems Workshop Archive Listing

Back to Full Workshop Archive Listing