Supplementary plots

<p>To comprehensively evaluate our proposed methodology for lightweight and effective compression ratio estimation, we present experimental results on multiple datasets. We consider three different lossy compressors: SZ3, ZFP, and SPERR, each with error bounds of 1e-3 and 1e-6.</p> <p>For the Hurricane, Miranda, and NYX datasets, we perform a 12-fold cross-validation using our proposed method. The median approximate percentage error (MedAPE) from each validation fold is used to create a distribution of MedAPE values, as depicted in the respective plots. These experiments exemplify the &quot;in-sample prediction accuracy&quot; mentioned in the main paper. The color-coded crosses in the plots represent the maximum MedAPE&nbsp;observed across all the fields for each of the compressors.</p> <p>Regarding the CESM datasets, we focus on the first 10 time snapshots, each containing multiple 2D and 3D fields. We specifically consider the following 3D fields with dimensions of 3600X1800X26 due to their spatial diversity: CLOUD, CONCLD, FICE, GCLDLWP, ICLDIWP, ICLDTWP, OMEGA, OMEGAT, RELHUM, T, U, UU, V, VT, VU, VV, Z3. Similar to the in-sample prediction experiments, we employ a 12-fold cross-validation and illustrate the distribution of MedAPE values obtained from each validation fold.</p> <p>Lastly, for the Qmcpack dataset, which consists of 280 time-points, each containing a 3D dataset with dimensions of 69X69X115, we assess the out-of-sample prediction accuracy. We train our model using datasets from time points 1:20, 40:60, 80:100, ..., 240:260, and then use the remaining fields as the test data. The MedAPE (%)&nbsp; is plotted on the y-axis, while the x-axis represents the corresponding time points as the test dataset.</p>