PI: Franck Cappello (MCS)
Description: Our research is focused on understanding how deep learning can be used to improve lossy compression of scientific data from simulations and instruments. Deep Learning will be used in the prediction stage of the SZ compressor developed at Argonne. We plan to investigate several aspects:
- Is deep learning good enough for accurate prediction of is a combination with linear predictor necessary to reach high prediction accuracy?
- What type of DNN provide the best prediction performance?
- What are the hyper-parameters for the DNNs that provide the best prediction for compression?
We will adapt the SZ compressor to include the deep learning prediction stage. We will apply this research to the ECP HACC, NYX, QMCPACK, NWCHEM, LATICEQCD, EXAALT and EXAFEL data. We may also explore the CESM ATM dataset and other non ECP application datasets. We plan to publish papers presenting the results of this research.
Testbed: Nvidia DGX, Voltas