1 of 1

�SuperBench: A Super-Resolution Benchmark Dataset

for Scientific Machine Learning�

Scientific Achievement

Due to the absence of standardized benchmark datasets for comparing and validating super-resolution (SR) methods in SciML, we introduce a new high-resolution benchmark dataset, SuperBench, for super-resolution of scientific problems. SuperBench empowers researchers to advance super-resolution methods specifically tailored for scientific tasks.

Significance and Impact

SuperBench includes four distinct datasets of HR simulations, surpassing the resolution of typical scientific datasets used in SciML. It features challenging problems in fluid flows, cosmology, and climate science. It aims to push the performance limits of existing methods and facilitate the development of innovative SR methods for scientific applications.

Technical Approach

We benchmark existing super-resolution methods on SuperBench, such as state-of-the-art Swin Transformer.
We investigate a range of degradations tailored for scientific data, including bicubic/uniform downsampling and using low-resolution simulations.
We analyze the performance of various super-resolution approaches using both data and physical-centric metrics.

PI(s)/Facility Lead(s): Lenny Oliker (LBL)

Collaborating Institutions: ICSI, UC Berkeley, University of Tennessee, Knoxville

ASCR Program: SciDAC RAPIDS2

ASCR PM: Kalyan Perumalla (SciDAC RAPIDS2)

A) Illustration of the super-resolution problem. B) Example snapshots of weather data. The task is to recover the high-resolution (HR) representation from the corresponding low-resolution (LR) input by a factor of × 16. All SOTA methods reconstruct a blurred approximation that washes out important multi-scale and fine-scale features. C) Showcasing baseline SR methods on cosmology data with LR simulation data as inputs. The results are based on ×8 up-sampling. D) Showcasing baseline SR methods on turbulent fluid flow data (Re = 16000) under bicubic down-sampling (×16 up-sampling).

LOCAL LAB POC: Ben Erichson

TALKING POINTS:

Background: Super-resolution techniques have emerged as powerful tools for enhancing data resolution and retrieving fine-scale structures. These techniques find application in diverse fields such as image restoration and enhancement, medical imaging, astronomical imaging, remote sensing, and forensics. Despite the remarkable achievements of deep learning-based super-resolution methods developed primarily in the computer vision community, their application to scientific tasks has certain limitations. In particular, these methods often struggle to capture intricate fine-scale features accurately and to preserve fundamental physical properties and constraints inherent in scientific data.

Datasets: there are four datasets of high-resolution simulations in fluid flows, cosmology, and climate science. We provide two fluid turbulence datasets with two different high Reynolds numbers. For the cosmology data, we provide a pair of low-resolution and high-resolution simulations. The weather data is modified from ERA5 dataset. The entire SuperBench datasets take up over 250 gigabytes.

Technical approach: We evaluate the existing super-resolution methods on SuperBench datasets, such as convolutional neural networks, ResNet-based architectures, and state-of-the-art Swin Transformers. For different data degradations, we consider the general computer vision setup with bicubic downsampling, mimicking the experimental setup by using uniform downsampling and potential noise, and directly using low-resolution simulation data as inputs. For the evaluation metrics, we consider pixel-wise differences, such as relative reconstruction accuracy. We also consider some metrics for measuring human-level perception, like SSIM. In addition, since we are focusing on scientific data, it is also important to measure the preservation of physical properties.

Significance and impact: SuperBench includes four distinct datasets of HR simulations, up to 2048 x 2048, surpassing the resolution of typical scientific datasets used in SciML. It features challenging problems in fluid flows, cosmology, and climate science. It aims to push the performance limits of existing methods and facilitate the development of innovative SR methods for scientific applications.

This paper is a collaborative and interdisciplinary work with UC Berkeley, International Computer Science Institute (ICSI), and University of Tennessee, Knoxville. The topics of AI for Science and SR for accelerating extreme-scale simulations align well with the strategy goals and research priorities of DOE and Berkeley Lab.

METADATA:

Name of the associated awarded project: SciDac-5 Institutes RAPIDS2

PI name(s): Leonid Oliker

Name of the program manager: Kalyan Perumalla (RAPIDS)

CITATIONS:

P. Ren, N. B. Erichson, S. Subramanian, O. San, Z. Lukic, and M. W. Mahoney, “Superbench: A super-resolution benchmark dataset for scientific machine learning.” arXiv preprint arXiv:2306.14070 (2023).

AWARDS:

If this work has received any awards at conference or otherwise, please note them here.