Using Neural Networks for Two Dimensional Scientific Data Compression

Continual advances in high-performance computing have enabled the development of higher resolution and more realistic simulations of a wide variety of scientific phenomena. As a result, many computational science communities are increasingly constrained by the massive volumes of data produced, for example, strict storage constraints often force reductions in the number of output variables, data output frequency, or simulation length. Accordingly, modelers across many scientific domains are beginning to adopt purpose-built scientific data compression techniques as an effective mitigation for these challenges. The origins of scientific data compression tools every so often lie in image and video compression. Recently, compression researchers have achieved state-of-the-art performance using neural networks for natural image compression, but this achievement has yet to be adapted to scientific data. This paper assesses the performance of an existing autoencoder neural network compression algorithm on two sets of two-dimensional floating-point scientific data. Compared to state-of-the-art scientific data compression algorithms SZ and ZFP, this out-of-the-box neural network achieves higher peak signal-to-noise ratios at low bit rates, and remains competitive in controlling maximum point-wise error. This preliminary assessment paves the way for future research into neural network compression on floating-point scientific data.