Quantifying the Impact of Lossy Compression on Neural Generative Surrogate Modeling

The study presents a method for quantifying the impact of lossy compression on generative surrogate models used in scientific simulations. By analyzing the uncertainty in neural network training, the authors demonstrate that their approach allows for data storage reductions of up to 39x and training time reductions of up to 3x, with minimal effects on model accuracy. This work is significant for practitioners as it addresses storage and I/O challenges in high-fidelity models, enabling more efficient training processes without compromising quality.