Separable Conditional Neural Fields for In-Situ Compression of High-Fidelity Spatiotemporal Turbulence Simulation Data

High-fidelity eddy-resolving turbulence simulations, such as direct numerical simulation and large-eddy simulation, produce immense spatiotemporal datasets due to the need for fine spatial resolution and frequent temporal sampling to resolve multiscale turbulent structures. As solver capabilities advance, enabling larger and more detailed simulations, the volume of data generated has become a primary bottleneck, overwhelming storage systems, limiting I/O bandwidth, and constraining downstream analysis. In-situ compression, performed during simulation runtime, offers a practical path forward but remains technically challenging due to the complex, anisotropic, and nonlinear nature of turbulence.

This work introduces separable conditional neural fields (SepCNF) designed for in-situ spatiotemporal encoding of turbulence data. SepCNF integrates tensor decomposition with conditional neural field representations to construct compact, continuous latent embeddings of flow fields in both space and time. Its separable architecture exploits low-rank structure while conditioning on physical context to preserve key flow dynamics under aggressive compression. Numerical experiments demonstrate that SepCNF achieves superior compression ratios and high reconstruction fidelity, preserving accurate coherent structures and spectral content. This framework provides a scalable and accurate solution for in-situ spatiotemporal data compression.