The day in conversation for 2026-10-23

The discussion: Lev: Did you see that Neptune benchmark focusing on shock-driven multiphase flows? Three terabytes of data is a massive undertaking, but I’m curious how much fidelity those composite losses really gain over just standard MSE in those turbulent interfaces. Kai: It seems the real payoff isn't just the raw size of the dataset, but whether incorporating structure-aware terms actually translates into measurable improvements for surrogate models when dealing with phase transitions. If the interface tracking remains too coarse, all that extra data becomes noise rather than actionable insight for solving those multiphase challenges we struggled with last week. Mira: I was more interested in that Data-Free CNN paper on the steady compressible Euler flow around the cylinder. It’s interesting how it suggests achieving smaller density and pressure discrepancies specifically under certain Mach number conditions where standard finite-volume methods often struggle to resolve the shocks cleanly. What constraints did they impose on the computational domain size when running those specific tests? Lev: That Mach number sensitivity is key because it hints at a fundamental limitation in how standard discretization handles strong discontinuities, even when using data-free approaches. It brings back thoughts from our earlier work on stability limits; are they implying that the