← back to site

QGP freeze-out surface & FNO surrogate

Interactive 3D view of the quark-gluon plasma particlization hypersurface in Pb–Pb collisions at √sNN = 5.36 TeV, and how a Fourier Neural Operator approximates it on a coarse 61×61 grid.

hot QGP core (ε >> ecrit) freeze-out surface (T ≈ 0.155 GeV) cool hadronic edge FNO 61×61 bin (averaged)
30°
on
40°
on
5.0 fm
4
on
freeze-out cells (Cornelius)
~65,000 irregular 3D pts
FNO grid resolution
61 × 61 = 3,721 bins
FNO T accuracy
1.5% of Tfo
ε₂ correlation
0.993 FNO vs vHLLE
GPU speedup
~200× vs vHLLE CPU
training time
17 min Tesla T4, 1999 events
Physics: ∂μTμν = 0  ·  τπ D πμν/dτ + πμν = 2η σμν  (η/s = 0.08)
FNO operator: 𝒢 : ε(x,y,τ₀) → Tμν(x,y,τf)  ·  each layer: f(l+1)(x) = σ( ℱ-1[R(l)·ℱ[f(l)]](x) + W(l)f(l)(x) )
Code: github.com/phymanshu/FNO_Hydro  ·  vHLLE (Karpenko et al., CPC 185, 3016, 2014)  ·  Pb-Pb √sNN = 5.36 TeV