melloyello/graphtrans

Cylinder wake at Re=100 — rotary-control training dataset

Training data for a mesh-transformer dynamics surrogate of the 2D cylinder wake (HydroGym / Firedrake, Re=100, M=8,744 mesh vertices), actuated by cylinder rotation. Built for a GNN-MPC pipeline: the surrogate learns the one-step map (x_k, Omega_k) -> x_{k+1} at the control period dt = 0.1 (10 solver substeps of dt=0.01, matched sampling).

Contents

• dataset/traj_XXX.npz — 220 trajectories, 300 control steps each (66,000 transitions). Per file: v [301, 8744, 2] float32 CG1 velocity snapshots (vertex order matches mesh/), cmd_omega [300], meas_omega [301] (wall-measured rotation — integrity check vs the actuator-lag model holds to 5.8e-5 across all files), CL, CD [301], meta (JSON job spec incl. RNG seed — every trajectory exactly regenerable).
• dataset/manifest.json — job specs and the frozen train/val/test split (192/14/14, split by trajectory). Test = special trajectories (unforced replays + sinusoids), quarantined for evaluation.
• dataset/norm_stats.npz — normalization statistics, train split only.
• mesh/cylinder_re100_dataset.npz — vertex positions [8744,2], triangles [17258,3], node types (0 interior, 1 cylinder, 2 inlet, 3 outlet, 4 freestream), plus a 201-frame uncontrolled reference movie.
• constants/actuation_constants.json — measured plant constants: R=0.5, U_inf=1, actuator lag tau=0.0556, sign conventions, base-flow growth rate sigma=0.1506, St_linear=0.139, St_nonlinear=0.172, Omega_max=2.0.
• checkpoints/ — Firedrake restart files: settled limit cycle, unstable steady base flow qB (Newton+continuation), and the six growth-run amplitude rungs used as corridor initial conditions.
• scripts/ — full generation + training code (HydroGym 1.0 / Firedrake).

Trajectory composition

kind	count	initial condition	forcing
cycle_ou	100	limit cycle, random phase	clipped OU walk (a=0.93), tiered sigma
growth_ou	90	growth-run amplitude rungs	clipped OU walk
qB_nudge_ou	16	base flow + fresh random nudge	gentle OU walk
unforced / sine	14	cycle & corridor	none / sinusoidal Omega

Excitation follows Eberhard et al. (arXiv:2602.17601): random-walk inputs at the MPC sampling period; volume calibrated to Garnier et al. (arXiv:2508.18051).

Regenerability

scripts/generate_dataset.py --id N with dataset/manifest.json reproduces any trajectory bit-for-bit (deterministic solver, stored seeds) given a HydroGym 1.0 + Firedrake environment and the checkpoints/ restart files.