Campaign

Continuity-prior PINN for CSI crowd density — does a 1D mass-conservation prior regularize the CSI->density inverse?

c-continuity-prior-csi-density · exp-csi-crowd

Archive snapshot, as of 5 h ago — the run corpus is rebuilt once a day, so this page is not a live reading. The fleet panel is the live one; it refreshes every 30 s.

Sessions

Session	State	Runs	Synthesis	Criticism	Figures	Verdict
F9WXGYFC 2026-06-25T20:22	finished	5	✓	✓	—	Mismatched-FD ablation REFUTES the FD-specific reading of session-1's win. At rho>=1 (n=105), the matched Weidmann closure (1.23), a constant closure (1.24) and an anti-physical increasing closure (1.24) give IDENTICAL RMSE -> the continuity gain is FD-agnostic, not the specific physics. And plain smoothing (0.91) BEATS every PINN variant -> the continuity content is a suboptimal generic regularizer, not physics. The session-1 -11% was confounded exactly as the critic feared. Demote continuity-prior-regularizes-density: the 1D-continuity-with-FD-closure prior earns no specific support here; a generic smoothness prior is better. Regularization helps the under-determined CSI->density inverse, but the fluid physics is not what helps.
3YTXJXVJ 2026-06-25T20:01	finished	5	✓	✓	—	Necessary in-silico signal, confounded. With a crush testbed (peak rho 4.98, n=110 high-regime cell-frames) the continuity-PINN beats the unconstrained MLP in all regimes incl. rho>=1 (-11%) with no loss at low density. BUT the FD closure (Weidmann) matches JuPedSim's own velocity law, so the gain cannot be separated from leaking the generator's FD; no L2/smoothing baseline; single seed. Nudges continuity-prior-regularizes-density speculative->plausible at most — not confirmation. Needs: mismatched-FD ablation, a generic-regularizer baseline, multi-seed error bars.

Brief

Question

The thesis's central unifying claim continuity-prior-regularizes-density is that a 1D continuity-equation prior (with a fundamental-diagram velocity closure) regularizes a CSI-based crowd-density estimator better than a purely data-driven model, in the high-density regime (rho >= ~1 ped/m^2). It has never been tested. The coupled chain supplies exactly the controlled testbed the hypothesis note asks for: JuPedSim ground-truth density on a real floor + co-located synthetic CSI. This campaign runs that settling experiment.

Framing discipline. In-silico, JuPedSim density truth + self-authored synthetic CSI on a sparse-link layout (the regime where a physics prior should help most). Necessary-not-sufficient: a sim gain licenses the approach; real-channel transfer + synthetic-CSI faithfulness stay open (synthetic-csi-sim-to-real-transfer / IP-106). Update the hypothesis strength only on the defeater outcome.

Unblock (2026-06-25)

This campaign needs the python-ml runner (CPU torch 2.8.0 + lightgbm). It is CPU-only by design — no GPU required; the continuity PINN is a small MLP that trains in minutes on CPU. The macOS dual-libomp segfault is a native-only hazard (torch's libomp + lightgbm's colliding); running inside the python-ml Linux container (single libgomp1) avoids it entirely. The image builds from dockerfiles/runners/python-ml/ (docker build locally; CI via docker-runners.yml).

Testbed

test-lab-synth-floor-0 — small, with a corridor (1D continuity's natural geometry) and the csi-synth-corridor-link Tx/Rx layout (a deliberately sparse link set — the under-determined CSI->density inverse a prior is meant to regularize). A crowd-size sweep {12, 24, 36, 60} drives the density up; the crush (n=60, tight funnel) run is there to sustain rho >= 1 long enough to populate the high regime (a brief transient peak is not a testable population — the open issue from the first session).

What the supervisor does

Fan out exp-csi-crowd over the crowd-size sweep (no Doppler — keep the frame axis aligned to the trajectory; no BLE — density truth comes from the trajectory).
Run continuity_density_pinn (runner python-ml): build per-cell density truth + velocity from the trajectory, CSI link features per frame; train unconstrained vs PINN; report RMSE by regime x AP-sparsity + the defeater verdict.
Synthesise with the necessary-not-sufficient framing; update the hypothesis only on the defeater.

Figure render request

continuity_density_pinn — density RMSE by regime (unconstrained vs PINN bars), the RMSE-vs-AP-sparsity curve, and a predicted-vs-true density field strip at the densest frame.

Out of scope

Real-channel transfer / synthetic-CSI faithfulness — IP-106 / exp-csi-calibration.
2D velocity field / full Navier-Stokes — the claim is scoped to 1D density (hypothesis note).
lightgbm GBDT baseline — the unconstrained MLP (same architecture minus the physics term) is the controlled ablation; a separate model class would confound architecture with the prior.

Expected interpretation

PINN lower RMSE at rho >= rho_hi, no loss at rho < rho_lo -> confirms (in-silico) — upgrade continuity-prior-regularizes-density speculative -> plausible; the AP-sparsity curve shows where it matters most.
PINN loss at low density -> the defeater's low-regime clause fires; the fluid prior mis-helps the dominant indoor regime. Demote.
High regime never populated (peak rho < rho_hi or < 5 cell-frames) -> a testbed finding: residential floors don't sustain crush density; the claim needs a bottleneck/corridor inflow geometry (EXP-S1 territory) before it can be settled.

Sessions

01KW05Y7MGBVXXZN8N3YTXJXVJ — crush testbed; first pass looked like a win (continuity-PINN −11 % at ρ≥1), but the critic flagged the FD-leak confound.
01KW074MN1Q3M0JM2VF9WXGYFC — mismatched-FD ablation; FD-agnostic + smoothing beats every continuity variant → refutes the FD-specific reading on the low-directed-flow corridor. Demote-recommend; door left open only for strong directed flow.
01KW1WSRM93FHQM00GS1XPW00V (2026-06-26, local-manual) — strong-directed-flow re-test. Wired directed_flow → sionna-csi-runner (parametric scene), single-exit bottleneck density sweep, REAL coupled CSI. Result REPRODUCES session 2 in the last-open regime: FD-agnostic, smoothing beats all continuity variants (0.61 vs ~0.73–0.76 at ρ≥1, robust over seeds + λ), and the new measured-v divergence variant also fails to beat smoothing. Field-level flux signal does not transfer to the CSI inverse. Defeater met → continuity-prior-regularizes-density candidate → deferred.

Campaign outcome: the 1D continuity (FD-closure and measured-v) prior is, across both flow regimes, a generic smoother in disguise — beaten by plain smoothing on the under-determined CSI→density inverse. The "crowd-as-liquid" continuity prior is not warranted as the thesis central claim. Hypothesis deferred.