BDXXN/amortized-inference-training

Amortized Inference Training

Training trajectories for demo-conditioned material inference (Design 4: g_ψ(demo) → material code m, with m conditioning once-trained constitutive nets V(state; m) / D(state; m)). All dynamics come from the PAC-NeRF ecosystem MPM engine (GIC taichi MLS-MPM, dt = 1/4800 s); geometry is the PAC-NeRF torus (L3 will add Objaverse shapes). No self-made simulators.

The three levels

Train one encoder per level and compare — the deltas isolate where difficulty lives (IC-robustness vs material reading vs shape transfer).

Level	Folder	Contents	Purpose
L1	L1/	100 trajectories — same object (torus), same material (E=1e6, ν=0.3), varied ICs, 24 fps (59 frames)	observation view (Stage-B input rate); pipeline overfit pilot
L1 dense	L1dense/	100 trajectories — same object+material as L1, varied ICs, 48 fps (118 frames)	physics view: denser states for material supervision
L2	L2/	150-target (148 delivered; 2 soft tori splat unrecoverably) — same object, many materials (E, ν sweep), varied ICs, 24 fps	material diversity: makes latent codes m meaningful
L3	(pending)	Objaverse shapes × material sweep	shape diversity

The L1/L1dense material is certified against the shipped PAC-NeRF torus demo (corresponded particle error 4.6% of span at E=1e6, ν=0.3; the E=1e5 alternative rejected at 34%). The real 59-frame PAC-NeRF demo is reserved for evaluation and is not part of this dataset.

File schema

{MODE}_shard{S}_traj{NNNN}.npz (raw trajectory):

field	shape	meaning
traj	(F, 30000, 3) float16	particle positions; fixed subsample of the 381,414-particle PAC-NeRF torus cloud (exact correspondence)
sub_idx	(30000,)	indices into the 381,414-particle cloud
E, nu, rho	scalars	ground-truth material (EVALUATION ONLY — never train on them)
vel, omega, rot	(3,), (3,), (3,)	initial condition
fps	scalar	export rate

{MODE}_shard{S}_traj{NNNN}_prep.npz (chart preprocessing):

field	shape	meaning
t, R	(F,3), (F,3,3)	per-frame Kabsch rigid pose (world translation; body→world rotation)
xi	(F, 12)	chart coordinates — transformer-chart inversion (warm-started Adam; mean residual ~1.0–1.4% of span)
inv_err	(F,)	per-frame inversion residual, fraction of cloud span
E, nu, vel, omega, fps	—	copied from the raw file

Units (important)

Positions are meters. Material parameters in frame units scale with fps²; the canonical yardstick is demo units (time unit = 1/24 s), where the L1 material is μ=0.325, λ=0.291. When consuming 48 fps data, scale rates by fps/24 and use dt_sub = (24/fps)/NSUB so learned parameters are fps-invariant.

Loading

import numpy as np
d = np.load("L1dense/L1dense_shard0_traj0000.npz")
traj = d["traj"].astype(np.float32)          # (F, 30000, 3) exact MPM clouds
p = np.load("L1dense/L1dense_shard0_traj0000_prep.npz")
t, R, xi = p["t"], p["R"], p["xi"]           # floating-frame + chart states

Provenance / integrity gates

• Engine: GIC taichi MLS-MPM (the PAC-NeRF ecosystem), elasticity material, dt = 1/4800 s invariant across export rates.
• Per-trajectory CFL check by the engine; crashing ICs are excluded (L2: 2/150).
• Ground-truth material for L1 certified against the shipped PAC-NeRF torus trajectory (gate above).
• Observation augmentations (noise, frame subsets) are intentionally NOT stored — derive on the fly.