[bbarad]

Problem

On a high-quality 9.98 Å/px OMM dataset, measure_thickness returned NaN for 47–84% of triangles even after refinement. At ~1 nm/px the two leaflets merge into one flat-topped peak. The whole-surface average is unresolved too, so _global_bilayer_prior returned None, and the recovery tier for merged triangles never ran. Every failure was "single peak, no prior". The tomograms are not low-passed (power extends to Nyquist).

Changes

• Opt-in thickness_measurements.force_bilayer_prior (default false; CLI --force_bilayer_prior/--no_force_bilayer_prior). Fits an unresolved surface average as a bilayer anyway, accepted only if it passes the recovery checks (R², separation clear of the floor, comparable leaflet amplitudes), and uses it as the prior.
• New per-triangle forced_bilayer_prior column (1/0): 1 where the thickness relies on the forced prior. thickness stays a single column; together with bilayer_resolution, the flag identifies low-confidence regions. The column is always written so the schema is stable.
• Hint when the average is unresolved: measure_thickness now prints a message suggesting a less-binned tomogram, or the new option.
• density_sampling.interpolation: cubic (new default) or linear, used by sample_density and refine_mesh. Linear interpolation's between-voxel blur is enough to merge leaflets at ~1 nm/px. Spline coefficients are computed only for the block the surface spans, and out-of-volume samples stay NaN.
• radius_neighbors replaces the five cKDTree.query(k=500) neighbor queries in measure_thickness and refine_mesh. On dense meshes those silently shrank the averaging radius (~8 nm instead of 12 nm; ~16 nm instead of 20 nm on the example OMM). The surface is now thinned uniformly with a fixed seed so neighborhoods span the full radius at the same cost.
• fit_triangle_chunk now returns (thickness, offset, resolution, recovered).

Results

Four 9.98 Å/px OMM surfaces, 1,500 sampled triangles each, 12 nm radius:

Surface	Before	New default	+ force_bilayer_prior (flagged)
lam12	22%	36%	85% (49%)
lam25	15%	25%	81% (56%)
lam26_1	49%	58%	86% (28%)
lam10	41%	54%	87% (34%)

• Median thickness with the new default: 0.05–0.27 nm lower than before. Don't pool measurements made with and without this change; interpolation: linear reproduces earlier sampling.
• Bias caveat for forced values: these read thinner than strict fits. In a synthetic blur test, though, the strict survivors read thick (they are selected for wide separation) and the forced values matched the unblurred truth. The docs therefore say "lower-confidence" without claiming a direction.
• Well-resolved example surface: thickness is unchanged (median 2.92 → 2.90 nm), and the share of strictly resolved triangles rises from 54% to 67%. Cubic sampling had no meaningful time or memory cost (8.9 s vs 8.0 s; 2.1 GB vs 2.4 GB peak).

Refinement check

A 4-iteration refine_mesh on the example OMM was run with main, with this branch using linear sampling, and with this branch using cubic sampling. All three produce nearly identical surfaces:

• vertices differ by a median of 0.02–0.03 nm (95th percentile ≤0.15 nm);
• convergence curves overlap;
• measured fraction, centering, bilayer dip depth, and roughness match, with cubic marginally best.

The neighbor cap was hit for 92% of triangles at 20 nm. It barely matters for centering here, likely because offsets are smoothed again over 25 nm with an uncapped query.

Tests

119 pass:

• new test_forced_bilayer_prior.py, test_cubic_sampling.py, test_radius_neighbors.py;
• updated test_thickness_worker.py for the 4-tuple.

sample_density + measure_thickness were also run end to end on the example tomogram, both with and without the forced prior (the latter on deliberately blurred profiles).

Release

The CHANGELOG entries are folded into the 2.0.0b6 section (version stays 2.0.0b6), so the beta can be re-cut from main after merge.

🤖 Generated with Claude Code