Multi-Pixel#

CHAI-class receivers have multiple pixels per array (e.g., 7 HFA pixels) and multiple arrays (HFA, LFA, 4G1–4G4). The engine handles this in two ways.

Vectorized Path (Primary)#

resolve_calibration_load_full() computes gamma, T_rec, and bad channels for all pixels simultaneously as [C, R, A] arrays. The calibration loop (calibrate_full) broadcasts these across the dump and subscan axes — no per-pixel loop needed.

This is the path used by cal-cli and cal-io::pipeline.

Per-Pixel Path (Python/Dask)#

cal-py uses a per-pixel strategy for Dask compatibility:

        graph TD
  P0[Pixel 0,0] -->|calibrate_scan| R0[Result 0,0<br/>+ fitted PWV]
  R0 -->|extract PWV| PWV[Common PWV]
  PWV --> P1[Pixel 0,1]
  PWV --> P2[Pixel 1,0]
  PWV --> PN[Pixel N,M]
  P1 -->|calibrate_scan| R1[Result 0,1]
  P2 --> R2[Result 1,0]
  PN --> RN[Result N,M]
  R0 --> MR[merge_pixel_results]
  R1 --> MR
  R2 --> MR
  RN --> MR
  MR --> CR[CalibrationResult<br/>full R,A]

  1. slice_pixel(0, 0) extracts pixel 0 as a single-pixel ScanData

  2. Calibrate pixel 0 → extract fitted PWV

  3. For all other pixels: calibrate with PwvSource::External(pwv)

  4. merge_pixel_results() stitches single-pixel results into full [C, D, R, A, S]

Common PWV Strategy#

By default (pwv_per_pixel = false), PWV is fitted on pixel 0 only and propagated to all other pixels. This ensures consistent atmospheric correction across the focal plane and reduces computation by \(\frac{1}{N_{pixels}}\).

Set pwv_per_pixel = true to fit each pixel independently (useful for diagnosing receiver-dependent systematics).

Implementation:

  • cal-core/src/scan.rs → ScanData::slice_pixel(), merge_pixel_results()

  • cal-io/src/pipeline.rs → common_pwv_source()

  • cal-py/src/lib.rs → calibrate_one_scan()