exoALMA. II. Data Calibration and Imaging Pipeline

This paper, “exoALMA. II. Data Calibration and Imaging Pipeline” (Loomis et al. 2025), addresses a practical but scientifically central question: how can ALMA interferometric data be calibrated and imaged with sufficiently high fidelity that faint, spatially localized deviations from Keplerian rotation in protoplanetary disks are not artifacts of miscalibration, misalignment, or imaging choices? The exoALMA Large Program is designed to search for subtle kinematic signatures—potentially produced by embedded planets—using high angular and spectral resolution Band 7 observations of three molecular lines (12CO J=3–2, 13CO J=3–2, and CS J=7–6). Because the analysis is image-plane and channel-map based, the “noise floor” for kinematic features is set not only by thermal noise but also by calibration and imaging systematics that can mimic real velocity residuals. Thus, the pipeline must preserve image fidelity across multiple execution blocks (EBs), multiple array configurations (12-m compact/extended and, for some targets, ACA 7-m), and multiple spectral windows.

The study is not a statistical experiment with a conventional sample size; instead, it is a methodological paper describing the end-to-end processing of a large observational dataset. The exoALMA program targeted 15 protoplanetary disks, observed between October 2021 and May 2023, with 95 12-m execution blocks and 27 7-m execution blocks. The observations used four spectral windows in Band 7: three centered on the molecular lines with 15.3 kHz sampling (native velocity spacing 13.5 m s1 and effective velocity resolution of 26 m s1) and 3840 channels per spw, and a fourth wideband spw (1.875 GHz bandwidth) for deep continuum and self-calibration support. The program’s sensitivity goal was 3 K in a 0.1 beam over a 150 m s1 channel at the representative 12CO frequency. For targets with large angular sizes (disk diameters in 12CO above 6), ACA data were included for 7 sources to recover large-scale emission.

Methodologically, the authors build on the ALMA pipeline calibration (CASA-based, using CASA v6.2.1.7) but explicitly add custom steps for (i) alignment between EBs, (ii) self-calibration to correct phase decoherence, (iii) flux rescaling when needed, and (iv) careful imaging with masking strategies designed to avoid biasing results toward Keplerian expectations. The pipeline begins with ALMA pipeline-calibrated measurement sets, followed by manual quality checks. They report that most EBs were fine, but some were semi-pass flagged for specific issues: in one case (LkCa 15) the phase calibrator for long-baseline observations was spatially resolved, so those data were excluded; in another (J1615-3255) maser frequency lock issues produced time-variable LO frequency instability with a standard deviation of 10 kHz. For the latter, they used per-scan centroiding of a match-filtered impulse response spectrum of 12CO to re-center frequencies, finding no structural imaging artifacts.

A key part of the methodology is diagnosing three classes of inter-EB problems: relative alignment errors, phase decoherence (decorrelation), and flux differences. The authors use continuum image comparisons and visibility-domain diagnostics (deprojected baseline amplitude trends and amplitude-vs-time “waterfalls”) to identify decoherence. They then implement a staged self-calibration workflow. First, they create pseudo-continuum measurement sets by flagging 15 km s1 around the systemic velocities for each line and averaging the remaining data into 250 MHz channels (chosen to avoid frequency smearing at Band 7). They also note that in a few disks, complex organic molecule emission may not have been flagged, but testing showed no notable impact on the self-calibration outcome.

The self-calibration proceeds in multiple iterations and at multiple data-combination levels. For each individual EB, they perform one round of phase-only self-calibration using tclean with Briggs weighting (robust parameter 0.5), stopping CLEAN at a 66 RMS threshold to avoid inserting spurious signal. They report that this step increased peak SNR in the resulting EB images by up to 50% relative to the original images. Next, they align EBs in the uv-plane rather than relying on image-plane Gaussian fits, because disk morphologies can be asymmetric and ring/gap dominated. Their uv-alignment grids visibilities onto a common uv-grid with natural weighting and minimizes the difference only over overlapping uv-cells, solving for positional shifts (phase shifts) while separating flux scaling into a later step. They emphasize that the order of operations matters when self-calibration is involved: misalignment can create artificial asymmetries that could be misinterpreted as non-Keplerian kinematic features. Empirically, they find the most robust order is EB-based self-calibration, then alignment, then iterative group self-calibration from shorter to longer baselines.

After spatial alignment, they perform flux alignment. They diagnose flux offsets as either constant multiplicative scaling errors (absolute flux calibration differences) or baseline-dependent offsets indicative of residual phase decoherence. They apply a single flux correction when offsets exceed 4% and there is no evidence of decoherence; otherwise, they use a modified group self-calibration approach. In cases with decoherence, they first self-calibrate ACA+SB to reduce baseline-dependent decorrelation, then assess flux offsets and rescale any EB with offsets exceeding 4%, repeating the phase self-calibration after rescaling. They report that after this second iteration, flux offsets are verified to be all <4%.

For group-level self-calibration, the pipeline progressively adds longer baselines, similar to DSHARP’s strategy. For sources with ACA, they start with two rounds of ACA phase self-calibration: the first on EB-long intervals combining scans but not spws/polarizations, and the second on 30 s intervals after combining spws. They then concatenate ACA and SB and run multiple phase self-calibration rounds with progressively shorter solution intervals (360 s, 120 s, 60 s, 30 s, 18 s), again CLEANing down to 66 RMS each time. They report that peak SNR improvements can be dramatic: improvements of >300% in most cases during the ACA+SB phase self-calibration. They also report that decoherence improvements were observed down to 18 s solution intervals, and that amplitude-vs-time “waterfalls” were significantly reduced.

If flux offsets remain significant even after phase self-calibration (an example is shown for PDS 66), they revert to un-self-calibrated datasets, apply uniform flux rescaling for the offset EBs, and re-run the phase self-calibration. Finally, they add long-baseline (LB) data and perform phase self-calibration (with multiple rounds) followed by two rounds of amplitude+phase self-calibration. For amplitude solutions, they use a conservative SNR threshold of 5 and flag outliers with amplitude solutions <0.8 or >1.2. The CLEAN model for amplitude rounds is generated deeper (down to 16) to include low-level flux on relevant angular scales. They again report substantial peak SNR and improved noise structure in several disks.

Once the final self-calibration is complete, they produce “fiducial” measurement sets: a time-averaged continuum set (30 s binning) and line measurement sets for each spectral window (30 s binning), including continuum-subtracted versions. They apply continuum subtraction using uvcontsub with solint=1 and fitorder=1. The authors stress that the order of applying gain solutions, phase shifts, and flux alignments is non-commutative and must match the pseudo-continuum pipeline order.

The imaging section addresses two additional fidelity threats: non-Gaussian PSFs from combined array configurations and spectral artifacts that can map into spurious spatial distortions. They discuss three restoration strategies from the literature (re-weighting visibilities, JvM residual scaling, or using a larger restoring beam). For exoALMA, they avoid the JvM correction to prevent confusion in statistical significance when point-like features coexist with extended emission. They also avoid beam enlargement because it would blur small-scale kinematic features. Instead, they use Briggs robust weighting and moderate uv-tapering to produce PSFs “Gaussian enough” for kinematic analysis, and they record a PSF conditioning metric  (defined as restoring beam power divided by main-lobe dirty beam power) in FITS headers. They caution that for some ACA-including images,  indicates beams that remain significantly non-Gaussian, so flux measurements in low-SNR regions should be treated carefully.

For line imaging, the pipeline uses fiducial measurement sets and produces channel maps with specified channel spacing (typically 100 m s1 for CO and 200 m s1 for CS) and target beam sizes (often 0.150.15 for fiducial cubes). A central methodological choice is the CLEAN masking strategy: because the science targets non-Keplerian morphologies, they do not use Keplerian masking. Instead, they generate CLEAN masks based on the observed emission morphology in each channel by convolving the previous-step CLEAN model with a relatively wide Gaussian kernel (0.5–0.7) and thresholding to create a binary mask. They then perform deep iterative CLEANing within the mask, producing multiple products at different CLEAN thresholds (e.g., 6, 5, 4, 36 RMSfin).

The paper also evaluates spectral effects: ensuring adequate sampling relative to the narrow linewidth along a single line of sight, accounting for ALMA’s Hann-window spectral response and channel covariance, and considering spectral regridding from the topocentric frame into the source co-moving frame. They conclude that their spectral sampling is high enough to mitigate these effects for the spatio-kinematic scales relevant to embedded planets.

Although the paper’s primary deliverable is a pipeline rather than a set of astrophysical results, it provides concrete performance indicators for the produced images. For continuum images, Table 2 reports achieved RMS values and peak SNRs; for example, DM Tau has achieved RMS 26.6 cJy beam11 with peak SNR 128.6, while PDS 66 reaches peak SNR 1232.5 with RMS 26.0 cJy beam11. For fiducial line cubes (Table 3), the achieved RMS and peak SNR vary by source and molecule; for instance, MWC 758 CO 3–2 has RMS 4.39 mJy beam11 and peak SNR 37.73, while HD 135344B CO 3–2 has RMS 2.98 mJy beam11 and peak SNR 67.11. High-resolution cubes (Table 4) show that some sources have reduced SNR due to resolution demands; e.g., LkCa 15 CO 3–2 at high resolution has RMS 2.53 mJy beam11 and peak SNR 14.48.

Limitations are mainly methodological and acknowledged implicitly through conservative choices. The pipeline avoids aggressive amplitude self-calibration early and uses calonly modes to prevent flagging low-SNR data. It also notes that PSF non-Gaussianity can persist for some ACA-including images, requiring caution when measuring fluxes at high resolution and low SNR. Additionally, while the paper tests the impact of unflagged COM emission on self-calibration, it does not claim that all astrophysical complexity is fully captured; rather, it demonstrates that the calibration outcome is robust to at least that specific complication.

Practically, the results matter to anyone using exoALMA data for kinematic analyses of planet-disk interactions. The pipeline’s main contribution is a set of calibrated and aligned measurement sets and fiducial images designed to minimize spurious non-Keplerian artifacts. Researchers studying embedded planets, disk dynamics, or instabilities should care because the pipeline explicitly targets the dominant systematic risks for channel-map residual studies: phase decoherence, EB misalignment, flux scaling errors, and CLEAN/restoration artifacts. The authors also release the calibration scripts and QA figures via the exoALMA data release, enabling reproducibility and future improvements. In broader terms, the paper provides a template for high-fidelity interferometric processing when the scientific inference depends on subtle spatially localized deviations rather than on total flux or point-source detection.