CHIME: Canadian Hydrogen Intensity Mapping Exp.

• CHIME is a multipurpose radio interferometer that uses a fixed, redundant cylindrical array to map 21 cm cosmological signals while tracking fast radio bursts and pulsars.
• It employs an FX digital processing architecture with FPGA-based channelization and GPU nodes for real-time spatial correlation, enabling high-resolution surveys over 400–800 MHz.
• Its advanced calibration, beam characterization, and data provenance techniques ensure sub-percent accuracy for cosmology studies and transient radio science.

The Canadian Hydrogen Intensity Mapping Experiment (CHIME) is a multipurpose, stationary radio interferometer sited at the Dominion Radio Astrophysical Observatory (DRAO) in British Columbia, Canada. Designed to map large-scale cosmological structure via neutral hydrogen (H I) 21 cm intensity mapping over the redshift range $0.8 \leq z \leq 2.5$, CHIME simultaneously serves as a high-cadence observatory for fast radio bursts (FRBs), pulsar timing, and precision radio transient science. Key to its conceptual and technical advancement are a highly redundant cylinder-array design, real-time digital back-end, and rigorous systematics control across a broad 400–800 MHz frequency window. The following sections synthesize CHIME's architecture, calibration, data infrastructure, principal scientific results, and operational strategies, with particular emphasis on details relevant to high-precision cosmology, software-defined beamforming, FRB and pulsar pipelines, and data provenance.

1. Instrument Design and Digital Architecture

CHIME consists of four parallel, fixed parabolic cylindrical reflectors, each 100 m (north–south) × 20 m (east–west), oriented along the local meridian. Spaced by 2 m, the cylinders focus sky emission onto a total of 1024 dual-polarization “cloverleaf” feeds, with 256 per cylinder, each spaced at 0.3048 m in the north–south direction. This configuration yields a full feed grid of 2048 signal paths and provides instantaneous Nyquist sampling in the focal plane over much of the 400–800 MHz band. Analog signals are first amplified by room-temperature LNAs (20–25 K system noise), transmitted via 50–60 m coaxial runs, and then digitized at 800 MS/s with 8-bit ADCs operating in the 2nd Nyquist zone (Collaboration et al., 2022).

Digital processing follows an FX architecture. The F-engine, realized by 128 ICE FPGA boards, applies a polyphase filter bank and 1024-point FFT to channelize the 400 MHz bandwidth into 1024 frequency bins (Δf ≈ 390 kHz), with each FPGA processing 16 input signals (Bandura et al., 2016). Data are requantized to 4+4 bits and reorganized by a hierarchical “corner-turn” network (full-mesh FPGA mesh: intra-/inter-crate, QSFP+ links), efficiently aggregating all antennas for a subset of frequencies per GPU node.

The X-engine comprises 256 compute nodes (each equipped with four AMD S9300x2 GPUs) and computes the full $N^2$ spatial correlation matrix and digital beamformed products in real time. The system is designed for a sustained throughput of ≈8.4×10¹⁴ cMAC/s and an aggregate data flow of 6.6 Tb/s from F- to X-engine (Denman et al., 2020). Backend products include 10 simultaneous, steered dual-polarization beams (pulsar tracking), FFT fan beams (FRB survey), full-visibility matrices for intensity mapping, and real-time baseband triggers for VLBI/FRB outriggers (Collaboration et al., 2020).

2. Calibration, Beam Characterization, and Systematics Control

The extreme foreground-to-signal dynamic range (≈10⁵) for cosmological H I mapping necessitates calibration and beam knowledge at the ≲0.1% level per feed (Reda et al., 2022, Newburgh et al., 2014). Complex gain calibration is derived daily from bright source transits (Cas A, Cyg A, Tau A, Vir A), with flux scales anchored to standards (Perley & Butler). Phase drifts and delay errors from clock offsets, focal-line thermal expansion, and cable length variations are modeled and corrected via in situ weather/thermal monitoring, enabling amplitude and phase stability at ≲1% and ≲0.003 rad, respectively, post-correction (Collaboration et al., 2022).

The CHIME beam is characterized through complementary methods:

• Radio Holography: Signals from each feed are correlated with a tracking reference (26 m John A. Galt telescope), allowing direct measurement of the Jones matrix response per feed, for both co- and cross-polarizations, with amplitude and phase mapping across –12°≤δ≤59° (Amiri et al., 2024, Reda et al., 2022). Beam FWHM is found to range from ≈2° at 400 MHz to ≈1° at 800 MHz; standing-wave ripples of ≈30 MHz (amplitude up to ±5%) are observed.
• Sun Drift Scans: The Sun's declination drift provides a two-dimensional map of the primary beam over ≈7200 deg² without telescope movement, enabling assessment of the main lobe and sidelobes with ≲10% uncertainty where B > 10⁻³ (Collaboration et al., 2022).
• Point-Source Meridian Tracks and Data-Driven SVD: Meridian raster scans and model-based decompositions constrain the beam envelope and provide cross-validation.
• Feed-to-Feed and Polarization Effects: Holography and solar scans show per-feed main-lobe centroid wandering at the ≈0.15° (1 cm) level along cylinders. Cross-polarization leakage is observed at ≈5–10% of co-polar amplitude in main lobes, with far sidelobe leakage up to ≈40% under some conditions (Amiri et al., 2024).

Performance of these strategies is assessed via in-band, drift-scan, and out-of-band comparisons; beam nonuniformities and polarization leakage are directly injected into cosmological simulations to determine impact on foreground subtraction and 21 cm power-spectrum bias (Newburgh et al., 2014, Berger et al., 2016).

3. Data Acquisition, Management, and Provenance

CHIME's correlator produces up to 100 TB/day of output, comprising raw visibilities, real-time beams, and system logs. Data are divided into acquisitions (∼3 hr segments), with each file registered by size and checksum in a MySQL database (Hincks et al., 2014). The hardware configuration is modeled as a time-indexed directed graph $G(t)$ whose nodes represent physical elements (antennas, cables, amplifiers) and edges encode their connections. Every alteration is logged as a discrete event (add/remove/modify connection), with full support for time-stamping, versioning, and calibration association. Data ingestion is managed via the alpenhorn daemon, enforcing ≥2 distributed replicas per file.

A combined event-driven hardware registry and data-indexing schema ensure full provenance: each data product is tied to the exact analog-digital signal path and configuration in force at capture time, supporting reproducibility and diagnostics at sub-component granularity. Software architecture employs Python+peewee ORM and NetworkX, with web-based and barcode-driven hardware logging interfaces (Hincks et al., 2014).

4. Survey Modes, Beamforming, and Science Backends

CHIME's digital backend supports commensal operation for all principal science cases, through software-selection of correlation products and dynamic beam steering:

• 21 cm Intensity Mapping (Cosmology): All $N_{\mathrm{feed}}^2$ visibilities are integrated, redundantly averaged, flagged, and used for 3D mapping of brightness temperature fluctuations and baryon acoustic oscillation (BAO) scale measurement (Collaboration et al., 2022, Collaboration et al., 24 Nov 2025).
• FRB Science: Real-time FFT beamforming along the north–south axis (256 feeds) and east–west (cylinders) delivers 1024 fan beams, each channelized to 16 384 spectral bins (24.4 kHz) at 0.983 ms cadence. The FRB pipeline achieves 2–42 FRB/sky/day sensitivity (surveyed above 1 Jy), utilizing a scalable four-stage detection pipeline (“bonsai” blocked tree dedispersion, clustering, classification, and action) (Collaboration et al., 2018).
• Pulsar Science: Up to 10 simultaneous, fully-polarized tied-array beams track pulsars in real time via beamforming on the 2048-voltage input grid. The pulsar backend performs coherent dedispersion, folding (1024 phase bins, 1024 frequency channels, full Stokes), and maintains a dynamic scheduling system for high-cadence observations (>700 sources daily, near-daily PTA monitoring) (1711.02104, Collaboration et al., 2020).
• H I Absorber Searches: Fine channelization (down to 3 kHz) and multi-year integrations enable blind and targeted discovery of narrow H I 21 cm absorbtion systems at $0.78 \leq z \leq 2.55$ (Collaboration et al., 12 Jun 2025).
• VLBI/Outriggers: For milliarcsecond localization of FRBs and cosmological analysis, remote Outrigger stations (KKO, HCRO, GBO) operate as a very-long-baseline interferometric array, duplicating CHIME’s feed architecture and correlator, achieving ≲50 mas FRB localization with robust ionospheric calibration (Collaboration et al., 7 Apr 2025).

5. Cosmological Results and Sensitivity Benchmarks

CHIME has produced decisive (>11σ) detections of the cosmic 21 cm signal via both cross-correlation and auto-correlation, the former by stacking filtered maps on the positions of LRGs, ELGs, and QSOs from optical surveys (Bayes factor ln B ≳ 10.8–56.3), and the latter by measuring the 21 cm auto-power spectrum at $z ≈ 1.08–1.24$ (12.5σ, $0.4 < k < 1.5\, h\,{\rm Mpc}^{-1}$) (Collaboration et al., 2022, Collaboration et al., 24 Nov 2025). These analyses deploy advanced RFI masking (radiometer and fringe-rate tests), achromatic beamforming (frequency-independent synthesized beam), and delay-domain foreground filters (DAYENU) prior to time-averaging. Residual calibration errors are monitored and corrected via a hybrid foreground-residual subtraction (HyFoReS) pipeline.

The principal astrophysical parameter constrained is the effective H I clustering amplitude:

$$\mathcal{A}_\mathrm{HI} \equiv 10^3\, \Omega_\mathrm{HI}\, [b_\mathrm{HI} + \langle f\mu^2 \rangle]$$

where $\Omega_\mathrm{HI}$ is the cosmic neutral hydrogen fraction, $b_\mathrm{HI}$ its bias, and $\langle f\mu^2 \rangle$ the mean Kaiser RSD factor. Representative results: $\mathcal{A}_\mathrm{HI} = 1.68^{+1.10}_{-0.67}$ (QSO, $z=1.20$), with modeling uncertainties inflated by foreground and small-scale bias marginalization (Collaboration et al., 2022).

FRB surveys achieve >200 FRBs/year with reliable flux and position calibration; pulsar backend delivers daily DM and RM time series for PTA arrays, mode-switching, and glitch monitoring (Collaboration et al., 2018, Collaboration et al., 2020). The absorber survey identifies H I absorbers at $z > 2$, demonstrating sensitivity to optical depths $\tau > 0.03$ on $\lesssim 100\, {\rm mJy}$ continuum backgrounds (Collaboration et al., 12 Jun 2025).

6. Data Validation, System Performance, and Operational Metrics

Routine performance includes a geometric collecting area of 8000 m², a system temperature $T_\mathrm{sys} ≈ 50\,\mathrm{K}$, instantaneous survey speed covering 3/4 of the northern sky daily, and field of view ≈ 250 deg² (1711.02104). The correlator operates at ≈220 kW including all cooling, with GPU and FPGA components achieving high-duty cycle via heat-managed liquid cooling and Faraday-cage shielding (Denman et al., 2020). Data losses stem primarily from daytime Sun contamination, weather, and feed downtime (cumulatively ≳69% for cosmology data) (Collaboration et al., 2022).

Pipeline validation includes:

• Statistical χ² and null tests on imaginary-power spectra, even–odd splits, Stokes Q, and sky-field splits, all yielding p-values >0.2 and excluding significant contamination (Collaboration et al., 24 Nov 2025).
• Quantified residuals from foreground subtraction, delay filters, and beam models, demonstrating stability under varying sub-band and time-field masks.
• Inter-benchmarking with eBOSS stacking and external surveys; all results agree within 1σ uncertainties.

Complementary pipeline improvements—per-integration RFI excision, sub-nanosecond cable delay calibration, and daily gain normalization—ensure robustness against spectral leakage and systematics; these are directly tied into the event-based hardware/data provenance chain for diagnostic and retrace purposes (Hincks et al., 2014).

7. Outlook and Scientific Impact

CHIME's pioneering array design, digital architecture, and systematic calibration have established high-significance H I intensity mapping in the post-reionization universe and precision daily monitoring of radio transients over a broad sky and frequency window. Its data infrastructure and calibration strategies set methodological standards for future transit telescopes and large-scale survey arrays. Ongoing improvements in beam mapping (holography, Sun drift, model-based approaches), RFI and polarization leakage control, and high-throughput intercontinental VLBI—with Outriggers and baseband recording—are expected to translate to transformative advances in cosmological constraint on BAO, dark energy, IGM baryon location (via FRBs), and radio transient astrophysics (Amiri et al., 2024, Collaboration et al., 7 Apr 2025).

The CHIME experiment thus defines both a model and a testbed for low-frequency digital aperture radio astronomy at extreme data and calibration scale, with a uniquely integrative commensal science program spanning cosmology, time-domain radio astronomy, and system engineering for high-volume observational data streams.