LiteBIRD Simulation Framework (LBS)

• LiteBIRD Simulation Framework is a Python-based tool that models the entire data acquisition process for CMB observations with a focus on instrument modeling and systematic effects.
• It employs advanced scanning strategies, detailed noise and HWP modeling, and real-time data handling to support mission-level analyses and requirement setting.
• The open-source infrastructure integrates with end-to-end pipelines and external Bayesian tools, ensuring reproducibility and cross-validation in cosmological inference.

The LiteBIRD Simulation Framework (LBS) is a Python package designed for the implementation of pipelines that model the outputs of the data acquisition process from the three instruments on the LiteBIRD spacecraft—LFT, MFT, and HFT—and, more broadly, a reusable simulation infrastructure for LiteBIRD end-to-end analyses, specialized systematic-effect studies, and reduced map-based approximations (Tomasi et al., 7 Jul 2025). In the literature it is presented both as a software framework and as the backbone of mission-scale numerical pipelines: it simulates scanning strategy, sky observation, instrumental noise, and selected systematic effects; it interfaces with the LiteBIRD Instrument Model (IMo); and it supports propagation from detector-level timelines to maps, angular power spectra, and, in some studies, biases on the tensor-to-scalar ratio $r$ (Bortolami et al., 8 Jul 2025, Stellati et al., 3 Sep 2025).

1. Definition, mission role, and scope

LBS was developed for LiteBIRD’s requirement-driven simulation program. LiteBIRD is a space mission aimed at measuring CMB $B$-mode polarization and constraining primordial inflation, with a target sensitivity of $\delta r \leq 10^{-3}$ (Tomasi et al., 7 Jul 2025). The framework was created because the collaboration required not a single one-off simulator, but a common infrastructure able to support multiple pipelines with shared mathematical models, file formats, provenance tracking, and instrument descriptions (Tomasi et al., 7 Jul 2025).

In this role, LBS is described as supporting several distinct simulation modes. These include full end-to-end pipelines, studies of half-wave plate (HWP) systematics, transient-source simulations, and map-based approximations (Tomasi et al., 7 Jul 2025). The first official LiteBIRD simulation release was realized with an end-to-end pipeline developed using LBS, and that release was explicitly tied to the mission objective of detecting primordial CMB $B$-modes with a target uncertainty on the tensor-to-scalar ratio of $\delta r \sim 0.001$ (Bortolami et al., 8 Jul 2025).

The framework’s scope extends beyond nominal sky-plus-noise production. It is used to establish requirements, optimize observing strategy, validate ground-segment software and data storage, and create realistic data sets for Bayesian and component-separation analyses (Bortolami et al., 8 Jul 2025). In systematic-effect papers, the same framework is used as an end-to-end bridge from instrumental perturbations to cosmological consequences, for example in studies of beam far side-lobe uncertainty and reflective-HWP misalignment (Leloup et al., 2023, Stellati et al., 3 Sep 2025).

A plausible implication is that LBS occupies the boundary between instrument modeling, scan simulation, and cosmological inference rather than belonging exclusively to any one of those layers. That interpretation is consistent with its repeated use as a pipeline backbone in studies that begin with detector or optical perturbations and end with $\delta r$-level requirements.

2. Software architecture, data model, and provenance

LBS is implemented in Python, runs on 64-bit Unix systems, and is distributed as an open-source package, litebird_sim, on PyPI (Tomasi et al., 7 Jul 2025). The code uses NumPy for array and matrix operations, Numba for CPU-intensive kernels, AstroPy for time and coordinate handling, Healpy for HEALPix maps, and HDF5/FITS for I/O (Tomasi et al., 7 Jul 2025). Version 0.11.0 is the release identified as the one used in the companion end-to-end study, although the software evolved further through versions 0.12.0, 0.13.0, 0.14.0, and 0.15.1 (Tomasi et al., 7 Jul 2025).

The central object is the Simulation object, which acts as a container for mission-level parameters, the instrument model connection, the random-number generator, allocated observations, and the provenance record used for reporting (Tomasi et al., 7 Jul 2025). This object stores both inputs and generated outputs and serves as the main anchor for assembling a pipeline. The framework is therefore organized around a persistent simulation state rather than isolated scripts.

Time-ordered data (TOD) are treated as a core data structure. A TOD is stored as a $2\mathrm{D}$ matrix with rows corresponding to detectors and columns to time samples (Tomasi et al., 7 Jul 2025). Because LiteBIRD-scale simulations are large, LBS supports MPI-based distribution of these matrices across processes. The split can be by time, by detectors, or by both dimensions, controlled through n_blocks_det and n_blocks_time, with redistribution available through Observation.set_n_blocks() (Tomasi et al., 7 Jul 2025). An Observation may hold multiple TOD-like matrices simultaneously, allowing users to preserve separate components such as sky signal, white noise, $1/f$ noise, CMB, dust, or synchrotron as distinct arrays (Tomasi et al., 7 Jul 2025).

Detector metadata are kept alongside the timelines. Examples explicitly listed include wafer name, white-noise level, $1/f$ knee frequency, and noise slope (Tomasi et al., 7 Jul 2025). Pointing information is stored as an $N \times M \times 3$ array corresponding to colatitude, longitude, and orientation. In version 0.11.0, orientation and HWP angle were combined into a single polarization angle; later versions separated them, and version 0.14.0 no longer automatically folded polarization angle into beam orientation when using $B$0 beams (Tomasi et al., 7 Jul 2025). This evolution reflects a shift from idealized pencil-beam modeling toward more realistic beam and HWP treatments.

A major architectural element is the versioned Instrument Model database, IMo, implemented via InstrumentDB backed by SQLite and exposed through local or remote access (Tomasi et al., 7 Jul 2025). The IMo stores versioned information about instruments, channels, detectors, and scanning parameters. LBS records every IMo query and the version used, and it prefers a local exported copy of the IMo for large simulations because querying the remote server would be too slow and could overload the SSDC server (Tomasi et al., 7 Jul 2025). Output is written as HDF5 for timelines and pointings and FITS for HEALPix maps; in version 0.11.0, each MPI process writes its own HDF5 file (Tomasi et al., 7 Jul 2025).

Reproducibility is treated as a first-class requirement. LBS can generate a Markdown/HTML report automatically, copy the input TOML parameter file to the output directory, record the Git commit hash, save git diff output if the working tree is dirty, and require explicit initialization of the random-number generator through random_seed (Tomasi et al., 7 Jul 2025). The framework also includes unit tests, integration tests, and end-to-end tests, along with assertions and explicit error checks for incorrect MPI splitting, inconsistent TOD/map handling, wrong module call order, frequencies outside bandpass, and inconsistent scanning inputs (Tomasi et al., 7 Jul 2025).

3. Scanning, sky, dipole, noise, and HWP modeling

LBS models the LiteBIRD scanning strategy as a composition of rotations: spin around the spin axis, precession of the spin axis around the Sun–Earth axis, and Earth’s revolution around the Sun (Tomasi et al., 7 Jul 2025). Quaternions encode these rotations at low sampling rate, and spherical linear interpolation (slerp) is used to interpolate pointings to detector sample times (Tomasi et al., 7 Jul 2025). In the first simulation release, the scanning strategy parameters loaded through the IMo include a boresight angle $B$1, a spin rate of $B$2 rpm, a precession angle $B$3, and a precession period of about $B$4 min; the HWP rotation rates are $B$5 rpm for LFT, $B$6 rpm for MFT, and $B$7 rpm for HFT (Bortolami et al., 8 Jul 2025).

The sky-simulation layer includes Galactic diffuse foregrounds, extragalactic signals, the CMB, and the dipole. LBS uses the Map-Based Simulation module, MBS, as a wrapper around PySM (Bortolami et al., 8 Jul 2025). In the first release, Galactic foregrounds include thermal dust, synchrotron, anomalous microwave emission (AME), free-free emission, and CO rotational line emission; extragalactic signals include thermal and kinetic Sunyaev–Zel’dovich effects, the cosmic infrared background, lensing convergence, and frequency-dependent radio sources from WebSky simulations and related catalogs (Bortolami et al., 8 Jul 2025). The CMB is generated as 500 Monte Carlo realizations using Planck 2018 cosmological parameters and then lensed using the WebSky convergence map via lenspix (Bortolami et al., 8 Jul 2025). These sky components are convolved with circular Gaussian beams using channel-specific FWHM values before timeline generation (Bortolami et al., 8 Jul 2025).

For map scanning, version 0.11.0 assumes ideal pencil beams with essentially zero FWHM and no sidelobes, so scanning a sky map reduces to sampling the HEALPix map at the detector boresight location for each time sample (Tomasi et al., 7 Jul 2025). The paper explicitly notes that full beam convolution was not yet available in 0.11.0, while later versions introduced $B$8 beam convolution using Ducc (Tomasi et al., 7 Jul 2025). In systematic-effect studies requiring sub-pixel accuracy, LBS can use bilinear interpolation with four nearest neighbors, for example when a wedge-like HWP misalignment induces tiny pointing offsets smaller than the pixel scale (Stellati et al., 3 Sep 2025).

The dipole model is treated explicitly. The moving-frame CMB temperature is written as

$B$9

with $\delta r \leq 10^{-3}$0, $\delta r \leq 10^{-3}$1 the line of sight, and $\delta r \leq 10^{-3}$2 defined in the cited implementation (Tomasi et al., 7 Jul 2025). In the first simulation release the dipole is included under the LBS nomenclature TOTAL_FROM_LIN_T, with

$\delta r \leq 10^{-3}$3

where

$\delta r \leq 10^{-3}$4

(Bortolami et al., 8 Jul 2025).

Noise modeling includes white noise and $\delta r \leq 10^{-3}$5 correlated noise. In the framework paper the noise power spectrum is written as

$\delta r \leq 10^{-3}$6

where $\delta r \leq 10^{-3}$7 is white-noise power, $\delta r \leq 10^{-3}$8 the knee frequency, $\delta r \leq 10^{-3}$9 the slope, and $B$0 a regulator at low frequency (Tomasi et al., 7 Jul 2025). In the first official release, LBS simulates white-noise timelines and two mixed white-plus-$B$1 timelines per detector with $B$2 mHz and $B$3 mHz, the former intended to be realistic and the latter pessimistic (Bortolami et al., 8 Jul 2025). The frequency-domain implementation is given as

$B$4

with $B$5 (Bortolami et al., 8 Jul 2025).

The HWP treatment depends on the application and software version. In version 0.11.0, the HWP model is ideal and only rotates the detector polarization orientation (Tomasi et al., 7 Jul 2025). The broader LiteBIRD simulation program, however, repeatedly exploits the fact that polarized incident radiation is modulated at four times the HWP rotation frequency, shifting polarized signal above the $B$6 knee and helping separate sky signal from low-frequency noise (Bortolami et al., 8 Jul 2025). In scan-optimization work, the continuously rotating HWP changes the scan-design problem by removing the need for rapid spacecraft spin to modulate polarization directly, while introducing a coupling between scan speed and HWP rotation rate (Takase et al., 2024).

4. End-to-end pipelines and the first simulation release

The first official LBS-based end-to-end pipeline is built around LiteBIRD’s baseline configuration: a satellite at the Sun–Earth $B$7 point, observing the full sky in 15 frequency bands from 34 to 448 GHz for a 3-year mission, with 4508 detectors sampled at 19.1 Hz (Bortolami et al., 8 Jul 2025). The nominal mission performance cited there is an effective polarization sensitivity of $B$8 and angular resolution of 31 arcmin at 140 GHz (Bortolami et al., 8 Jul 2025). The first release itself is a reduced pathfinder configuration using about one-third of the full mission duration and about one-third of the detectors, corresponding to 1678 simulated detectors out of 4508 (Bortolami et al., 8 Jul 2025).

The released products include 500 full-sky simulated maps at HEALPix $B$9, and, for the first simulation only, one year of TOD for approximately one-third of LiteBIRD’s total detectors, along with pointing information and noise covariance matrices (Bortolami et al., 8 Jul 2025). The channels cover the three LiteBIRD telescopes and use 22 channels from the IMo database; the beam FWHM values in the release span roughly from 70.5 arcmin at 40 GHz down to 17.9 arcmin at 402 GHz (Bortolami et al., 8 Jul 2025).

The pipeline architecture follows a modular sequence from IMo ingestion to map production. The IMo supplies detector counts, beam FWHM, detector efficiencies, noise levels, HWP rotation rates, and detector-orientation quaternions (Bortolami et al., 8 Jul 2025). Sky maps are prepared component by component at $\delta r \sim 0.001$0, scanned into detector TOD using the LBS pointing machinery, combined with dipole and noise components, and then converted into full-sky maps by a simple binner or naive map-maker (Bortolami et al., 8 Jul 2025). The standard map-making relation is given as

$\delta r \sim 0.001$1

with pixel-space covariance

$\delta r \sim 0.001$2

(Bortolami et al., 8 Jul 2025). In the release, the covariance is generated per channel and has $\delta r \sim 0.001$3 blocks for $\delta r \sim 0.001$4, $\delta r \sim 0.001$5, $\delta r \sim 0.001$6, and the cross-terms $\delta r \sim 0.001$7, $\delta r \sim 0.001$8, $\delta r \sim 0.001$9 (Bortolami et al., 8 Jul 2025).

The paper reports extensive validation. TOD are checked visually and through power spectra; noise covariance matrices are compared against binned-map residuals; and full-sky spectra of output-minus-input maps are averaged over 500 realizations (Bortolami et al., 8 Jul 2025). In polarization, residual-map spectra are consistent with white noise, which the authors interpret as evidence that the HWP suppresses $\delta r$0 noise effectively, whereas in temperature residual low-$\delta r$1 power increases for $\delta r$2 mHz relative to 30 mHz (Bortolami et al., 8 Jul 2025). One limitation is explicitly acknowledged: a bug in the noise timestream generation caused an unintended increase in $\delta r$3 noise, potentially affecting temperature simulations, though the impact is regarded as minor given the high signal-to-noise ratio in temperature (Bortolami et al., 8 Jul 2025).

The computational footprint of this first release is large but concrete. The total disk footprint is about 35 TB, and the production and validation required roughly 600 kcpu.h on 370 nodes of the MARCONI-CINECA cluster (Bortolami et al., 8 Jul 2025). The authors anticipate that full focal-plane simulations over the full three-year mission will require about 1 Mcpu.h (Bortolami et al., 8 Jul 2025).

5. LBS as a framework for systematic effects and requirement setting

A defining feature of LBS in the literature is its repeated use as an end-to-end engine for requirement derivation. In the far side-lobe study, LBS functions as the organizing pipeline connecting beam simulations, sky simulations, scanning, map-making, beam correction, component separation, and cosmological inference (Leloup et al., 2023). The pipeline generates a sky model, produces beam-convolved data, applies beam correction and Galactic masking, runs a parametric component-separation method, and then converts residual $\delta r$4-mode contamination into a bias $\delta r$5 (Leloup et al., 2023). Under the assumptions of that study, the key quantity is the integrated fractional power mismatch in the far side-lobes, which must be known at approximately the $\delta r$6 level to keep the beam-induced bias below

$\delta r$7

(Leloup et al., 2023).

In the reflective-HWP misalignment study, LBS is used as the end-to-end numerical engine that propagates a deliberately injected optical misalignment from the instrument level to a final cosmological-parameter bias (Stellati et al., 3 Sep 2025). The systematic is a small constant tilt between the HWP rotation axis and the optical axis, parameterized by a wedge angle $\delta r$8. Because the reflective HWP is assumed ideal and plane-parallel, the misalignment acts like a physical wedge; the reflected beam deviation is $\delta r$9, and the resulting sky displacement is a circle of angular diameter $2\mathrm{D}$0 (Stellati et al., 3 Sep 2025).

In that study, the detector signal in the idealized modulation model is

$2\mathrm{D}$1

while for orthogonal detector pairs with perfectly matched detectors the differential timestream is

$2\mathrm{D}$2

(Stellati et al., 3 Sep 2025). LBS is used to generate CMB-only $2\mathrm{D}$3 maps, scan them into TOD with wedge-modified pointing using the LiteBIRD scan strategy, reconstruct maps using the nominal scan model without knowledge of the wedge, and compare recovered maps and spectra to inputs (Stellati et al., 3 Sep 2025). Because the wedge-induced offsets are tiny, bilinear interpolation with four nearest neighbors is used to evaluate the sky at sub-pixel displaced positions (Stellati et al., 3 Sep 2025).

The principal result is that the contamination mainly resembles lensing $2\mathrm{D}$4-modes rather than primordial tensor modes (Stellati et al., 3 Sep 2025). The likelihood-based systematic bias is defined as

$2\mathrm{D}$5

and is compared to a mission systematic budget with

$2\mathrm{D}$6

From a fourth-order polynomial fit, the study infers a maximum allowable wedge angle

$2\mathrm{D}$7

for the two-detector case, corresponding to a mounting displacement of about $2\mathrm{D}$8 mm for a 500 mm HWP (Stellati et al., 3 Sep 2025). When the number of detectors increases, the allowable wedge becomes less stringent: $2\mathrm{D}$9 arcmin for 2 detectors, $1/f$0 arcmin for 4 detectors, and $1/f$1 arcmin for 6 detectors (Stellati et al., 3 Sep 2025).

These studies illustrate a general pattern. LBS is not restricted to producing nominal simulations; it is also used to isolate specific perturbations, reconstruct data under deliberately mismatched assumptions, and translate residual power into mission-level tolerances. This suggests that its scientific importance lies as much in requirement setting and systematic budgeting as in synthetic data generation.

6. Scan optimization, external Bayesian pipelines, and computational feasibility

LiteBIRD scan strategy is itself treated as a simulation input that can be optimized quantitatively. The scan-optimization study parameterizes the strategy by

$1/f$2

with geometric and kinetic constraints derived from full-sky coverage, Sun-shield and thermal requirements, and HWP modulation conditions (Takase et al., 2024). It introduces metrics directly relevant to an LBS-style scan simulator: integrated planet visibility time, hit-map uniformity,

$1/f$3

and the cross-link factor,

$1/f$4

(Takase et al., 2024). The same work introduces Falcons.jl, a fast spacecraft scanning simulator, and validates a map-based method against TOD-based results for time-independent systematics with differences around $1/f$5 for pointing offset and negligible differences for HWP non-ideality (Takase et al., 2024). The standard LiteBIRD configuration

$1/f$6

is identified as the best practical compromise within that analysis (Takase et al., 2024).

LBS outputs also interface with external Bayesian analysis software. In the Cosmoglobe/Commander3 feasibility study, the simulations analyzed are first-generation TOD produced by LBS via its map-based simulation module (Aurvik et al., 7 Jul 2025). LBS generates HDF5 TOD products that are converted into Commander3-compatible format, ingested, compressed in memory using Huffman compression, and analyzed with Gibbs sampling (Aurvik et al., 7 Jul 2025). The study treats this as the first non-LBS application of these TOD simulations and uses it to verify that the simulations can be readily used by external codes (Aurvik et al., 7 Jul 2025).

The computational estimates from that work quantify the scale of full Bayesian end-to-end analysis. For the reduced data set analyzed there, the uncompressed TOD volume is 1.55 TB and the compressed volume is 470 GB; extrapolated to the full three-year mission, the corresponding estimates are 238 TB and 70 TB (Aurvik et al., 7 Jul 2025). The full-mission estimate for one Gibbs sample is approximately 3000 CPU hours, implying about 3 million CPU-hours for a 1000-sample production run (Aurvik et al., 7 Jul 2025). The study concludes that such requirements are within the capabilities of future high-performance computing systems, but it also emphasizes that the current simulations are based on an ideal instrument model and include only correlated $1/f$7 noise (Aurvik et al., 7 Jul 2025).

Future development directions are stated explicitly across the LBS literature. Planned extensions include the full focal plane, the full three-year mission duration, and additional instrumental systematics such as gain drifts, downtime, cosmic-ray impacts, beam systematics, and HWP systematics (Bortolami et al., 8 Jul 2025). The framework paper notes that realistic HWP simulators exist in later versions beyond 0.11.0 (Tomasi et al., 7 Jul 2025), while the Bayesian feasibility study identifies half-wave plate non-idealities, gain fluctuations, beam modeling, sidelobe modeling, and more complicated foreground components as required future additions for a fully realistic mission analysis (Aurvik et al., 7 Jul 2025). A plausible implication is that LBS is intended to evolve from a mission simulation framework into the principal environment for continuous cross-validation between instrument design, data analysis, and cosmological inference.