Batched GPU simulation¶

For large injection catalogues (e.g. a year of events) the JAX/ripple backend can generate and project the whole catalogue on device in batched, vmap-ped calls, then assemble the signals into uniform-duration data-segment files. The single-event CPU path (Strain injection, CBCSimulator) is unchanged and remains the default; this page covers the batched device path.

Two independent durations¶

A catalogue mixes two notions of "duration" that this API keeps separate:

• Generation buffer — how long each waveform must be to contain its inspiral from minimum_frequency. It is estimated from the post-Newtonian chirp time. A vmap needs one fixed shape, so the batch uses a single worst-case buffer sized for the longest (lowest-mass) event, unless you fix it with RippleBackend(segment_duration=…).
• Data-segment duration — the fixed length of each output data file. It is independent of any signal's length: a long signal simply spans several consecutive segments.

End to end in one call¶

import numpy as np
from gwmock_signal.jax_batch import simulate_cbc_catalogue

# Canonical gwmock-pop parameter names as a struct-of-arrays (one entry per event).
catalogue = {
    "detector_frame_mass_1": np.array([40.0, 36.0]),
    "detector_frame_mass_2": np.array([31.0, 29.0]),
    "luminosity_distance": np.array([400.0, 410.0]),
    "spin_1z": np.array([0.3, 0.0]),
    "spin_2z": np.array([-0.2, 0.0]),
    "inclination": np.array([0.9, 0.4]),
    "coa_phase": np.array([0.3, 0.0]),
    "right_ascension": np.array([1.375, 0.5]),
    "declination": np.array([-1.211, 0.3]),
    "polarization_angle": np.array([2.659, 0.0]),
    "coa_time": np.array([1_126_259_500.0, 1_126_259_600.0]),
}

segments = simulate_cbc_catalogue(
    "IMRPhenomD",
    ["H1", "L1"],
    sampling_frequency=4096.0,
    minimum_frequency=20.0,
    parameters=catalogue,
    segment_duration=64.0,
    start_time=1_126_259_456.0,
    end_time=1_126_259_968.0,
)
# segments: list[DetectorStrainStack], one per 64 s segment, zero-noise + injected signals.
for stack in segments:
    ...  # write each stack (see Multichannel strains)

simulate_cbc_catalogue is a convenience over the two building blocks below; use them directly when you need non-zero backgrounds or your own segment tiling.

Building blocks¶

simulate_cbc_batch — batched, on-device generation and projection. Returns a BatchedDetectorStrain: a (n_events, n_detectors, n_samples) JAX array (on device) plus per-event timing metadata (coa_time, epoch, sampling_frequency). Each event/detector row is a signal with coalescence near the segment end; its first sample is at epoch + coa_time[event].

assemble_segments — host-side scatter of those signals onto a tiling of fixed-duration segments. Each signal is injected into every segment it overlaps via inject_strains_sequential, which crops it to the segment — so a signal longer than one segment contributes its overlapping part to each segment it spans. Segments are zero-noise by default; pass backgrounds (one mapping per segment) to inject into existing data, e.g. coloured noise from gwmock-noise.

from gwmock_signal.jax_batch import simulate_cbc_batch, assemble_segments

batch = simulate_cbc_batch(
    "IMRPhenomD", ["H1", "L1"],
    sampling_frequency=4096.0, minimum_frequency=20.0, parameters=catalogue,
)
segments = assemble_segments(
    batch,
    segment_duration=64.0,
    segment_start_times=[1_126_259_456.0, 1_126_259_520.0, ...],
    # backgrounds=[{"H1": ts, "L1": ts}, ...],  # optional, aligned with the starts
)

Conventions and current limitations¶

• Earth rotation. The device path evaluates the antenna pattern and geocenter delay once per event at the segment midpoint (earth_rotation=False), matching project_polarizations_to_network. The time-dependent (earth_rotation=True) response is not available on device.
• Sidereal time. Sidereal time uses an on-device IAU model with default leap seconds and dut1 = 0, accurate to ~1e-4 rad; this is the small difference from the Astropy-based CPU path.
• Worst-case buffer. A wide mass range makes the shared generation buffer as long as the lowest-mass event needs. For very heterogeneous catalogues, simulate in chirp-mass groups to bound memory.
• Validation. Every event/detector of the device path is cross-checked against the CPU pipeline (white overlap > 0.999); the small gap is the sidereal-time default above.