IMRPhenomD on GPU

IMRPhenomD is a fully vectorised, torch.compile-compatible implementation of the IMRPhenomD aligned-spin frequency-domain waveform model. It computes batches of \(h_+\) and \(h_\times\) polarisations from a batch of physical parameters.

Note

In production, signal generation is handled through IMRPhenomPv2, which extends IMRPhenomD to precessing spins and wraps parameter sampling, waveform generation, multi-detector projection, and SNR rescaling into a single callable. See IMRPhenomPv2 on GPU for the recommended production usage.

Parameters

Each waveform is characterised by 11 physical parameters:

Index	Parameter	Description
0	`m1_msun`	Primary mass in solar masses
1	`m2_msun`	Secondary mass in solar masses
2	`chi1`	Aligned spin of the primary (−1 to +1)
3	`chi2`	Aligned spin of the secondary (−1 to +1)
4	`dist_mpc`	Luminosity distance in Mpc
5	`tc`	Time of coalescence in seconds (relative to segment start)
6	`phic`	Coalescence phase in radians
7	`inclination`	Inclination angle in radians
8	`polarization_angle`	Polarisation angle in radians
9	`ra`	Right ascension in radians
10	`dec`	Declination in radians

Comparison with LALSuite

Sage’s IMRPhenomD is validated against lalsimulation.SimInspiralChooseFDWaveform. The two agree to floating-point precision (residuals ≈ 0) for the plus polarisation:

import lal, lalsimulation as lalsim

approximant = lalsim.SimInspiralGetApproximantFromString("IMRPhenomD")
hp_lal, hc_lal = lalsim.SimInspiralChooseFDWaveform(
    30.0 * lal.MSUN_SI, 30.0 * lal.MSUN_SI,
    0.0, 0.0, 0.99, 0.0, 0.0, 0.99,
    440e6 * lal.PC_SI, 0.5, 0.5,
    0.0, 0.0, 0.0,
    1.0 / 20.0, 20.0, 2048.0, 20.0,
    None, approximant,
)

# Residual between Sage and LAL
import matplotlib.pyplot as plt
plt.plot(hc_lal.data.data[:40_000] - hc[0].detach().cpu().numpy()[:40_000])
# Residual is at absolute zero (bit-identical within float64 precision)

Note

LALSuite zero-pads its output to the next power of two. A requested length of 40 961 samples (20 s at 2048 Hz + DC bin) will be returned as 65 537 samples. Sage returns exactly the requested length.