Waveform catalog and support

Choose a model by its physical content, then check the evidence and execution restrictions for the options you need. The catalog records implemented families; a familiar approximant name does not establish equivalence to an external model. For availability in other packages, use the dated package comparison. For the mapping from every public dispatcher name to a mathematical specification status, see specification coverage. The waveform library contains the first three scoped, versioned definitions and their component evidence.

Model catalog

Family	Approximant(s)	Harmonics / Multipoles	Physical Features & Regimes
Post-Newtonian	TaylorF2 TaylorF2_HM TaylorF2_Eccentric	\((2,2)\) Eight modes through \(\ell=4\) \((2,2)\)	3.5PN point-particle & spin phasing, BNS tidal deformability Leading-order higher-mode amplitudes Moore et al. (2016) 3PN stationary-phase eccentric corrections
Aligned-Spin IMR	IMRPhenomD IMRPhenomHM IMRPhenomXAS IMRPhenomXHM	\((2,2)\) \((2,\pm 2), (2,\pm 1), (3,\pm 3), (3,\pm 2), (4,\pm 4)\) \((2,\pm 2)\) \((2,\pm 2), (2,\pm 1), (3,\pm 3), (3,\pm 2), (4,\pm 4)\)	Aligned-spin dominant mode, 5-collocation intermediate Multi-harmonic PhenomD calibration with mode mixing 4th-generation aligned-spin dominant mode (Pratten et al. 2020) 4th-generation higher multipoles (García-Quirós et al. 2020)
Native Time-Domain IMR	IMRPhenomT, IMRPhenomTHM, IMRPhenomTP, IMRPhenomTPHM	Up to \((5,\pm 5)\)	Aligned or precessing black hole binaries; see the T-family guide
Precessing IMR	IMRPhenomP / Pv2 IMRPhenomXP IMRPhenomXPHM	\((2,\pm 2)\) twisted \((2,\pm 2)\) twisted Multipolar twisted	Generic 3D spins with NNLO PN precession angles & Wigner \(d\)-matrices 4th-gen precessing carrier with MSA analytical precession Multipolar precession supporting MSA (v223), SpinTaylor (v320), and PNR (v330)
Tidal & Matter	IMRPhenomD_NRTidal IMRPhenomD_NRTidalv2 IMRPhenomNSBH	\((2,2)\) \((2,2)\) \((2,2)\)	Dietrich et al. (2017) rational Padé BNS tidal phasing (NRTidal v1) Dietrich et al. (2019) NRTidal v2 with self-spin & Planck windowing Neutron star–black hole model with tidal disruption & remnant torus
Effective-One-Body	SEOBNR_ROM / SEOBNRv4_ROM	\((2,2)\)	External SEOBNRv4 ROM data with a native Torch/JAX evaluator
Ringdown & Bursts	KerrRingdown SineGaussian MemoryWaveform	\((2,2,0), (l,m,n)\) Burst packet Supplied \((2,\pm2)\) modes	Damped sinusoids from Kerr quasinormal modes (time & frequency domain) Standard unmodeled sine-Gaussian burst wavelet Versioned dominant-pair direct-class memory through evaluate_22; the older polarization-only carrier is diagnostic
Specialized Models	ContinuousWave EMRIKludge ParameterizedPostEinstein	Monochromatic Multi-harmonic Modified phasing	Continuous waves from triaxial rotating neutron stars with Doppler modulation Extreme mass ratio inspirals (LISA) with Kerr geodesic orbital frequencies Model-independent testing of General Relativity (\(\beta_{\rm ppE}, b_{\rm ppE}\))
Time-Domain Tools	get_td_waveform HybridMECO	Supported FD models Diagnostic	FFT conditioning and time-domain waveform generation Minimum Energy Circular Orbit hybrid velocity & frequency root solver

Generating a waveform

Select a frontend with backend="torch" or backend="jax" in the unified diffgw API, or call torchwave / jaxwave directly. For example, IMRPhenomXP is available through get_fd_waveform in both frontends, just as TaylorF2, IMRPhenomD and IMRPhenomXAS are. The generation guide explains the shared interface; its examples illustrate that interface rather than enumerate every available model.

Generation task	Guide
Generate frequency-domain polarizations	Frequency grids and batching; IMRPhenomXP example
Set precessing spins, modes, tides or eccentricity	Model parameters and examples, parameter conventions
Generate native T / THM / TP / TPHM time series	T-family generation guide; these models require get_td_waveform
Convert a supported frequency-domain model to a time series	Time-domain generation using a conditioned inverse Fourier transform

The catalog also includes model classes and diagnostic tools, which have their own entry points. Consult the PyTorch or JAX API reference for these; a catalog entry does not imply support through both get_fd_waveform and get_td_waveform.

Execution restrictions

Generation, batching, compilation and differentiation are separate capabilities. Use the execution guide for device, precision and transform setup, then check the restrictions of the model and options you select:

• T family: native time-domain evaluation only. Parameter batches share a time grid. Fixed grids support first derivatives; higher derivatives through frequency inversion are unsupported. Automatic duration selection is eager; supply a fixed duration inside JAX transformations. See the grid and derivative restrictions.
• SEOBNR_ROM / SEOBNRv4_ROM: native evaluation requires a selected external data asset. Its identity and provenance determine which surrogate you are evaluating.
• Precessing and higher-mode paths: mode selection and the precession prescription can change batching and derivative support. Check the precession examples and numerical-precession restrictions.
• Precision and devices: PyTorch MPS does not support the FP64 examples; use CPU on Apple Silicon. JAX's non-symmetric Kerr angular eigensolver uses a host CPU callback. See hardware guidance and model-specific limitations.

Validation coverage

Evidence answers what was checked, with which settings. It does not define which waveforms can be generated. The published timing campaigns cover selected models and settings; that is a measurement scope, not the full supported catalog.

• Measured execution and waveform agreement: the FP64 campaign covers TaylorF2, IMRPhenomD and IMRPhenomXAS with Torch eager/compiled and JAX JIT on CPU/CUDA for specified grids and populations. Its accuracy qualifications include failed configurations. These measurements do not establish other models' performance or support on other accelerators. A separate PhenomXPy comparison measures T/THM on CPU with Torch eager and JAX JIT; failed strict agreement checks are marked as diagnostic timings.
• Model-specific reference agreement: Differences from LAL records configurations, deliberate differences and unresolved discrepancies, including XP's precession, final-spin and reference-phase choices. Independent all-eight leading amplitudes for TaylorF2_HM agree with a checksum-pinned author supplement. Its signed leading terms, nonspinning phase, and Eq. (11) normalization also agree for an all-eight unwindowed complex raw-mode slice with f_ref=None. Exact external all-eight complex-waveform parity remains unprovable because that symbolic asset is parsed by this project and does not independently generate the aligned-spin, projection, reference-frequency, and cutoff contract. Physical validation remains unavailable. ROM evidence applies to the selected asset. The T-family report is a separate record.
• Derivatives: derivative support and evidence states the checked parameters, derivative orders and restrictions. Waveform-value agreement and framework autodiff do not establish derivative accuracy.

The validation overview distinguishes mathematical-reference checks, frontend consistency, external-reference agreement and physical validation. The manifest records selected configurations, not a complete support matrix; it does not yet include the T family. Treat a missing check as unestablished evidence, not as evidence that generation is unavailable.

The unvalidated native SEOBNR surrogate, SEOBNR_ODE / SEOBNRv4Published assemblies, TEOBResumS waveform, SpinPrecessionODE and experimental XHM phase connections are development-branch work and are unavailable in releases. The retired-route record preserves former names, manifest metadata, and research provenance. SEOBNR_ROM, including direct class construction, consistently selects the official external-data evaluator. Independently tested EOB equation components have a separate scope: the retained generic 3PN and TEOB Hamiltonians are nonspinning; the TEOB component supports vacuum parameters only. These components do not constitute validated complete EOB waveform models.

Parameter domains and conventions

The campaign populations are tested domains, not calibration boundaries. Restricted cases and component checks must not be extrapolated to a model's entire physical domain. Where a calibration range or option is not established by the linked model documentation and provenance, this catalog makes no domain-wide accuracy claim.

The waveform guide describes generation. The parameters and conventions reference covers units, reference phase/frequency, spin frames, sampling and mode selection. The T-family guide specifies its time origin, duration, conventions and rejected options. Other model-specific conventions are recorded in the manifest and differences from LAL.