ElasticTTI2nd¶
Governing equations¶
ElasticTTI2nd solves the second-order displacement equations of a tilted
transversely isotropic solid in the \(x\)-\(z\) plane (P-SV only), following
Oh et al. (2020, eqs 11–12). With the displacements \(u_x\), \(u_z\) (ux, uz)
and the density \(\rho\) (rho),
The stresses are not carried as wavefields; they are formed inside each step.
Stiffness from the model parameters. The VTI constants follow the
parametrization of Oh et al. (2020, eq. 19), with the horizontal P velocity
\(v_h\) (vh), the vertical S velocity \(v_s\) (vs), \(\rho\), Thomsen's
\(\varepsilon\) (epsilon) and the anellipticity
\(\eta = (\varepsilon - \delta)/(1 + 2\delta)\) (eta):
Here \(v_h^2/(1+2\varepsilon)\) is the vertical P velocity squared and
\(v_h^2/(1+2\eta)\) the NMO velocity squared, so \(C_{13}\) agrees with Thomsen's
\(\delta\). The tilt \(\theta\) (theta) rotates the symmetry axis from \(z\)
toward \(x\), to \((\sin\theta, \cos\theta)\), through the 2-D Bond rotation
(Oh et al. 2020, eq. 10). With \(c = \cos\theta\), \(s = \sin\theta\),
with \(A = C_{13} - C_{11} + 2C_{55}\) and \(B = C_{13} - C_{33} + 2C_{55}\).
These are the \(\phi = 0\) entries of the rotation used by
ElasticTTI. Unlike there, \(\theta = 0\) is not
special-cased, so the gradient with respect to theta is also defined in
untilted regions. There is no \(\gamma\) or \(\phi\): the SH system is not
modelled.
Discretisation¶
Both displacement components live on the same grid nodes. Each term
\(\partial_a (\tilde C\, \partial_b u)\) is evaluated as
\(D_a^{-} \big[\tilde C\, D_b^{+} u\big]\), a forward staggered difference
\(D^{+}\) followed by a backward one \(D^{-}\), both of order spatial_order
(default 4) and without interpolation. The discrete stiffness operator is then
symmetric, \(K = -D^{\mathsf T} \tilde C D\), and has no checkerboard null space;
the mixed terms (\(a \neq b\)) carry a diagonal half-cell offset. Time stepping
is the second-order central difference,
Oh et al. (2020) solve the same equations with a frequency-domain finite element method; the time-domain scheme here is not theirs.
Boundary conditions¶
The default pml_type="cpmls" applies a CPML to every first derivative: the
inner forward differences use the half-node profiles, the outer backward
differences the integer-node profiles (see
Absorbing boundaries). For this second-order
displacement system the plain CPML grows at late times, which Li and
Bou Matar (2010) report for the same CPML and which follows from the
stability condition of Bécache et al. (2003). The profiles are therefore made
multiaxial (Meza-Fajardo and Papageorgiou, 2008, in the form of Li and
Bou Matar, 2010, eq. 23):
where \(d\) is the damping profile, \(r\) is the constructor argument
mpml_ratio (default 0.05) and the frequency-shift \(\alpha\) of each axis is
kept. mpml_ratio=0 restores the plain CPML. A free surface is not
implemented: free_surface=True is refused at construction, because the
isotropic image method would not give a stress-free surface for an
anisotropic medium.
References¶
- Bécache, E., S. Fauqueux, and P. Joly, 2003, Stability of perfectly matched layers, group velocities and anisotropic waves: Journal of Computational Physics, 188(2), 399–433, doi:10.1016/s0021-9991(03)00184-0.
- Li, Y. F., and O. Bou Matar, 2010, Convolutional perfectly matched layer for elastic second-order wave equation: The Journal of the Acoustical Society of America, 127(3), 1318–1327, doi:10.1121/1.3290999.
- Meza-Fajardo, K. C., and A. S. Papageorgiou, 2008, A nonconvolutional, split-field, perfectly matched layer for wave propagation in isotropic and anisotropic elastic media: Stability analysis: Bulletin of the Seismological Society of America, 98(4), 1811–1836, doi:10.1785/0120070223.
- Oh, J.-W., Y. Shin, T. Alkhalifah, and D.-J. Min, 2020, Multistage elastic full-waveform inversion for tilted transverse isotropic media: Geophysical Journal International, 223(1), 57–76, doi:10.1093/gji/ggaa295.
sweep.equations.ElasticTTI2nd ¶
Bases: sweep.equations.base.FirstOrderEquation
Displacement-based 2-D elastic TTI wave equation (Oh et al. 2020).
Second-order-in-time P-SV formulation rho u_tt = div(C~ : grad u)
(GJI 223, eqs 11-12) with the hierarchical VTI parametrization of
their eq. (19): horizontal P velocity vh, vertical S velocity
vs, density, Thomsen epsilon, anellipticity eta and the
tilt theta — all six differentiable. prepare_models builds
the four VTI constants and Bond-rotates them (their eq. 10) into the
six TTI entries consumed by the update, so autograd chains the
stiffness gradients back to the physical parameters.
Differences from :class:ElasticTTI/:class:ElasticTTISG (first-order
velocity-stress, 2-D-3C, 8 parameters): two displacement components
only (no SH — no gamma/phi), vh/eta instead of
vp0/delta, and a leapfrog second-order time stepping whose
stiffness operator is discretely self-adjoint.
Models (constructor input order)
vh(m/s): Horizontal P-wave velocity (VTI frame).vs(m/s): Vertical S-wave velocity.rho(kg/m^3): Density.epsilon: Thomsen epsilon.eta: Anellipticity eta = (epsilon - delta) / (1 + 2 delta).theta(rad): Tilt angle of the TI symmetry axis.
Wavefields
ux(aliases:displacement_x): Horizontal displacement; default receiver.uz(aliases:displacement_z): Vertical displacement; default source and receiver.ux_pre: Previous-step horizontal displacement (internal).uz_pre: Previous-step vertical displacement (internal).m_gxux: CPML memory for dux/dx (internal).m_gzux: CPML memory for dux/dz (internal).m_gxuz: CPML memory for duz/dx (internal).m_gzuz: CPML memory for duz/dz (internal).m_sxxx: CPML memory for dsxx/dx (internal).m_sxzz: CPML memory for dsxz/dz (internal).m_sxzx: CPML memory for dsxz/dx (internal).m_szzz: CPML memory for dszz/dz (internal).
Defaults
source_type:['uz']receiver_type:['ux', 'uz']pml_type:'cpmls'
Build the displacement-based 2-D elastic TTI operator.
Parameters:
-
spatial_order–FD accuracy order of the staggered first-derivative operator — e.g.
spatial_order=4is fourth-order accurate. Internally the half-stencil width isM = spatial_order // 2(used for loop bounds and PML padding). Must be an even integer (2, 4, 6, 8, 10, …). Performance note (impl='c'on CUDA): the compiled kernels ship template specialisations only forspatial_order ∈ {2, 4, 6, 8}. Above 8 the dispatcher drops to a generic runtime path (order = -1insrc/sweep/csrc/cuda/equations/elastic_tti_2nd2d/forward.cu) which uses more registers and runs noticeably slower. The PyTorch eager path is unaffected. Defaults to 4. -
device–Device for the operator's static gradient kernels. Use
'cuda'/ atorch.devicefor GPU runs so the propagator can follow without a host↔device copy. Defaults to'cpu'. -
backend–Array / programming backend,
'torch'or'jax'. When you later wantimpl='c', leave this on'torch'. Defaults to'torch'. -
mpml_ratio–Multiaxial-PML mixing ratio (Li & Bou Matar 2010 eq. 23). Plain CPML is unstable for this second-order displacement system — see the module docstring — and this is the knob that fixes it: each damping profile receives this fraction of the other axis'.
0restores the plain CPML and with it the late-time growth (envelope past the direct arrival at 9.6 s, NaN at 19.5 s on a uniform TTI medium). Larger values are safer but absorb worse: measured boundary artefact against a domain no reflection can reach is -56.3 dB at 0 (unstable), -53.2 dB at 0.05, -49.9 dB at 0.10, -44.5 dB at 0.25. Defaults to :data:DEFAULT_MPML_RATIO(0.05).
C_HAS_RECURSIVE_CKPT
class-attribute
¶
bool(x) -> bool
Returns True when the argument x is true, False otherwise. The builtins True and False are the only two instances of the class bool. The class bool is a subclass of the class int, and cannot be subclassed.
C_NAME
class-attribute
¶
str(object='') -> str str(bytes_or_buffer[, encoding[, errors]]) -> str
Create a new string object from the given object. If encoding or errors is specified, then the object must expose a data buffer that will be decoded using the given encoding and error handler. Otherwise, returns the result of object.str() (if defined) or repr(object). encoding defaults to sys.getdefaultencoding(). errors defaults to 'strict'.
prepare_models_for_c
class-attribute
¶
bool(x) -> bool
Returns True when the argument x is true, False otherwise. The builtins True and False are the only two instances of the class bool. The class bool is a subclass of the class int, and cannot be subclassed.
supports_free_surface
class-attribute
¶
bool(x) -> bool
Returns True when the argument x is true, False otherwise. The builtins True and False are the only two instances of the class bool. The class bool is a subclass of the class int, and cannot be subclassed.
takes_invariants
class-attribute
¶
bool(x) -> bool
Returns True when the argument x is true, False otherwise. The builtins True and False are the only two instances of the class bool. The class bool is a subclass of the class int, and cannot be subclassed.
init_abc ¶
Build the CPML profiles, then mix them multiaxially.
The mix has to happen here rather than in step: that is exactly
the property of the multiaxial layer that makes it cheap — the update
never learns about it. The profiles come back as full 2-D fields even
at mpml_ratio=0 so the compiled kernel has a single index path.
prepare_models ¶
(vh, vs, rho, epsilon, eta, theta) -> [rho] + 6 rotated stiffnesses.
VTI constants from Oh et al. (2020) eq. (19) — note C13 uses the
NMO velocity vh^2/(1+2 eta) — then the closed-form 2-D Bond
rotation of their eq. (10). Pure smooth tensor ops: autograd
differentiates every physical parameter, including theta (no
VTI-fallback masking needed — the polynomials are exact at
theta = 0).