Elastic full-waveform inversion gets a neural field: on Marmousi, Vp MAE falls from 412.02 to 216.90 with fewer crosstalk artifacts
Synopsis
Multiparameter elastic full-waveform inversion estimates several elastic properties at once, and crosstalk between them leaves artifacts in the recovered models. INR-EFWI replaces independent grid-point values with a coordinate neural network that generates the subsurface model continuously, and pairs it with a Vp, P-wave impedance Ip and Vs/Vp parameterization whose radiation patterns overlap less. On the Marmousi and Overthrust synthetic models and the Volve ocean-bottom dataset, it yields lower model error and better structural continuity than automatic-differentiation EFWI, though the authors stress that this parameterization is not universally optimal.
Interpretation
Multiparameter elastic full-waveform inversion now represents the subsurface with a coordinate neural network and inverts for Vp, P-wave impedance Ip and the Vs/Vp ratio, reducing crosstalk artifacts and improving interface continuity in the recovered models. Conventional elastic full-waveform inversion represents each physical parameter independently on a discrete spatial grid, so every grid point is an independent degree of freedom unless correlations are introduced through regularization; that high-dimensional representation favors short-wavelength artifacts and noise and does not intrinsically encode the spatial continuity expected of geological models. On the Marmousi model, starting from the same deliberately simplified linear initial models and running 500 iterations each, INR-EFWI lowers the Vp mean absolute error from 412.02 to 216.90 relative to automatic-differentiation EFWI, raises model signal-to-noise ratio from 12.84 dB to 15.71 dB and structural similarity from 0.32 to 0.53.
Spatial coordinates pass through a shared encoder, and three independent decoder branches output Vp, impedance and Vs/Vp; the waveform residual is backpropagated through the elastic wave-equation solver into the network weights. The shared encoder concatenates a multiresolution hash encoding with a SIREN sinusoidal representation into one set of spatial features; the hash encoder has 16 resolution levels with two features per level and a base resolution of 50, while the SIREN pathway has three sinusoidal layers of widths 256, 256 and 128. The three decoder branches share an architecture but not parameters, each with two hidden layers of 128 and 64 neurons and a linear output layer, with GELU activations between hidden layers; each branch output is rescaled to physical units and added to the initial model, so each parameter's gradient returns only through its own branch.
Perspective
The demonstrated scope is limited to 2D isotropic elastic media: the Marmousi and Overthrust synthetic models, and the Volve ocean-bottom dataset filtered to 5–12 Hz with 48 shots, 197 traces per shot and a 25 m trace interval; the dataset is called ocean-bottom-node (OBN) data in the abstract and introduction but ocean-bottom-cable (OBC) data in Section 3.3, and this report does not resolve that disagreement. The field data have no reliable ground-truth model, so performance is assessed only by how well forward-modeled shot gathers match the observed data.
The authors explicitly do not treat Vp–Ip–Vs/Vp as a universally optimal parameterization; recoverability depends on radiation patterns, scattering angles, acquisition aperture, component availability, wave modes, frequency content, background properties and geological structure. Reconstruction accuracy degrades as expected when frequencies below 5 Hz are missing, and quality also degrades modestly at a data signal-to-noise ratio of 1 dB.