Skip to main content
Reading viewAll insights →
RESEARCH7 min read · 19 Nov 2026

Physics-Informed Neural Networks for the Elastic Wave Equation

A PINN fits a wavefield by minimising a weighted sum of data misfit and PDE residual, and one scalar sets the exchange rate between them. That scalar is not a hyperparameter to inherit from a paper. The usable basin is narrow, it moves with noise and receiver coverage, and a weight calibrated on clean dense data can be a factor of 30 too small on noisy sparse data while the field still looks perfectly smooth.

Tannistha Maitiby Tannistha Maiti · Head of Engineering
subsurface-aiphysics-informed-mlseismic-imaginggeophysicsuncertainty-quantification

The elastic wave equation is the forward model behind every synthetic seismogram: given a density and a stiffness tensor, it maps a source to the displacement field recorded at receivers. Classical solvers discretise space and time and march the field forward, and they are the right tool when the medium is known and the problem is forward.

The inverse problem is harder, and it is where a physics-informed neural network becomes interesting [1,2]. Represent the displacement field by a network. Impose two losses: a data term measuring mismatch against the recorded traces, and a physics term evaluating the strong-form residual of the wave equation at interior collocation points, with the derivatives supplied by automatic differentiation rather than by a stencil or a mesh. Train. The network is the solution; training is the solve.

What makes it work is also what makes it fragile. The physics is a soft constraint, and the strength of a soft constraint is a free parameter.

One scalar decides whether the output is physical

The total loss is the data misfit plus lambda times the PDE residual, and lambda sets the exchange rate between them. The two terms have different curvatures, different gradient magnitudes and different sensitivities to the network's spectral bias [4]. Wavefields are oscillatory and networks are slow to fit high frequencies, so the residual for a wave equation is genuinely hard to drive down without help. Balancing it against the data term is the entire method.

At small lambda the network is a flexible interpolator. It fits the recorded trace, including its noise, and leaves a large residual: a curve through the samples that does something arbitrary between them and is not a physical wavefield. At large lambda the residual collapses and the field is smooth and wave-equation-consistent, but the data term is outvoted and the prediction pulls away from the recorded amplitudes.

Neither end is the answer, which is well known. What is less often stated is where the answer actually sits, and what moves it.

The basin is narrow, and it moves

Physics weight against conditions
The usable weight is a window, and the conditions move it.
1 step outfrom the usable basin, for the 0.1 a published setup hands you
A block diagram of nine stations on a plinth. Noise on the recorded traces runs left to right, from clean to noisy; receiver coverage runs front to back, from dense to sparse. Up a post at each station runs the physics weight lambda on a log scale, 0.01 at the bottom to 30 at the top, one bead for each of 8 weights. A glass tablet on each post is the usable basin there, the weights for which both the held-out misfit and the PDE residual are inside tolerance. An amber pane at the inherited weight, 0.1, is level across the whole map. At clean traces and dense receivers the basin is 0.1, 0.3, 1 and the pane passes through it; at noisy traces and sparse receivers it is 3, 10, 30 and the bed has climbed clear of the pane. Selected: ordinary field noise, standard coverage. The usable basin is 0.3, 1, 3. The inherited weight 0.1 is 1 step outside it: the nearest usable weight is 3 times the inherited value.
Usable basin here
0.3, 1, 3
Held-out misfit at 0.1
0.328 against 0.34
PDE residual at 0.1
0.475 against 0.4
To reach the basin
1 step, 3 times the inherited value
Map: noise left to right, clean to noisy. Receivers front to back, dense to sparse. Up a post: the weight, log scale.usable basin: both diagnostics inside toleranceinherited weight, 0.1steps from the inherited weight to the basin, read on the mastone bead per weight in the sweep
Both diagnostics across the sweep, at ordinary field noise and standard coverage. A weight is usable only when both are inside tolerance.
lambdaheld-out data misfitPDE residualusable
0.010.9441.25no
0.030.5950.788no
0.1 (the inherited value)0.3280.475no
0.30.1890.3yes
10.1520.181yes
30.2230.114yes
100.4160.069no
300.6970.043no
A demonstration of the mechanism our ES-3110 note describes, not a training run: no network was fitted and every quantity is dimensionless. Tolerances are 0.34 for the held-out misfit and 0.4 for the residual. Height in the drawing is the weight, not the defect.
The physics weight sweep for a PINN on the elastic wave equation, across every combination of trace noise and receiver coverage. Each of the nine posts is one set of conditions and height up it is the weight lambda on a log scale, from 0.01 to 30, one bead for each value swept. The glass tablet on a post is the usable basin there, the weights for which both the held-out data misfit and the PDE residual are inside tolerance. The amber pane at 0.1 is the weight a reader inherits from a published setup, and it is level across the whole map. Calibrated on clean traces with dense receivers the basin includes 0.1. Switch to noisy traces and sparse coverage and it has slid to 3, 10 and 30, leaving the inherited value three steps outside it, a factor of 30 too small. Split the two diagnostics to see each tablet pulled apart into the band the misfit allows and the half-line the residual allows. A demonstration of the mechanism our ES-3110 note describes, not a training run. Drag to turn, pinch or hold Ctrl and scroll to zoom, double-click to reset.

Each of the nine posts in the block is one set of conditions, and height up a post is lambda over more than three orders of magnitude. The glass tablet on a post is the usable basin there, and a lambda is usable only when both diagnostics are inside tolerance: the held-out data misfit and the residual on a held-out collocation grid. Either alone is easy to satisfy and neither alone means anything. The amber pane at 0.1 is the weight a reader inherits, and it is level across every set of conditions.

Start on clean traces with dense receivers, which is the setting a published demonstration is most likely to have run under. The basin includes 0.1. Now switch the noise to field level and thin the receivers to sparse. The basin has slid to 3, 10 and 30. The inherited value of 0.1 is three steps outside it, a factor of 30 too small, with the held-out misfit well past tolerance and the PDE residual, at 0.57 against a tolerance of 0.4, past it as well.

Nothing about the network changed. No architecture, no data, no optimiser. The conditions changed, and the conditions are what set where the exchange rate has to sit. Noisier traces need the physics weighted harder, because more of what the data term is fitting is noise. Sparser receivers need it harder too, because there is less data holding the field between sensors.

The practical consequence is a reproducibility one. A PINN reported without its lambda is under-specified, and two groups can run what they both describe as the same PINN on the same elastic wave problem and obtain materially different fields purely because one weighted the physics an order of magnitude harder. The good news is that this is structural rather than mysterious: lambda can be calibrated against held-out residual and held-out misfit, and a defensible protocol schedules or adaptively weights it, annealing the physics term up as training proceeds or normalising the two gradients against each other [4], so the network finds the data first and is then pulled onto the physics.

The failure that looks like success

The reason to care about this specifically, rather than filing it with other hyperparameters, is what the failure looks like from the outside.

A lambda that is too small produces a smooth, plausible, well-fitting curve through every recorded sample. It will pass a visual check. It will report a low data misfit on the receivers it trained on. The thing it is wrong about is the field between the receivers, which is the only reason anyone built it. An unexamined regularisation dial is where silent failure hides, and here the dial is the one that decides whether the output obeys physics at all.

Two limits worth stating

Spectral bias is the first. A plain PINN struggles with high-frequency wavefields, so realistic broadband seismograms may need Fourier-feature encodings or domain decomposition before the residual can be driven down at all [3,5]. The lambda trade-off sits on top of that harder problem rather than instead of it.

The strong-form residual is the second. It assumes the medium is smooth enough to differentiate, and sharp velocity contrasts, a Moho or a sediment-basement interface, are exactly where the coefficients jump. A naive collocation PINN will fight those discontinuities.

Both are reasons to treat the PINN as a differentiable, data-assimilating complement to classical solvers rather than as a replacement for them. The value is real where it is real: a mesh-free differentiable field you can push data into and pull gradients out of, provided its physics weight is treated as a quantity to calibrate rather than a constant to inherit.

Key takeaways

  1. A PINN's entire behaviour on the elastic wave equation is governed by one scalar, the weight between the data misfit and the PDE residual.
  2. At small weights the network is an interpolator that fits the noise and violates the wave equation between receivers; at large weights the field is physical and ignores the data.
  3. The usable basin is narrow and it moves: calibrated on clean traces with dense receivers it includes 0.1, and on noisy traces with sparse coverage it has slid to 3, 10 and 30.
  4. A weight inherited from a published setup can therefore be a factor of 30 too small for your data, with nothing about the network changed.
  5. A PINN reported without its physics weight is under-specified, and the weight should be calibrated against held-out residual and held-out misfit, or scheduled and gradient-normalised.
  6. The failure mode looks like success: a smooth curve through every recorded sample that is wrong about the field between them, which is the only part anyone built it for.

References

  1. M. Raissi, P. Perdikaris, G. E. Karniadakis. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686 to 707, 2019. doi:10.1016/j.jcp.2018.10.045
  2. G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang. Physics-informed machine learning. Nature Reviews Physics, 3, 422 to 440, 2021. doi:10.1038/s42254-021-00314-5
  3. B. Moseley, A. Markham, T. Nissen-Meyer. Solving the wave equation with physics-informed deep learning. 2020. arXiv:2006.11894
  4. S. Wang, Y. Teng, P. Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing, 43, A3055 to A3081, 2021. doi:10.1137/20M1318043
  5. M. Rasht-Behesht, C. Huber, K. Shukla, G. E. Karniadakis. Physics-informed neural networks (PINNs) for wave propagation and full waveform inversions. Journal of Geophysical Research: Solid Earth, 127, e2021JB023120, 2022. doi:10.1029/2021JB023120
  6. Our earlier piece on the wave equation as an unsupervised label: Physics-informed neural networks for seismic noise attenuation.
Tannistha Maiti
Tannistha Maiti

Head of Engineering

More from EarthScan

Related research

All insights →
Below Tuning, Thickness and Contrast Are the Same Number
Insight

Below Tuning, Thickness and Contrast Are the Same Number

Physics-Informed Neural Networks for Seismic Noise Attenuation: When the Wave Equation Is Your Only Label
Research

Physics-Informed Neural Networks for Seismic Noise Attenuation: When the Wave Equation Is Your Only Label

Amortized Bayesian Inversion of Crustal Structure
Insight

Amortized Bayesian Inversion of Crustal Structure

Stay ahead

EarthScan insights, in your inbox.

Field-tested research on subsurface and energy-transition AI. About twice a month. No noise.

We use your email only for this newsletter. Unsubscribe anytime Privacy.