arxiv: 2604.26938 · v2 · submitted 2026-04-29 · ⚛️ physics.geo-ph

Recognition: unknown

Meta-learning-enhanced implicit full waveform inversion

Zefeng Wang , Shijun Cheng , Weijian Mao , Wei Ouyang , Huanhuan Tang

Authors on Pith no claims yet

Pith reviewed 2026-05-07 08:54 UTC · model grok-4.3

classification ⚛️ physics.geo-ph

keywords meta-learningfull waveform inversionimplicit neural representationsseismic velocity modelSIRENgeophysical inversiongeneralizationconvergence acceleration

0 comments

The pith

Meta-learning lets an implicit neural network adapt to new seismic inversion tasks after pretraining on several velocity models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes Meta-IFWI to address slow convergence and lack of knowledge sharing in implicit full waveform inversion. It represents the subsurface velocity as a continuous function via a neural network and uses meta-learning to pretrain that network across multiple inversion tasks so it acquires shared priors and fast adaptation rules. Once pretrained, the network requires only a small number of gradient steps on new seismic data to produce an accurate velocity model. Experiments on layered synthetics, the Overthrust model, and the out-of-distribution Marmousi 2 model show higher accuracy, fewer iterations, and lower cost than standard implicit inversion. A reader would care because the approach could make full-waveform methods practical for routine exploration by cutting per-area computation and improving reliability across geological settings.

Core claim

By pretraining a single implicit neural network with periodic activations on multiple velocity-inversion tasks, the network learns shared inversion priors and rapid adaptation strategies; for any new task the pretrained network can be fine-tuned to the observed seismic data with only a few gradient updates, yielding improved accuracy, faster convergence, and stronger cross-area generalization than conventional implicit full waveform inversion.

What carries the argument

A meta-learning strategy applied to an implicit neural network with periodic activation functions that parameterizes the subsurface velocity model as a continuous function of spatial coordinates, enabling the network to acquire task-shared priors during pretraining.

If this is right

Inversion accuracy improves on both in-distribution models such as layered synthetics and the Overthrust model and on out-of-distribution models such as Marmousi 2.
Convergence occurs in substantially fewer iterations than conventional implicit full waveform inversion.
Computational cost per inversion task decreases because fewer gradient updates are required after pretraining.
The method shows greater robustness and stronger generalization across different geological areas.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same meta-learning pattern could be tested on other inverse problems in geophysics that currently optimize a separate model for each data set.
If the learned priors remain effective on real data, field workflows might shift from building custom initial models to light adaptation of a shared meta-network.
Extending the pretraining distribution to include more varied synthetic and field-derived models would test how far the generalization benefit can be pushed.

Load-bearing premise

Priors learned from a limited collection of synthetic velocity models will transfer to real field data and to geological structures substantially different from the training set without large loss of performance.

What would settle it

Apply the pretrained Meta-IFWI network to a real field seismic dataset from a previously unseen geological province and compare the final model misfit and iteration count against a standard implicit inversion run from the same starting point.

Figures

Figures reproduced from arXiv: 2604.26938 by Huanhuan Tang, Shijun Cheng, Weijian Mao, Wei Ouyang, Zefeng Wang.

**Figure 1.** Figure 1: Layered model inversion results. (a) represents the true velocity model, (b) shows the IFWI result, and (c) view at source ↗

**Figure 2.** Figure 2: Comparison of velocities at (a) 0.5 km, (b) 1.5 km, and (c) 2.5 km in the layered model. The black solid line view at source ↗

**Figure 3.** Figure 3: Comparison of convergence curves for the layered model. (a) Data misfit convergence and (b) velocity model view at source ↗

**Figure 4.** Figure 4: Overthrust model inversion results.(a) represents the true velocity model, (b) shows the IFWI result, and (c) view at source ↗

**Figure 5.** Figure 5: Comparison of velocities at (a) 1.0 km, (b) 2.0 km, and (c) 3.5 km in the Overthrust model. The black solid view at source ↗

**Figure 6.** Figure 6: Comparison of convergence curves for the Overthrust model. (a) Data misfit convergence and (b) velocity view at source ↗

**Figure 7.** Figure 7: Marmousi 2 model inversion results.(a) represents the true velocity model, (b) shows the IFWI result, and (c) view at source ↗

**Figure 8.** Figure 8: Comparison of velocities at (a) 1.5 km, (b) 2.0 km, and (c) 2.5 km in the Marmousi 2 model. The black solid view at source ↗

**Figure 9.** Figure 9: Comparison of convergence curves for the Marmousi 2 model. (a) Data misfit convergence and (b) velocity view at source ↗

read the original abstract

Implicit full waveform inversion (IFWI) introduces implicit neural representations to parameterize the subsurface velocity model as a continuous function of spatial coordinates, which alleviates the dependence on the initial model and improves inversion flexibility. However, IFWI still requires a large number of iterative updates for each new exploration area, leading to slow convergence, high computational cost, and a lack of mechanisms to share prior knowledge across different geological settings, thereby limiting its efficiency and generalization capability. To further accelerate convergence and enhance cross-area generalization, we propose a meta-learning-based implicit full waveform inversion method, referred to as Meta-learning-enhanced implicit full waveform inversion (Meta-IFWI). In this framework, the subsurface velocity model is represented using an implicit neural network with periodic activation functions (SIREN), while a meta-learning strategy is employed to pretrain a single network on multiple velocity inversion tasks. Through this process, the network learns shared inversion priors and rapid adaptation strategies across different geological scenarios. For a new inversion task, the pretrained Meta-IFWI model can be efficiently adapted to the observed seismic data with only a few gradient updates, significantly reducing the number of iterations required for inversion. Numerical experiments conducted on in-distribution models, including layered synthetic models and the Overthrust model, as well as out-of-distribution complex models such as Marmousi 2, demonstrate that, compared with conventional IFWI, the proposed Meta-IFWI achieves improved inversion accuracy while substantially accelerating convergence and reducing computational cost. Moreover, Meta-IFWI exhibits enhanced robustness and stronger cross-area generalization capability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Meta-learning a SIREN network across synthetic velocity tasks lets Meta-IFWI adapt with fewer steps than plain implicit FWI, but only on the synthetics they tested.

read the letter

The paper's main contribution is showing that meta-learning can cut the iteration count for implicit full waveform inversion when the velocity model is represented as a SIREN. They pretrain one network on several inversion tasks so that a new task needs only a small number of gradient updates to match the seismic data. On the layered synthetics, Overthrust, and Marmousi 2, the adapted model reaches higher accuracy with less compute than a fresh IFWI run each time. That is a clear, practical difference inside the synthetic setting they chose. The experiments directly compare in-distribution and one out-of-distribution case, which supports the generalization claim within that regime. The method itself is straightforward: shared priors from meta-training plus quick fine-tuning. No circularity or internal contradiction shows up in the setup. The soft spot is that every result stays inside synthetic velocity models. The abstract and experiments do not include field data, so we still do not know whether the learned priors survive real noise, irregular acquisition, or geological structures far from the training distribution. Details on the exact number of meta-training tasks and inner-loop steps are also thin in the high-level description, though the full text presumably fills them in. This is useful reading for geophysicists who already work with neural representations for seismic inversion and want to reduce per-survey compute. A reader looking for empirical evidence that meta-learning helps FWI convergence on standard benchmarks will find the numbers worth checking. The work is coherent enough on its own terms to deserve referee time rather than a desk reject. I would send it out, with the note that reviewers will almost certainly ask for real-data tests or a clearer discussion of the synthetic-to-field gap.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Meta-IFWI, which parameterizes the subsurface velocity model via a SIREN implicit neural network and uses meta-learning to pretrain the network across multiple synthetic velocity inversion tasks. This allows the model to learn shared priors and rapid adaptation strategies, so that a new task can be solved with only a few gradient updates. Numerical experiments on in-distribution models (layered synthetics, Overthrust) and an out-of-distribution model (Marmousi 2) are presented to claim higher accuracy, faster convergence, lower computational cost, and stronger cross-area generalization relative to conventional IFWI.

Significance. If the experimental claims are substantiated, the work would constitute a useful advance in accelerating full-waveform inversion by transferring meta-learned priors across geological settings. The combination of implicit representations with meta-learning is timely, and the explicit out-of-distribution test on Marmousi 2 provides a concrete check on generalization within the synthetic regime. The approach directly addresses the per-task optimization burden that currently limits IFWI scalability.

major comments (2)

[Numerical Experiments] Numerical Experiments section: the central claims of improved accuracy, substantially accelerated convergence, and reduced cost are not supported by any tabulated quantitative metrics (e.g., RMSE or SSIM of the recovered velocity models, or iteration counts to reach a target misfit). Without these numbers the magnitude of the reported gains cannot be assessed and the comparison to conventional IFWI remains qualitative.
[Method] Method section: the meta-training protocol (number of tasks, distribution of training velocity models, inner-loop adaptation steps, and meta-optimizer hyperparameters) is described only at a high level. These quantities are load-bearing for the rapid-adaptation claim; their omission prevents reproduction and makes it impossible to judge whether the reported performance is robust or sensitive to the particular meta-training choices.

minor comments (2)

[Abstract] Abstract: the summary paragraph would be strengthened by inserting at least one concrete quantitative result (e.g., “reduces iterations by a factor of X on the Overthrust model”) to anchor the qualitative claims.
[Figures] Figure captions: several captions are terse and do not state the inversion parameters or the number of meta-adaptation steps used for the displayed results, reducing clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments highlight important areas for improvement in substantiating our claims and ensuring reproducibility. We address each major comment below and have revised the manuscript accordingly.

read point-by-point responses

Referee: [Numerical Experiments] Numerical Experiments section: the central claims of improved accuracy, substantially accelerated convergence, and reduced cost are not supported by any tabulated quantitative metrics (e.g., RMSE or SSIM of the recovered velocity models, or iteration counts to reach a target misfit). Without these numbers the magnitude of the reported gains cannot be assessed and the comparison to conventional IFWI remains qualitative.

Authors: We agree that the absence of tabulated quantitative metrics leaves the magnitude of the improvements qualitative. In the revised manuscript we have added a new table (Table 2) in the Numerical Experiments section reporting RMSE and SSIM values for all recovered velocity models (layered synthetics, Overthrust, and Marmousi 2) for both Meta-IFWI and conventional IFWI. We have also included a supplementary table listing the iteration counts required to reach successive misfit thresholds (e.g., 10^{-3} and 10^{-4}) together with wall-clock times. These additions provide direct numerical evidence for the claimed gains in accuracy, convergence speed, and computational cost. revision: yes
Referee: [Method] Method section: the meta-training protocol (number of tasks, distribution of training velocity models, inner-loop adaptation steps, and meta-optimizer hyperparameters) is described only at a high level. These quantities are load-bearing for the rapid-adaptation claim; their omission prevents reproduction and makes it impossible to judge whether the reported performance is robust or sensitive to the particular meta-training choices.

Authors: We acknowledge that the meta-training protocol was presented at an insufficient level of detail. The revised Method section now contains a new subsection (Section 3.3) that explicitly states: (i) the number of meta-training tasks (50), (ii) the sampling distribution used to generate the training velocity models (randomized layered and faulted models with velocity ranges 1500–4500 m/s), (iii) the inner-loop adaptation steps (5 steps per task), and (iv) the meta-optimizer hyperparameters (MAML with Adam, outer-loop learning rate 1e-4, inner-loop learning rate 1e-3). These specifications enable full reproduction and allow readers to evaluate sensitivity to the chosen protocol. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes an empirical meta-learning procedure that pretrains a SIREN-based implicit network across multiple synthetic velocity inversion tasks and then adapts the network to new tasks with few gradient steps. The central claims rest on direct numerical comparisons (accuracy, convergence speed, cost, robustness) between Meta-IFWI and standard IFWI on listed synthetic benchmarks (layered models, Overthrust in-distribution; Marmousi 2 out-of-distribution). No equations or derivations are presented that reduce to their own inputs by construction, no uniqueness theorems are imported via self-citation, and no fitted quantities are relabeled as predictions. The reported gains are therefore falsifiable experimental outcomes rather than tautological restatements of the training procedure.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the standard assumption that SIREN networks can faithfully represent subsurface velocity fields and on the meta-learning premise that shared priors across synthetic tasks transfer to new tasks. No explicit free parameters or invented entities are named in the abstract.

free parameters (1)

Meta-training task count and inner-loop adaptation steps
These control the meta-learning process but are not quantified in the abstract.

axioms (1)

domain assumption Implicit neural representations with periodic activations can accurately parameterize subsurface velocity models as continuous functions.
This underpins the IFWI component of the method.

pith-pipeline@v0.9.0 · 5585 in / 1295 out tokens · 63716 ms · 2026-05-07T08:54:16.120226+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 2 canonical work pages

[1]

A strategy for nonlinear elastic inversion of seismic reflection data.Geophysics, 51(10):1893–1903,

Albert Tarantola. A strategy for nonlinear elastic inversion of seismic reflection data.Geophysics, 51(10):1893–1903,

1903
[2]

Implicit full waveform inversion with energy-weighted gradient

S Wang and T Alkhalifah. Implicit full waveform inversion with energy-weighted gradient. In86th EAGE Annual Conference & Exhibition, volume 2025, pages 1–5. European Association of Geoscientists & Engineers,

2025
[3]

16 Meta-learning-enhanced implicit full waveform inversionA PREPRINT Diederik P Kingma and Jimmy Ba

URLhttps://doi.org/10.5281/zenodo.3829886. 16 Meta-learning-enhanced implicit full waveform inversionA PREPRINT Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

work page doi:10.5281/zenodo.3829886
[4]

On First-Order Meta-Learning Algorithms

Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm.arXiv preprint arXiv:1803.02999, 2(3):4,

work page Pith review arXiv