acceptodds
Under review as a conference paper at ICLR 2027

Joint 3D Gravity and Magnetic Inversion via Latent Rectified-Flow Posterior Sampling

Abstract

Joint gravity and magnetic inversion, which recovers 3D subsurface density and magnetic susceptibility from surface measurements, is among the cheapest and most widely used tools in mineral exploration, but it is severely non-unique: many subsurface models explain the same data. Classical inversions resolve this ambiguity with hand-crafted regularization, and learning-based inversions are trained on simplified synthetics. However, both return a single model. We instead sample from the posterior under a learned geological prior. We train a 3D variational autoencoder and a latent rectified-flow prior on paired density–susceptibility volumes from Noddyverse, a physics-based collection of one million 3D geological models, and condition the prior on both fields with flow-based posterior sampling at the full resolution. Evaluated against the 3D ground truth of held-out models from largely unseen geological histories, our method reduces susceptibility RMSE by 31–58% relative to twelve established inversions. The learned prior alone already outperforms hand-crafted regularization in susceptibility error. Our samples also counteract the shallow bias of classical inversion: on average, they place bodies within 100 m of their true depth, while the baselines place them 171–961 m too shallow.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Geophysical Inversion, High-dimensional modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4Hi9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies joint 3D gravity and magnetic inversion using a learned geological prior. The proposed pipeline combines a 3D VAE, a rectified-flow prior learned from paired density and susceptibility models, and FlowDPS-based posterior sampling with measurement-consistency guidance. The method additionally introduces a Ginzburg-Landau (GL) energy term to encourage compact, phase-separated susceptibility structures.

Critical strengths
  1. Meaningful application. Joint gravity-magnetic inversion is a practically relevant inverse problem, and learning a joint prior over density and susceptibility provides a natural mechanism for exploiting correlations between the two physical properties.
  2. Learned geological prior. Rather than relying only on handcrafted coupling constraints, the model can learn statistical dependencies between density and susceptibility from paired geological models.
  3. Potential resource value. The paper provides an anonymous repository and emphasizes release of associated resources, which could make the work useful to researchers studying generative approaches for geophysical inversion.
Critical weaknesses
  1. Limited novelty. The overall inference framework appears to be largely an application and integration of existing generative posterior-sampling techniques.
  2. The physics-inspired component is relatively weak. The gravity and magnetic forward models enter through the measurement/data-misfit term. This appears closer to a handcrafted structural/geological regularizer than a substantial integration of new physical principles into the generative model.
  3. The released repository currently lacks sufficient documentation. The paper provides an anonymous repository link for the associated resources. However, based on my inspection of the submitted repository, there is currently no README or clear reproduction documentation describing the released resources and how to reproduce the main experiments.
Questions for authors
  1. Beyond adapting FlowDPS to this problem, what do the authors consider the main methodological contribution of the proposed framework?
  2. Since the release is an important aspect of the contribution, could the authors provide complete documentation and reproduction instructions for the released data, models, and experiments?
Reviewer_RXadOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses joint 3D inversion of gravity and magnetic data — recovering a density–susceptibility volume from surface measurements. This is severely non-unique problem. The proposed method samples from the posterior under a learned geological prior. It trains a 3D VAE and a latent rectified-flow prior on paired density–susceptibility volumes from Noddyverse, then conditions on both fields via flow-based posterior sampling (FlowDPS) at full resolution. Against twelve established inversions on held-out models from largely unseen geological histories, it cuts susceptibility RMSE by 31–58%. The learned prior by itself — sampled with no data — already beats every baseline in susceptibility error.

Critical strengths
  1. The paper makes a sharp, correct point: the purpose of sampling a posterior in potential-field inversion is to expose ambiguity that a single reconstruction hides. That justifies the probabilistic approach on its own terms.

  2. The baselines are competitive. The authors compare against twelve established deterministic inversions: CGLS, LSQR, Tikhonov-CG, projected CGLS, four IRLS focusing variants, isotropic and joint TV, and cross-gradient inversion with a zero-coupling control. All of them use the same forward operator and depth weighting, and their regularization weights were set by data-tolerance or L-curve rules that never use the ground truth. This is a fair comparison.

  3. The authors report that they tried adding a Ginzburg-Landau phase-separation prior at inference time and that it did not measurably help. They discuss this in the main text rather than hiding it. Most papers would drop a negative result. Reporting it makes the argument that learned priors matter stronger, not weaker.

Critical weaknesses
  1. The paper presents prior-only beating all baselines as evidence that a learned prior is superior to hand-crafted regularization. But this result is expected given the setup, not impressive. A generative prior trained on Noddyverse samples from the Noddyverse distribution. The ground truth is also from Noddyverse. So a prior-only sample naturally sits inside the distribution of the test set. It should outperform data-fitting baselines that lack this distributional knowledge.

  2. The prior is a hard support constraint, not a soft one. This is double-edged. A hand-crafted regularizer such as smoothness or sparsity is a weak prior that any solution can satisfy. The learned prior only generates structures in the Noddyverse support. If a real subsurface structure is out-of-distribution, the prior will pull the posterior toward a Noddy-typical form. This is a systematic bias, not a neutral constraint. It is more dangerous than classical inversion being not accurate enough, because a user may read the posterior spread as genuine uncertainty when it only reflects uncertainty within the Noddy prior. The paper does not address this.

  3. The joint contribution is not validated against a separate-inversion control. The forward model is block-diagonal. Gravity responds only to density, and magnetics only to susceptibility. So the joint structure lives entirely in the learned prior over paired density and susceptibility volumes, not in the forward model. Yet the paper never runs the decisive control: is a joint prior over both fields better than two independent single-property priors? Without this, the claim that the coupling is learned from paired geological models is asserted but not shown to drive the gain. This matters because the paper's own results undercut it. On body localization, measured by susceptibility Dice, FlowDPS is on par with cross-gradient inversion, within 0.002. The paper also states that cross-gradient coupling does not improve. So the joint-specific benefit over an existing joint method is weaker than the framing suggests. The RMSE win may come from the strong prior's regularization rather than from the coupling.

  4. Figure 1 does not accurately represent the inference-time optimization. It depicts a single refine-per-step. But Algorithm 1 runs an inner refinement loop of K_ref equals 16 gradient steps at each of the 64 outer time steps. It then blends the prior endpoint estimate and the refined endpoint estimate before propagation. As drawn, it implies a simpler single-pass gradient correction than the algorithm actually performs. This could mislead a reader about the computational cost and the role of the inner loop.

Questions for authors
  1. Please add a control comparing the joint prior over both fields against two independent single-property priors, under otherwise identical settings. Its absence, combined with the cross-gradient results, makes the joint-specific benefit hard to attribute.
Reviewer_yRdqOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies joint 3D inversion of gravity and magnetic data. The goal is to recover subsurface density and magnetic susceptibility from surface measurements. Gravity data depend only on density, and magnetic data only on susceptibility, so a joint inversion needs a link between the two properties. Classical methods impose this link with hand-crafted terms, such as cross-gradient constraints that align the boundaries of the two properties. The proposed method instead learns it from paired geological models. Training data come from Noddyverse, a public set of one million simulated 3D geological models with their surface fields. The proposed method first trains a 3D VAE on the paired density and log-susceptibility volumes and then trains a diffusion-like rectified flow on the VAE latents. To invert the observed fields, it uses FlowDPS, which extends diffusion-style posterior sampling to rectified flows. The experiment results show that the method achieves lower susceptibility RMSE than classical inversions, and that much of this improvement already comes from the learned prior alone, without the data.

Critical strengths
  • The idea of joint inversion is well-motivated. The proposed method achieves this goal by learning the link between the two properties from paired geological models, without any hand-crafted coupling term.
  • The experiments are thorough and the results are promising. The method reduces the bias of classical inversions, and it improves susceptibility RMSE over every baseline on most test models. The paper also reports results that do not favor the method and discusses possible reasons for them.
  • The codes are public
Critical weaknesses
  • My major concern is limited novelty. VAE, the rectified flow, and FlowDPS are all used as in prior work, I do not see new insights other than combining these together.
  • It is unclear whether the joint setting helps. I expected this to be the interesting part. However, the paper does not compare with inverting each field separately using the same approach, or other baselines.
  • The improvement mainly reflects the prior. The prior alone, without any data, already beats every baseline on susceptibility error, and using the data reduces this error only slightly further.
Questions for authors
  • With magnetic data alone, does the recovered density improve over the prior? That would be a direct way to see whether the learned link between the two properties is doing anything.
  • Appendix F reports the VAE’s validation loss, but how do VAE reconstructions of the test models score on the metrics in Table 1? I wonder whether that limits the density results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.