Posterior Sampling Over Topology in 3D Acoustic Shape Inversion

When you bounce a sound wave off a hidden object and record only the intensity of the echo, often from a single direction and a single frequency, many different shapes explain the recording equally well. Some of these shapes differ in ways that matter: a solid lump versus a ring with a hole through it, one body versus two bodies sitting close together. Current 3D methods return a single answer and give no indication that the data could not decide. We propose to return a distribution over shapes. Each shape is stored as a neural signed distance function, a small network whose zero crossing traces the surface, so samples can differ in how many pieces they have or how many holes. The distribution comes from combining a learned shape prior with a physics model of how sound scatters. With it we can report calibrated probabilities for topological events, for example "82% chance this object has a through-hole" or "the two components may be fused," along with a surface uncertainty map.

The pieces exist separately but have not been assembled. One line of work solves physical scattering inverse problems with implicit surfaces and learned priors, but only in 2D and only as a single optimized answer. Another trains diffusion models (generators that learn to turn random noise into samples by denoising it step by step) over implicit 3D shapes, but never couples them to a physical measurement. A third runs 3D acoustic shape inversion, but as one feed-forward prediction of a point cloud. The new element in this proposal is posterior sampling. That is what turns the topological freedom of implicit surfaces from a way to get one answer into a way to measure how uncertain the topology is.

Core Claim

In limited-view, phaseless 3D acoustic scattering, topology is often underdetermined by the data. A sampler that combines a latent diffusion prior over signed distance functions with a differentiable boundary-integral likelihood produces calibrated probabilities for topological features: the number of connected components, and genus, which counts through-holes. Point estimates cannot produce these probabilities. Amortized conditional generators trained for one measurement setup lose calibration when the setup changes.

Relation to Prior Work

The motivating and method-supplying work is Implicit Neural Representation for Mesh-Free Inverse Obstacle Scattering. It represents a 2D obstacle as the zero set of an MLP signed distance function and evaluates the Helmholtz boundary integral over a thin band around that zero set using the implicit boundary integral method. It regularizes with a Real NVP normalizing flow over shape codes. We reuse the implicit boundary integral idea and extend it to 3D. We replace the flow-regularized point estimate with posterior sampling, which that paper leaves open.

Neural Inverse Scattering with Score-based Regularization shows that gradients from a diffusion prior improve neural-field scattering inversion over total variation (SSIM 0.7917 vs 0.6476 on the high-noise Cylinder phantom). It works on 2D permittivity images rather than surfaces, and it runs deterministic descent. That is evidence that score priors help physics inversion, but it says nothing about 3D geometry or uncertainty.

The prior comes from 3D-LDM. It trains a DeepSDF auto-decoder (DeepSDF) and then a diffusion model on the 256-dimensional shape codes, but it never conditions on physics. We use a global-code prior on purpose, even though grid-based priors such as SDFusion and AutoSDF preserve spatial locality better. A low-dimensional code keeps posterior sampling affordable when every step requires a 3D wave-physics gradient. We treat the resulting representation error as something to measure, not assume away.

The closest 3D competitor is ISSRNet. It maps phaseless far-field intensity directly into the 64-dimensional latent space of a PointNet VAE. It is very fast but gives one non-watertight point cloud, and it degrades on out-of-distribution particles. A Neural Network Warm-Start Approach constrains the design: feed-forward prediction alone fails at high frequency, while prediction followed by physics refinement works (1.32% vs 32.31% error for Gauss–Newton at $k=30$). Any competitive method therefore needs a physics-consistency stage, which ours has by construction. Its star-shaped parameterization is exactly what we avoid.

Posterior sampling with learned score priors has precedent only on 2D images: Denoising Score-Matching for Uncertainty Quantification for MRI and Proximal Nested Sampling for radio astronomy. The first reports difficulty annealing the sampler to low temperatures. That warns us that sampler convergence must be checked directly, not assumed.

Research Program

We first build the prior. A DeepSDF auto-decoder with an Eikonal penalty (which pushes the gradient norm of the field toward 1 so it behaves like a true distance) is trained on a shape corpus deliberately rich in topological variation. The corpus is ShapeNet cars, airplanes and chairs, since chair backs and legs create holes and thin handles. We add a synthetic set built from ISSRNet-style spherical-harmonic particles, which we combine by union, subtraction and near-contact placement to produce known genus and component counts. A diffusion model is then trained on the codes, following 3D-LDM.

Second, we build a differentiable 3D forward model. For a code $z$ the surface is $S(z)=\{x: f(z,x)=0\}$. We evaluate the far-field pattern of a sound-soft obstacle with a 3D implicit boundary integral: quadrature over points sampled in a band $|f|<\epsilon$, with normals $\nabla_x f/\|\nabla_x f\|$. The simulated measurement is $A(S(z))$, the phaseless intensity at receiver directions. We validate it against a conventional boundary-element solver on meshes extracted by marching cubes before trusting its gradients.

Third, we sample from

$$ p(z \mid y) \;\propto\; \exp\!\Big(-\tfrac{1}{2\sigma^2}\,\|y - A(S(z))\|^2\Big)\, p_{\text{diff}}(z), $$

where $y$ is the recorded intensity, $\sigma$ the noise level, and $p_{\text{diff}}$ the diffusion prior. We use diffusion posterior sampling, which adds a likelihood-gradient step to each reverse denoising step, and then run a short Langevin refinement at the final noise level. Each sample is meshed. We compute its component count $b_0$ and its genus from the Euler characteristic. We report the fraction of samples showing each topological event, together with a per-point occupancy probability map.

Fourth, we address representation error on out-of-distribution shapes. As an optional stage, each sample's decoder weights can be adapted for a few steps under the likelihood and the Eikonal penalty. We measure whether this changes the topology distribution or only sharpens geometry.

Baselines

The naive reference is physics-only inversion. We optimize a free neural signed distance function initialized as a sphere, using the same forward model with only Eikonal and total-variation-style smoothness penalties. This is the 3D analogue of the unregularized setting in Vlasic et al., and it shows how much any learned prior contributes.

The strongest existing 3D method is ISSRNet (ISSRNet). It reports Chamfer Distance 0.017 on airplanes, 0.0093 on cars and 0.034 on random particles. Because Chamfer normalization differs across pipelines, we retrain it on our exact simulated data and splits, with a matched tuning budget, rather than comparing to the printed numbers. We also train a variant that decodes into our DeepSDF space so the output is watertight. That separates the advantage of a better representation from the advantage of sampling.

The warm-start competitor follows A Neural Network Warm-Start Approach but uses our representation. The ISSRNet-style prediction initializes latent optimization under our forward model. This is a strong physics-consistent point estimate. For uncertainty we give it the fairest cheap treatment: an ensemble of runs from perturbed initializations.

The amortized competitor is a conditional latent diffusion model trained on (measurement, code) pairs, in the style of SDFusion's conditioning. It is given the same training compute as our prior plus sampler.

The key ablation replaces the diffusion prior with DeepSDF's isotropic Gaussian code prior, keeping the same sampler. This isolates whether the learned prior's shape of mass, rather than sampling alone, is what yields calibrated topology.

Beating a baseline means one of two things. Either we match its Chamfer and IoU within noise while achieving lower expected calibration error on topological events, or we win on geometry under measurement shift. There is no reported result on Lacuna for calibrated topology prediction in 3D scattering, so those numbers will be new.

Why not just…?

Why not just train ISSRNet with an implicit decoder and read off the answer? That fixes watertightness but still gives one answer. On ambiguous inputs it will predict the conditional average in latent space, which can be a topologically meaningless blend, for example a half-closed hole. ISSRNet's own errors on out-of-distribution particles and rounded roof corners suggest it cannot fix itself without physics. The retrained implicit-decoder variant in our baselines tests this directly.

Why not just run the warm-started optimizer from many initializations? This is a genuine threat. Multi-start optimization can find several modes cheaply. However, the fraction of runs landing in a mode reflects the size of that mode's basin of attraction, not its posterior probability, so there is no reason for it to be calibrated. Our calibration experiment reports expected calibration error for both methods at an equal count of forward-model evaluations. If the ensemble matches us, we will report that the diffusion prior is unnecessary for topology uncertainty.

Why not just train a conditional diffusion model on measurement–shape pairs? In distribution it may match us at a fraction of the test-time cost, and we expect that. The weakness is that it bakes in the aperture, frequency and noise level used in training. Our prior never sees measurements, so changing the sensor setup only changes the likelihood. The shift experiment below settles which approach is preferable, and we will state the in-distribution result plainly.

Why not just use total variation or Eikonal smoothness? NF-Score's comparison and the 1.2 dB gain in proximal nested sampling both indicate that hand-crafted priors underperform learned ones in scattering and imaging. More to the point, smoothness penalties give no mechanism to prefer one topology over another. The physics-only baseline quantifies this.

Proposed Experiments

We simulate phaseless far-field intensity from a single incident wave, matching ISSRNet's setup. We also run a multi-incidence full-aperture setting and a limited-aperture setting covering 90°. Noise is added at 30, 20 and 10 dB. Geometry is scored by Chamfer Distance, volumetric IoU, F-score and data residual. Topology is scored by accuracy of $b_0$ and genus for the most probable sample, and by expected calibration error of predicted event probabilities against ground truth across the test set.

The first decisive check uses constructed ambiguity pairs. These are synthetic shapes that differ in topology (a thin bridge present or absent, two particles touching or separated by a small gap) and whose simulated measurements differ by less than the noise level. On such inputs a correct posterior should split its mass between the two topologies. We check whether our sampler does this and whether each baseline collapses to one mode. If our sampler also collapses, the central claim fails.

The second check is measurement shift. The conditional diffusion baseline and ISSRNet are trained at full aperture and 30 dB. All methods are then tested at limited aperture, at a different frequency and at 10 dB. We expect our method's calibration to hold while the amortized methods degrade.

The third check validates the sampler. On a low-dimensional synthetic family where exact posterior mass can be computed by brute-force enumeration over the code grid, we compare the sampled topology probabilities to the exact ones. This follows the caution in Denoising Score-Matching for Uncertainty Quantification that annealed samplers can stall.

Finally, we report wall-clock cost per posterior sample against ISSRNet's 0.056 seconds per object. We state plainly the regimes in which the price of sampling is worth paying.