PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers

AI summary

Overview

Research area: Scientific machine learning, Bayesian inverse problems, and generative modeling — specifically the evaluation of diffusion- and flow-based samplers for physics-constrained inverse problems.

Technical level: Advanced. The paper assumes familiarity with Bayesian inference, diffusion models, neural operators, and PDE-constrained inverse problems, though the central argument (evaluate distributions, not single reconstructions) is broadly accessible.

Scope: The work introduces PosteriorBench, a benchmark that pairs four scientific inverse problems with high-fidelity reference posteriors and a five-metric distributional evaluation suite, then uses it to compare eight probabilistic inverse solvers.

What This Paper Is About

Scientific inverse problems — inferring an unknown physical field from sparse or noisy indirect measurements — are fundamentally ill-posed: many different fields can explain the same observations. Generative models are increasingly used to sample from the resulting posterior distribution, but the standard way of evaluating them asks only whether a single reconstructed field is accurate. This paper argues that such pointwise evaluation is misleading and builds a benchmark that instead measures whether a solver recovers the full posterior distribution, including its spread, shape, and spectral content.

Key Contributions

1. A distribution-centered evaluation protocol. The paper defines a benchmark for generative scientific inverse solvers that explicitly constructs, validates, and compares against reference posterior distributions rather than single ground-truth fields, shifting evaluation from point estimates to posterior matching.

2. Four diverse, physics-based inverse tasks under one framework. Darcy flow inversion (binary permeability), Poisson source recovery (smooth random fields), carbon capture and storage (geostatistical permeability from well columns), and light transport material inference (two-layer radiative medium) span different priors, observation patterns, forward maps, and ambiguity sources.

3. A five-metric posterior evaluation suite. Posterior-mean error, posterior-standard-deviation error, maximum mean discrepancy (MMD), sliced Wasserstein distance (SWD), and radially averaged power-spectrum (RAPS) error jointly assess pointwise accuracy, marginal uncertainty, distributional alignment, and global frequency fidelity.

4. Systematic empirical findings and ablations. The authors benchmark eight solvers and report distribution-matching gaps, a non-monotone tradeoff between pointwise error and distributional fidelity, out-of-distribution prior-coverage effects, guidance-weight calibration behavior, and resolution-robustness results.

Main Findings

• Function-space diffusion samplers are strong posterior samplers, but task-dependent. DDIS and FunDPS (and related guided diffusion methods built on function-space score priors) lead on Darcy flow and Poisson source recovery, while the classical ensemble data-assimilation baseline ES-MDA remains competitive or best on light transport material inference. No single method dominates across all tasks and metrics.

• Posterior-mean error alone is misleading. On the LTMI task, FNO with MC Dropout achieves lower posterior-mean error than FunDPS, yet its posterior-standard-deviation error, MMD, and SWD are substantially worse. Visual inspection shows the MC Dropout sample is over-smoothed relative to the reference structure, demonstrating that a low mean error can coincide with a poorly captured posterior.

• Pointwise and distributional metrics are not monotone. Comparing a traditional pointwise relative L2 metric against the five posterior metrics produces V-shaped trends: pushing samples toward a single reference field can improve pointwise error while degrading posterior variance and distributional alignment. Pointwise metrics alone can therefore misidentify over-fitted or over-concentrated solvers.

• Solvers systematically underestimate posterior variance. Error maps for both Poisson source recovery and Darcy flow inversion show predominantly negative standard-deviation errors, and posterior means pull extreme source values toward zero — indicating conservative, over-smoothed predictions that miss the true spread of the posterior.

• Broader prior coverage does not guarantee better posterior matching. In an out-of-distribution experiment on Poisson source recovery, training the Fun-DDPS prior on a wider Gaussian-random-field smoothness range gave the lowest pointwise L2 error but worsened std error, MMD, and SWD. Diagnostics showed more frequent PDE residual violations and biased recovery of the latent smoothness parameter in the broader-coverage setting.

• Optimal guidance weights track observation noise but are metric-specific. On Poisson source recovery, the guidance weight that minimizes each metric generally increases with inverse observation-noise variance, consistent with Gaussian likelihood theory. However, the mean-optimal and variance-optimal guidance weights differ at the same noise level, showing that a single scalar guidance coefficient trades off posterior-mean accuracy against uncertainty calibration.

• Function-space training improves resolution robustness. In a resolution ablation at 128×128, FunDPS consistently outperformed DiffusionPDE across training settings on mean error, standard-deviation error, MMD, SWD, and spectral metrics, with the best distributional alignment under multi-resolution training.

• Reference posteriors are expensive but reliable. Each benchmark case is paired with a reference posterior built via rejection sampling or Markov chain Monte Carlo, retaining roughly 100 weighted samples within a 3σ mismatch threshold and validated before use as an evaluation target.

Methodology in Plain English

The authors begin from Bayes' rule: given observations, the scientifically correct answer is a probability distribution over possible fields, not one field. To build a fair test, they design four inverse problems where this ambiguity is real and known:

• Darcy flow inversion: recover a binary permeability field from sparse pressure readings; several channel-like phase layouts can produce similar pressures.
• Poisson source recovery: recover a smooth source term from sparse potential measurements; the prior varies in correlation length and smoothness across cases.
• Carbon capture and storage: recover a subsurface permeability field from sparse well-column observations of CO2 saturation, using geostatistical priors and a reservoir simulator.
• Light transport material inference: recover two-layer optical material properties from low-resolution reflectance or transmittance measurements.

For each case, they generate a reference posterior by drawing a large candidate pool from the prior, simulating the forward physics for each candidate, and weighting candidates by how well they match the observations under an assumed Gaussian noise model. This is computationally expensive but transparent, and it gives an empirical target distribution to compare against.

They then run eight solvers on every task. Each solver returns an ensemble of samples, which is compared to the weighted reference ensemble using five metrics: mean error and standard-deviation error (marginal moment consistency), MMD with a multi-scale RBF kernel (higher-order spatial statistics), SWD after z-score normalization and projection onto smooth Gaussian-random-field directions (geometric alignment), and RAPS error (energy distribution across spatial frequencies).

Finally, they run ablations: varying observation noise and guidance strength to check calibration, varying the prior's training coverage to test out-of-distribution generalization, and varying resolution to test train–test robustness. They also compare all posterior metrics against a conventional pointwise relative L2 metric to probe whether the two families agree.

Why This Matters

Impact on research. The paper reframes how generative inverse solvers should be judged. It shows that a solver can be state-of-the-art on a pointwise metric while misrepresenting uncertainty, mode structure, or spatial texture — a failure mode that matters directly in scientific applications where posterior uncertainty drives decisions. By releasing reference posteriors, metrics, and code, it gives the field a common distributional yardstick and a reproducible pipeline for evaluating new methods.

Real-world applications:

• Subsurface and reservoir characterization: Groundwater management, oil and gas reservoir modeling, and geothermal exploration all depend on inferring permeability fields from sparse pressure or production data — exactly the ambiguity PosteriorBench measures.

• Carbon capture and storage monitoring: Inferring subsurface heterogeneity from a small number of monitoring wells directly affects forecasts of CO2 plume migration, pressure buildup, and leakage risk. Reducing required wells lowers field intervention and monitoring cost, but only if uncertainty is honestly quantified.

• Atmospheric and remote sensing retrieval: Recovering cloud and aerosol structure, and optical properties of layered media, from limited-viewpoint measurements is a light-transport inverse problem of the same form as the LTMI task.

• Biomedical and industrial imaging: Diffuse optical tomography of layered tissue and nondestructive optical inspection of semitransparent materials both require inferring internal material structure from low-resolution optical measurements.

Industry relevance. Uncertainty quantification is a prerequisite for regulatory acceptance and risk-based decision-making in energy, environmental monitoring, and medical imaging. The finding that current samplers systematically underestimate posterior variance is a direct warning that downstream risk assessments built on these methods may be overconfident. The benchmark also gives practitioners a principled way to choose between fast classical data assimilation and slower generative samplers per task.

Future Directions

• Conditioning mechanisms that jointly calibrate mean and uncertainty. Because the mean-optimal and variance-optimal guidance weights diverge, the paper motivates conditioning schemes that adjust posterior center and spread together rather than through a single scalar guidance coefficient.

• Closing the variance-calibration gap. Current solvers consistently underestimate posterior standard deviation, especially in high-gradient and high-uncertainty regions. Methods that better capture posterior tails and extremes remain an open problem.

• Latent-parameter and physical-consistency recovery. The out-of-distribution results show that plausible pixel-space samples can hide poor recovery of latent parameters (such as GRF smoothness) and increased PDE violations, suggesting a need for training objectives and evaluation that directly target latent structure and physical consistency.

• Broadening and scaling the benchmark. Extending PosteriorBench to more forward models, noise models, and observation geometries, and testing whether function-space priors retain their advantage at larger scales and higher dimensions, are natural next steps.

Target Audience

This paper is most valuable to researchers and practitioners working on generative models for scientific inverse problems, Bayesian uncertainty quantification, neural operators, and physics-informed machine learning. It is also directly relevant to domain scientists in geoscience, carbon storage, atmospheric remote sensing, and optical imaging who rely on inferred field reconstructions and their uncertainties, and to benchmark designers who need a distributional evaluation template beyond pointwise accuracy.

Authors’ abstract

Generative models are increasingly used to solve scientific inverse problems, but existing evaluations still focus primarily on whether a method can produce a single plausible reconstruction. This is insufficient for ill-posed problems, where multiple solutions may be consistent with the same sparse or noisy observations. In these settings, a method can achieve strong pointwise accuracy while still failing to capture the true posterior through mode collapse, overconfident uncertainty, or averaging incompatible solutions. We introduce PosteriorBench, a benchmark for evaluating the distributional accuracy of generative inverse solvers. PosteriorBench evaluates four physics-based inverse problems: Darcy flow inversion, Poisson source recovery, carbon capture and storage, and light transport material inference. For each task, we construct high-fidelity reference posteriors using computationally heavy but established procedures such as rejection sampling and Markov chain Monte Carlo, enabling direct assessment of whether solvers recover the full set of solutions rather than the single best sample. We pair these references with a five-metric posterior evaluation suite: posterior-mean error, posterior-standard-deviation error, maximum mean discrepancy, sliced Wasserstein distance, and radially averaged power-spectrum error. These metrics assess pointwise accuracy, marginal uncertainty, distributional alignment, and global frequency fidelity. The benchmark spans sparse sensing, low-resolution observations, nonlinear forward models, varying noise levels, and multimodal priors, with a unified pipeline for distribution matching and uncertainty quantification. Our experiments reveal substantial distribution-matching gaps across current solvers, while showing that neural operators improve resolution robustness, and guidance weights and generation noise are key to posterior-variance calibration.