PosteriorBench: A New Benchmark for Evaluating Generative Inverse Solvers

What changed

Researchers have introduced PosteriorBench, a novel benchmark for evaluating generative inverse solvers. Current evaluation methods often focus on single plausible reconstructions, which is inadequate for ill-posed scientific inverse problems where multiple solutions can be consistent with sparse or noisy observations. PosteriorBench aims to assess the distributional accuracy of these solvers, moving beyond pointwise accuracy to capture the true posterior distribution. The benchmark includes four physics-based inverse problems: Darcy flow inversion, Poisson source recovery, carbon capture and storage, and light transport material inference. For each task, high-fidelity reference posteriors are generated using established methods like rejection sampling and Markov chain Monte Carlo. These references are then used with a suite of five metrics—posterior-mean error, posterior-standard-deviation error, maximum mean discrepancy, sliced Wasserstein distance, and radially averaged power-spectrum error—to evaluate pointwise accuracy, marginal uncertainty, distributional alignment, and global frequency fidelity.

Why it matters for builders

This benchmark offers a more comprehensive way to assess generative models used for solving scientific inverse problems. Builders can leverage PosteriorBench to identify shortcomings in their models, such as mode collapse or overconfident uncertainty, and to ensure their solvers accurately represent the full spectrum of potential solutions. This leads to more reliable and trustworthy AI models for scientific applications.

Practical impact

PosteriorBench's evaluation suite allows for a deeper understanding of solver performance across various conditions, including sparse sensing, low-resolution observations, nonlinear forward models, and different noise levels. Initial experiments using the benchmark have revealed significant distribution-matching gaps in current solvers. The benchmark also suggests that neural operators enhance resolution robustness, and that guidance weights and generation noise play a crucial role in calibrating posterior variance.

Caveats and source limits

The provided source is a research paper abstract and associated metadata. While it introduces the PosteriorBench benchmark and its evaluation metrics, it does not contain specific details on the performance of individual solvers or release dates for the benchmark itself. The source mentions that code is available on GitHub, but does not provide direct metrics for repository activity or community engagement. The benchmark's scope is limited to the four physics-based inverse problems described.