PINN-accelerated diffusion models for amortized Bayesian inversion of chloride transport in concrete
Über dieses Buch
Accurate estimation of chloride transport parameters is critical for predicting chloride penetration and the residual service life of marine concrete structures. In-situ continuous monitoring using embedded solid-state chloride sensors (e.g., Ag/AgCl ion-selective electrodes) enables rich spatiotemporal observations at multiple cover depths over the entire service life; however, the resulting data streams are inevitably corrupted by sensor drift, temperature-induced potential shifts, and electromagnetic interference, making the inversion ill-conditioned. This study proposes a PINN-accelerated conditional score-based diffusion (PINN-Diff) framework that (i) trains a conditional Physics-Informed Neural Network (PINN) surrogate once over the full prior parameter space to serve simultaneously as a millisecond-scale forward solver and a high-throughput training data-generator, and (ii) performs amortized posterior sampling via a conditional score-based diffusion model equipped with Classifier-Free Guidance (CFG) and Feature-wise Linear Modulation (FiLM) conditioning, thereby enabling real-time Bayesian inference for any new observation without retraining. On synthetic benchmarks with noise levels σ ∈ {0.5%, 1%, 2%, 5%, 10%} across 20 test cases with 10 repetitions each, PINN-Diff achieves 26% lower posterior mean absolute error than MCMC-PINN (normalized MAE 0.040 vs. 0.054 at σ = 2%) and over 50× faster inference (0.08 s vs. 4.2 s per 1,000 posterior samples), while maintaining well-calibrated 90% credible intervals (Coverage_90 = 88 − 91%). Ablation studies on the noise schedule, observation sparsity, and PINN surrogate capacity confirm the robustness of the framework across a broad range of design choices.
