Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Online Bayesian Calibration under Gradual and Abrupt System Changes

NeurIPS 2026 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Simulation-based inference · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Bayesian Recursive Projected Calibration (BRPC), an online Bayesian calibration framework that separates a projected particle update for parameters from a conditional GP discrepancy update, with BOCPD- and CUSUM-based restart variants for abrupt regime changes.

Problem

Bayesian model calibration for digital twins must handle streaming data under simulator mismatch, where systems drift gradually or shift abruptly; classical offline calibration suffers from parameter–discrepancy confounding and assumes stationarity.

Approach

Online projected calibration: particle-filter update of calibration parameters using a discrepancy-free projected likelihood, followed by a conditional linear-Gaussian (GP) discrepancy update, combined with restart mechanisms (restarted BOCPD, wCUSUM on prequential scores, and residual re-anchoring) that reset the state at detected regime changes.

Key results

• On the synthetic sudden(3) scenario, B-BRPC-RRA reaches Precision@2 0.986 and Recall@2 0.999 with theta-RMSE 0.018 versus 0.166 for BC(80).
• On the plant-simulation drifting scenario, B-BRPC achieves theta-RMSE 0.800 versus 3.936 for BC(80) and 9.562 for DA.
• On plant-simulation sudden(5), B-BRPC-RRA reaches Precision@2 1.000 and Recall@2 1.000 with Response RMSE 0.128 versus 1.000 for DA.

Surprising findings

• Overly confident or overly adaptive discrepancy posteriors can hurt restart detection, since routine variation may look anomalous or the discrepancy may absorb regime shifts.
• B-BRPC attains high recall but excessive restarts, while C-BRPC trades recall for precision.

Limitations

• Restart mechanisms lack strong false-alarm control; B-BRPC produces many restarts.
• Experiments are on low-dimensional synthetic and one plant-simulation benchmark; high-dimensional systems are left to future work.
• B-BRPC-RRA is computationally costly because the discrepancy GP is refit offline at each update.

Experimental setup

Datasets: synthetic streaming benchmark adapted from static calibration example plant-simulation discrete-event factory benchmark

Benchmarks: synthetic streaming calibration benchmark plant-simulation benchmark

Metrics: theta-RMSE Response RMSE number of restarts Precision@2 Recall@2

Baselines: BC(80) sliding-window Bayesian calibration DA data assimilation calibration

Builds on: projected calibration Bayesian calibration with model discrepancy (Kennedy and O'Hagan) BOCPD restarted BOCPD CUSUM

Closest papers by meaning

• Bayesian Adaptive Calibration and Optimal Design (NeurIPS 2024 · similarity 0.86)
• Sampling-based Multi-dimensional Recalibration (ICML 2024 · similarity 0.83)
• Addressing Misspecification in Simulation-based Inference through Data-driven Calibration (ICML 2025 · similarity 0.82)
• Deep Bayesian Filter for Bayes-Faithful Data Assimilation (ICML 2025 · similarity 0.82)
• Conjugate Bayesian Two-step Change Point Detection for Hawkes Process (NeurIPS 2024 · similarity 0.82)
• Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting (NeurIPS 2025 · similarity 0.82)