Uncertainty Updating AutoGluon-Tabular Model for a Chemical Reactor Twin Based on Hybrid CFD and Device Measurement Data

Chemical reactor twins must reconcile high-resolution CFD evidence with device measurements that are sparse, delayed, and affected by drift. To make the study boundary explicit, the present work focuses on a jacketed stirred reactor with delay-corrected online measurements and CFD-derived transport descriptors. This study develops an uncertainty-updating AutoGluon-Tabular model for a jacketed stirred reactor by fusing 18,240 hybrid samples from CFD snapshots, temperature and pressure probes, shaft torque, feed-flow records, jacket data, and product analyzer measurements. The method converts CFD fields into regime-aware transport descriptors, aligns device signals through delay-corrected windows, learns CFD-measurement residuals with stacked tabular ensembles, and updates prediction intervals through recent calibration evidence. In the held-out campaigns, conversion mean absolute error decreased from 3.86 percentage points for the CFD baseline to 1.14 percentage points, hot-spot temperature error decreased from 6.72 K to 2.05 K, and product-concentration RMSE decreased from 0.084 to 0.031 mol L-1. The empirical coverage of the 90% prediction interval improved from 81.3% to 90.8%, while the average conversion interval width narrowed by 22.6%. The largest gains appeared during feed-ratio disturbances and impeller-speed transitions, where measurement-only learning lacked spatial transport evidence and uncorrected CFD retained systematic bias. The resulting workflow supports transparent monitoring decisions without replacing established CFD practice. The results show that AutoGluon-Tabular can serve as a practical uncertainty updater when CFD features, device evidence, quality indicators, and operating regimes are treated as coupled engineering information rather than independent inputs.