Title: Validation-Gated Multi-Agent Governance for Online Adaptation of Thermal-Hydraulic Surrogate Models under Operating-Regime Shift

Abstract:

While artificial intelligence surrogates enable second-by-second thermal-hydraulic forecasting, models that are selected and frozen during offline training often become condition-locked when deployed outside their pretraining domain. To address this, we present a guarded continual-adaptation framework applied to experimental thermal-hydraulic loop data. This system employs role-separated agents—specifically the Monitor, Diagnosis, Adaptation, Safety-Auditor, and Orchestrator—to identify error signatures, prioritize candidate model families, and oversee promotions. Final authority regarding model replacement remains with deterministic champion-challenger gates and background shadow learning.

We screened seven surrogate families using blocked three-fold cross-validation. A temporal Fourier neural operator was chosen as the initial champion for forecasting 10-second trajectories based on 60-second history windows across two held-out transients, utilizing three seeds per adaptive mode. In static deployment, the model yielded a channel-averaged Mean Absolute Error (MAE) of 7.06 and a warning-exceedance ratio of 56.8%. Rule-based adaptation lowered the MAE to 6.54, while shadow refresh alone kept performance close to the static baseline.

The MA-Full mode, which requires the role-separated multi-agent council to review every step of the evaluated stream, achieved the lowest mean error of 5.72 and a 35.8% exceedance rate. This represents a 19.0% improvement over the static deployment. Although paired bootstrap intervals against the static model excluded zero, intervals among the adaptive modes overlapped, and the limited number of six paired units precludes broad statistical claims. Validated promotions from the neural operator to Transformer and graph neural network architectures demonstrate that logged, gate-controlled adaptation supports auditable surrogate evolution, even as deterministic gates retain ultimate deployment authority.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC