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

Technical Summary: Validation-Gated Multi-Agent Governance for Online Adaptation of Thermal–Hydraulic Surrogate Models

Problem Statement
Real-time thermal–hydraulic (TH) forecasting is critical for experimental operation support, digital-twin monitoring, and passive-safety assessment. While high-fidelity system codes are too computationally expensive for second-by-second forecasting, Artificial Intelligence (AI) surrogates offer low-latency alternatives. However, a significant limitation exists: surrogates are typically trained, validated, and frozen offline within a specific pretraining envelope. When deployed in facilities where operating regimes shift (e.g., due to passive heat-removal pathway changes or control-policy adjustments), these "condition-locked" models can produce physically inaccurate trajectories despite smooth outputs.

Existing approaches to continual learning face a stability–plasticity dilemma. Aggressive online learning risks amplifying sensor faults, overfitting to recent windows, and inducing catastrophic forgetting of earlier regimes. Conversely, conservative deterministic rules (thresholds, cooldowns, rollback) constrain model changes but lack the reasoning capability to diagnose why errors occur, select appropriate model families for adaptation, or veto short-term improvements that mask tail-risk behaviors. The paper addresses the gap of how to integrate agent-based reasoning into safety-relevant model evolution without granting agents direct control over the real-time controller.

Methodology
The study proposes a guarded continual-adaptation framework that separates the real-time serving path from the adaptation path. The system employs a "champion–challenger" architecture where only a validated "champion" model serves predictions, while "challengers" are trained and evaluated in the background.

1. Multi-Agent Governance Structure:
   The framework utilizes five role-separated agents to support decision-making within a deterministic safety envelope:

   • Monitor: Summarizes rolling errors, warning streaks, clipping behavior, and dominant error channels.
   • Diagnosis: Maps error signatures to hypotheses regarding model staleness or coupling changes.
   • Adaptation: Proposes bounded fine-tuning, challenger refresh, or hold actions.
   • Safety-Auditor: Verifies that proposed updates do not violate hard gates or introduce tail-risk regression.
   • Orchestrator: Consolidates inputs into a final recommendation.
   • Crucially, these agents provide recommendations and vetoes, but deterministic validation gates retain final authority over model replacement.

2. Offline Screening and Initial Champion:
   Seven surrogate families (LSTM, GRU, Transformer, neural ODE, GNN, DeepONet, and Temporal Fourier Neural Operator) were screened using blocked three-fold cross-validation on offline data (normal operation, power reduction, and station blackout). The Temporal Fourier Neural Operator (Temporal-FNO) was selected as the initial champion due to the lowest cross-validation Mean Absolute Error (MAE).

3. Streaming Deployment and Safety Gates:
   The system was deployed on two held-out transients: a heat-pipe insertion configuration (altering passive heat removal) and a separate operating transient with an unseen schedule. The adaptation loop includes:

   • Deterministic Gates: Warning/critical thresholds, cooldown periods, prediction clipping against physical bounds, and bounded fine-tuning with weight blending.
   • Promotion Criteria: A challenger is promoted only if it improves mean error without degrading exceedance, tail-risk, or clipping behavior.
   • Rollback: Triggered by non-finite predictions or excessive clipping.

4. Experimental Modes:
   Seven operating modes were compared to isolate the contributions of different components:

   • Static: Frozen offline champion.
   • Rule-H: Deterministic adaptation rules without agents.
   • Shadow: Periodic background refresh of top families without agent review.
   • Single-H/Full: Single-agent planning (selective or frequent).
   • MA-H/Full: Role-separated multi-agent governance (selective or stepwise).

Key Results
The study evaluated performance across six paired statistical units (two stream splits × three seeds) using 95% bootstrap confidence intervals.

• Static Performance: The frozen Temporal-FNO champion yielded a channel-averaged MAE of 7.06 with a 56.8% warning-exceedance ratio, confirming the condition-locking problem in shifted regimes.
• Rule-Based Adaptation: The Rule-H mode reduced MAE to 6.54 (7.3% improvement) using only deterministic gates, establishing a baseline for adaptive deployment.
• Shadow Learning: Background refresh alone (Shadow) showed minimal improvement (MAE 6.96), indicating that generating challengers without governance is insufficient.
• Multi-Agent Governance: The MA-Full mode (stepwise multi-agent review) achieved the best performance with an MAE of 5.72 and a warning-exceedance ratio of 35.8%. This represents a 19.0% improvement over the Static mode.
• Statistical Significance: Paired bootstrap intervals for MA-Full versus Static excluded zero for both MAE and exceedance, indicating a statistically significant reduction in error. However, intervals among other adaptive modes overlapped, preventing definitive ranking among them.
• Model Evolution: In the MA-Full mode, the system successfully transitioned the serving model from the initial Temporal-FNO to Transformer and then to Graph Neural Network (GNN) challengers. These transitions were driven by stream-conditioned evidence (e.g., delayed thermal memory and hydraulic coupling) rather than arbitrary preference.
• Catastrophic Forgetting: Back-testing on the original offline pretraining windows showed that the adapted models retained performance on the original domain (MAE decreased from 2.91 to 1.96), indicating that replay-based updates did not cause catastrophic forgetting.

Significance and Claims
The paper claims that validated, gate-controlled adaptation can support auditable surrogate evolution in safety-relevant contexts. The primary contribution is demonstrating that multi-agent governance, when bounded by deterministic safety gates, can effectively guide continual surrogate evolution without acting as the real-time controller.

• Governance vs. Control: The study emphasizes that agents serve as a governance layer to diagnose error signatures, prioritize candidate families, and review risky promotions, while deterministic gates retain deployment authority.
• Adaptability: The results indicate that no single offline architecture is sufficient for all operating regimes. The ability to migrate from a Temporal-FNO to Transformer and GNN architectures based on real-time evidence allows the system to adapt to changes in thermal inertia and flow redistribution.
• Modest Scope: The authors explicitly note limitations, including the small number of paired statistical units (six) and the fact that the study does not establish universal superiority across all possible operating regimes. The findings are specific to the tested experimental facility and transients. The paper does not claim to solve all continual learning challenges but demonstrates a specific, guarded approach that balances adaptability with safety constraints.

In conclusion, the study validates a framework where role-separated agents and deterministic gates work in tandem to maintain surrogate model accuracy during operating-regime shifts, offering a pathway for more robust digital twins in nuclear thermal–hydraulic systems.