Digital Twin Off-Normal Fidelity Divergence (DTONFD)

GMI-NUC-002Nuclear OperationsInfrastructure-Dependent

Definition

Ratio of the digital twin's prediction error under transient, coastdown, or off-normal conditions to its prediction error under steady-state normal operations, measuring how much the twin's reliability degrades precisely when decisions depend on it most.

The Measurement Gap

Digital twins in nuclear plants are validated primarily against steady-state operating conditions, where the plant runs at rated power and parameters are stable. But the decisions that depend most on twin accuracy occur during off-normal conditions: transients, power reductions, coastdown to refueling, degraded-equipment operation, and LCO entries. No legacy metric measures how much the twin's prediction accuracy degrades when the plant moves away from steady state. Capacity factor measures generation output. WANO indicators track broad operational performance. The plant process computer records actual conditions. None of them compare the twin's predictions to reality under exactly the conditions where margin calculations matter most. The invisible failure mode is that the twin performs well during normal operations (building confidence) but its accuracy collapses during transients and off-normal conditions (where decisions depend on it), and no metric detects this divergence.

Decision Supported

Can the plant rely on digital twin margin predictions during transient or degraded-equipment operation, or does the twin's accuracy collapse under exactly the conditions where margins are tightest?

Formula

Plain language: Compute the twin's average prediction error during off-normal conditions. Compute the twin's average prediction error during normal steady-state. Divide. If the ratio is 1.0, the twin is equally reliable under both conditions. If it's 3.0, the twin is three times less accurate during off-normal events: precisely when accuracy matters most.

Notation:
DTONFD = MAPE_offnormal / MAPE_steadystate

Where:
MAPE_x = (1/N_x) × Σ |twin_prediction_i − actual_measurement_i| / |actual_measurement_i|

Variables:

  • twin_prediction_i = digital twin's predicted value for parameter i (units vary: temperature in °F/°C, pressure in psi/MPa, flow in gpm/m³/h, power in % RTP)
  • actual_measurement_i = plant instrumentation reading for parameter i at the same time point (same units)
  • N_offnormal = number of comparison points during transient/off-normal periods (count)
  • N_steadystate = number of comparison points during steady-state operation (count)
  • MAPE = Mean Absolute Percentage Error (dimensionless ratio)

Condition classification:

  • Steady-state: reactor power stable within ±2% RTP for >4 hours, no active LCOs on modeled systems
  • Off-normal: power transients (load follow, coastdown, trip recovery), degraded equipment configurations, LCO entries on modeled systems, post-trip stabilization

Sign convention: DTONFD = 1.0 means equal fidelity. Values >1.0 mean off-normal fidelity is worse (bad). Higher is worse. A twin with DTONFD of 5.0 is effectively useless for transient decision support.

Aggregation period: Per operating cycle (18–24 months), with off-normal events aggregated across the cycle.

Edge Cases

  • Very few off-normal events in a cycle: nuclear plants are designed to operate at steady state. Off-normal comparison points are scarce. Supplement with planned transients (power reductions, surveillance testing that perturbs parameters) and simulator-validated scenarios. Report sample size alongside DTONFD.
  • MAPE_steadystate approaches zero: if the twin is nearly perfect at steady state, even small off-normal errors produce large DTONFD ratios. Set a minimum MAPE_steadystate floor of 0.5% to avoid division-by-near-zero distortion.
  • Different parameters have different fidelity profiles: compute DTONFD per parameter group (thermal-hydraulic, neutronics, structural) and report separately, a twin may be excellent at predicting temperatures and poor at predicting vibration.

Confidence Grading (required alongside every DTONFD value):

Grade Basis Interpretation
A: Empirical 5+ observed events in measurement window provide direct tail observation Measurement grounded in observed outcomes; highest confidence
B: Hybrid 1-4 observed events plus Monte Carlo simulation with Bayesian updating Blends observation and simulation; moderate confidence
C: Simulated Zero observed events; derived entirely from Monte Carlo simulation Modeled, not observed; treat as indicative, not validated

A DTONFD of any value at Grade C is a simulation that agrees with itself. It tells you the model is internally consistent but not that it matches reality. Grade C scores should trigger proactive stress-testing against historical near-miss events and cross-asset analogues rather than acceptance as calibrated.

Proposed formula, pending validation pilots.

At a Glance

Unit
Percentage
Trend Direction
Higher
Reporting Frequency
Monthly
Category
AI Model Trustworthiness in Safety-Critical Decisions
← Work Package Scope Accuracy Index (WPSAI)PRA Input Currency Index (PRICI) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment