Hybrid Physics-ML Digital Twin

In Ansys Mechanical, build a composite wing spar model: 2m span, I-beam cross-section, carbon fiber/epoxy laminate using ACP (Ansys Composite PrepPost). Apply realistic boundary conditions: fixed at root, distributed aerodynamic loads plus a point mass at the tip representing fuel/payload. Define material degradation to simulate fatigue: stiffness reduction as a function of accumulated cycles.

Run static and modal analyses to validate the model: check that natural frequencies and deflections are physically reasonable. Generate the high-fidelity dataset you'll use to train the ML correction: simulate 50 damage states (increasing stiffness reduction from 0% to 30%) and log strain gauge readings at 8 sensor locations plus modal frequencies.

Use Ansys ROM Builder to create a Modal ROM: project the FEA solution onto the first 20 structural modes, capturing 99%+ of the response energy with a tiny fraction of the computational cost. The ROM takes inputs (loading conditions, temperature) and outputs responses at sensor locations in microseconds — fast enough for real-time operation.

Validate ROM accuracy: compare ROM vs. full FEA predictions for strain at all 8 sensor locations across all loading conditions. ROM error should be < 2% for the nominal (undamaged) structure. Verify that the ROM updates correctly when you pass it different loading conditions — it must generalize beyond the training snapshots.

Import the ROM into Ansys Twin Builder as a Twin Runtime. Create a schematic connecting: sensor data inputs (simulated strain and acceleration channels) → ROM component → ML correction → state estimator → health output signals. Twin Builder handles the data routing and time-stepping automatically.

Set up Twin Builder to call your Python ML correction script via the Python Scripting component. Verify the complete data flow with a simple test: input a known loading condition, confirm the ROM output is correct, and confirm the ML correction is applied. Log all intermediate signals for debugging.

The ML correction addresses two problems: (1) ROM modeling error at nominal health, and (2) health-dependent deviations from ROM predictions. Train an ensemble of gradient boosted trees (XGBoost) on the 50 damage state simulations: inputs are ROM predictions + loading conditions + sensor readings, output is the residual (sensor measurement - ROM prediction).

The residual is a function of the damage state. A structural health indicator emerges naturally: as damage accumulates, the residual pattern changes in characteristic ways at different sensor locations. Train a second model that maps the learned residual pattern to a damage severity score (0 = healthy, 1 = critical). Cross-validate using leave-one-damage-state-out validation.

Implement an Extended Kalman Filter (EKF) that fuses the ROM+ML correction predictions with noisy sensor measurements. State vector: strain at 8 sensor locations + first 3 modal frequencies. Measurement noise covariance R (calibrated from sensor specs). Process noise covariance Q (calibrated to match ROM uncertainty).

The EKF provides: (1) filtered state estimates smoother than raw sensor data, (2) uncertainty bounds (covariance matrix), and (3) innovation signals (measurement - prediction) that are diagnostic of sensor faults or anomalous structural behavior. Implement innovation monitoring: large innovations signal either sensor failure or structural damage outside the model's training range.

Run a 72-hour simulated structural health monitoring scenario: the spar starts at 0% damage, accumulates fatigue damage to 20% by hour 48, then has a discrete damage event (impact) at hour 60 that causes 5% sudden stiffness loss at a specific location. Feed synthetic sensor data (from high-fidelity FEA at each damage state) into the deployed hybrid twin.

Plot: damage indicator vs. time, sensor strain vs. predictions, and innovation signals. The hybrid twin should: (1) track gradual damage accumulation accurately, (2) detect the sudden impact event within 1–2 hours, and (3) localize the damage to within ±15% of the true impact location using the spatial pattern of sensor innovations. Document detection latency and localization accuracy.