Visualizing Uncertainty So Operators Can Judge It
The AI's confidence is never hidden behind a single number: distributions, bands, and hatching let operators see how much to trust each estimate on both machines.
Uncertainty is first-class, not a footnote
An estimate without its uncertainty is a trap: it looks equally authoritative whether the twin is confident or guessing. Every quantitative L6 surface therefore carries visible uncertainty. The twin's surrogate models and PINNs produce both a value and an uncertainty (via ensembles, dropout, or predictive variance), and L6 renders that uncertainty in a form matched to the display.
Encodings by context
| Where | Uncertainty encoding |
|---|---|
| Time series / trends | shaded confidence band around the line |
| 3D twin overlay | desaturation + diagonal hatch on low-confidence regions |
| KPI tiles | margin-to-limit shown with a band, not a point |
| Forecasts (predictive shadow) | fan chart widening with horizon |
| Anomaly alerts | explicit model confidence 0..1 on the card |
Aleatoric uncertainty (irreducible measurement noise) and epistemic uncertainty (the model being out of its training distribution) are shown differently, because they demand different responses. Wide aleatoric bands mean a noisy sensor; growing epistemic uncertainty means the machine has drifted into a regime the model has not seen — a cue to weight the AI less and physics/operator judgment more.
# ensemble gives value + both uncertainty kinds for display
preds = [m(x) for m in ensemble] # K surrogate members
mu = mean(preds)
epi = var(preds) # spread across members -> epistemic
alea = mean([m.noise_var(x) for m in ensemble]) # per-member noise -> aleatoric
return dict(value=mu, band=sqrt(epi+alea), epistemic=epi, aleatoric=alea)Calibration matters more than magnitude
A confidence display is only useful if it is calibrated — if 80% confidence is right about 80% of the time. Calibration is tracked continuously (reliability diagrams computed offline) and a miscalibrated model's confidence is visibly flagged so operators do not over-trust it. This ties directly to trust and override UX: the whole point of showing uncertainty is to make the override decision well-informed.
The encoding is also matched to how fast a decision must be made. A wide band on a slow process trend invites investigation; a wide band on the disruption-risk forecast demands immediate weighting of the human's own read of the plasma. Uncertainty is never rendered as a single scalar tucked in a corner, because a corner number is ignored under load. It is drawn on the same axes as the value it qualifies, at the same visual weight, so the operator cannot read the estimate without also reading its doubt. Provenance completes the picture — see confidence and provenance display.