Definition
Ratio of actual catalyst deactivation rate following a feed-quality upset to the AI-predicted deactivation rate for the same post-upset period: detecting whether the AI's steady-state-trained model accounts for the nonlinear damage that concentration spikes cause relative to baseline exposure.
The Measurement Gap
Catalyst life prediction models train on years of steady-state operating data. They learn how feed quality, temperature, and conversion drive gradual deactivation. They do not see, in proportion to the damage they cause, the spike events: four hours of sulfur breakthrough on a reformer, a contaminant slug from an upstream unit, a feed-water excursion that hits a hydrotreater. A spike at five times normal concentration can do twenty-five times the damage of an equivalent steady-state exposure because the damage function is nonlinear. The steady-state model rolls the spike into the running average and reports a small bump in predicted deactivation. The actual deactivation is much larger and shows up months later as conversion decline that arrives ahead of the predicted replacement date. Catalyst remaining-life prediction, conversion, selectivity, and weighted-average bed temperature all measure the rate, but the model's prediction error during the days following an upset is not measured anywhere unless this gap is named.
Decision Supported:
After a feed-quality upset, does the AI's remaining catalyst life prediction still reflect reality, or has the upset caused permanent damage that the steady-state model cannot see, making the prediction dangerously optimistic?
Formula
Plain language: After a feed-quality upset, measure the actual catalyst deactivation rate over the 30 days following the event. Divide by the AI's predicted deactivation rate for the same period. Above 1.0 means the model underestimated the damage.
Notation:
CUDF = Deactivation_rate_post_upset / Deactivation_rate_predicted_post_upset
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Deactivation_rate_post_upset | Measured catalyst deactivation rate in the 30 days following an upset, indicated by conversion decline, selectivity change, or required temperature rise for constant conversion | rate (e.g., % conversion loss / day) |
| Deactivation_rate_predicted_post_upset | AI-predicted deactivation rate for the same 30-day post-upset window using the model's standard methodology | rate (same units) |
| Upset definition | Feed contaminant (sulfur, heavy metals, chlorides, water) exceeding the 95th percentile of historical feed quality for more than 4 hours | threshold + duration |
Report CUDF per upset, plus a quarterly summary by contaminant type.
Proposed formula, pending validation pilots.