estimate future fouling rate or forecast end-of-run (EOR) timing with confidence. How AIO addressed it Imubit’s AIO platform combined several years of operating and crude-property data to train a time-dependent fouling model. The model learned from historical episodes of accel - erated fouling, correlating heater-duty profiles, pass veloci - ties, and feed blends with subsequent TMT rise (see Figure 1 ). In effect, it captured the dynamic trade-off between con - version severity and fouling rate that static heuristics cannot represent. Once deployed, AIO continuously: • Forecasts TMT drift and the days remaining to EOR under current firing, coil duty, and feed slate. • Diagnoses pass-level behaviour, distinguishing local hot spots or flow maldistribution. • Advises unequal pass balancing, departing from the con- ventional equal core outlet (COT) strategy, to redistribute heat flux and slow localised fouling. This proactive control approach prevents rapid degrada - tion in isolated passes before recovery is no longer possible. Results Guided by AIO’s forecasts, operators adjusted firing and pass-flow distribution to maintain target conversion without exceeding TMT limits. The site achieved a ~33% reduction in furnace-related downtime, smoother reliability planning, and higher average conversion across runs. By explicitly quantifying the conversion-versus-fouling relationship, the AIO model enabled the refinery to push conversion further while keeping fouling within recoverable limits, capturing additional margin without compromising coil integrity. Engineering insight Beyond operational control, the system delivered diagnos - tic intelligence. AIO’s variable-importance analysis revealed which factors, notably coil-duty distribution and specific
crude-blend properties, most strongly drove fouling rates. This capability bridges the traditional gap between process engineering and planning economics, allowing engineers to connect reliability outcomes directly to feedstock decisions, something unattainable with conventional linear control or heuristic monitoring. Case 3: Optimising feed composition across a In catalytic reactor systems, operators constantly bal- ance short-term yield against long-term catalyst health. Operating at higher severity improves immediate conversion but accelerates deactivation; running too gently preserves activity but sacrifices throughput and profit. The true opti - mum is dynamic: push when the catalyst is fresh, conserve as it ages, and opportunistically exploit conditions that offer yield gain with minimal impact on life. Traditional APC frameworks cannot address this problem. Their linear, memoryless models treat every hour of operation as identical, with no awareness of cumulative degradation or cycle position. As a result, control actions are optimised for the next few hours rather than for the entire catalyst cycle. catalyst cycle The challenge How AIO solved it As a specialty chemical producer, Imubit’s AIO platform developed a cycle-dependent yield model that explicitly cap - tured the evolving relationship between feed composition, product yield, and multi-day rate of change (ROC) in yield, used as a proxy for catalyst deactivation. To teach the model this long-term behaviour, the control window was expanded to five days (from the typical one- to two-hour horizon). This allowed AIO to observe the full context of yield evolution and to distinguish temporary yield improvements (caused by severity spikes) from true perfor- mance trends associated with degradation (see Figure 2 ). The model thus learned that while short-term increases in
Reactant ratio
Molar yield % Mean: -0.612 Median: -1.685 STD: 9.574 CV
MV
7.0
7.5
6.7
Step MV: + 1
Context (5 days control window)
6.3
6. 0
5.9
5.5
4.5
12:00 AM
12:00 PM
12:00 PM
12:00 AM
12:00 AM
Molar yield % CV
3.0
72. 6
CV response (5 days): -13
64.7
1.5
56.7
Context (5 days control window)
0
48.8
-2 6 .0 -13.2
-0.423
12.4
25.2
38.0
40.8
Gain (∆ Dependent / ∆ Independent)
12:00 PM
12:00 AM
12:00 PM
12:00 AM
12:00 AM
Figure 2 Neural network prediction showcasing how context window can be extended to days
43
Catalysis 2026
www.digitalrefining.com
Powered by FlippingBook