Between campaigns 5 and 6, we lowered the MTO reactor inlet temperature setpoint by 7 degrees C on the recommendation of the optimization model. The change seemed conservative, almost too minor to matter. Yet campaign 6 used 12 percent less energy per kilogram of jet-range product than campaign 5, while yield remained within measurement noise. Temperature alone did not explain the result. It revealed a coupled interaction among inlet temperature, space velocity, and inter-bed quench duty that we had not explicitly characterized, but that the model had inferred from campaign history. That was the point when we began using the optimization layer to make operating decisions, not simply to report on them.
We have completed 12 pilot campaigns at our Zurich facility. This is a candid account of how our run-to-run optimization system developed, what it improved, and what it did not deliver. The figures are internal pilot benchmarks, identified as such, and apply to our pilot reactor at its current scale.
What a Pilot Campaign Includes
At our facility, a pilot campaign is a sustained operating period, typically 10 to 21 days, covering the full MTJ process from methanol feed through fractionation to a jet-range product stream. Each campaign begins with a fixed catalyst load, a defined feedstock batch, and a target operating window established in advance.
Conditions change during a campaign. We adjust feed rate, temperature setpoints, quench flows, and fractionation cut temperatures as observations come in. The model recommends a setpoint for the next operating period, typically a 4-hour window, using current sensor readings plus the current campaign history and all prior campaigns.
Run-to-run learning is constrained by the catalyst reset between campaigns: each one starts with fresh or freshly regenerated catalyst. Early-campaign dynamics therefore differ from middle and late periods. The model must represent catalyst age explicitly, and correctly describing those age effects was a main challenge during the first five campaigns.
What the Optimization Layer Actually Does
Our optimization layer uses Gaussian process regression with an acquisition function that balances exploiting the current apparent optimum against exploring poorly characterized operating regions. This is a standard Bayesian optimization formulation, and the underlying algorithm is not proprietary. The application-specific element is the state representation. It includes current setpoints and derived process variables: a catalyst age proxy modeled from temperature history and regeneration cycles, upstream MTO conversion estimated from outlet gas composition, and an inter-bed thermal imbalance indicator derived from the multi-point thermocouple array.
At the end of each operating period, the model takes in energy consumed, the product yield estimate from the online GC, and the thermocouple array's temperature profile shape. It then updates the predictions behind the next recommendations. Operators see a suggested setpoint adjustment, a confidence interval, and an estimated outcome range, and can accept, modify, or override it. Acceptance rose from about 60 percent in early campaigns to around 85 percent in the most recent four campaigns, reflecting increased model accuracy and operator trust.
Energy Use Across 12 Campaigns
We measure total electrical and thermal energy consumed per kilogram of collected jet-range product fraction. This accounts for differences in campaign duration and throughput. The absolute figures are internal, but the campaign-to-campaign trajectory is clear.
Campaigns 1 through 4 focused mainly on collecting data. The model explored a wide range of operating regions to learn process behavior, so energy use was both variable and high on average. That was the expected cost of building the initial dataset.
Campaign 5 was the first in which the model had enough data to recommend exploitation meaningfully. We followed it, but the predicted improvement did not appear. The catalyst charge began with different activity from earlier charges because the regeneration procedure had changed. The model's catalyst-state assumption therefore differed from the actual state. We spent most of campaign 5 and part of campaign 6 diagnosing the mismatch.
After campaign 6, correcting the catalyst age model brought energy use down and stabilized it. By campaign 9, average energy per kilogram of jet-range product was approximately 14 to 18 percent below the campaign 1 to 4 average, according to our internal pilot benchmark data. Campaigns 10 through 12 stayed in that range, with less variance than earlier campaigns.
The 14 to 18 percent gain is not a dramatic engineering result. It comes from making many small setpoint choices correctly over hundreds of operating periods. Each choice is modest, but together they produce a meaningful reduction.
Process Consistency: Gains and Gaps
By process consistency, we mean how closely product quality and yield metrics cluster around their targets from run to run. Some variables improved meaningfully; others did not improve at all.
The largest gains were in variables directly affected by model recommendations: inlet temperature, space velocity, and quench flow. In our internal tracking, campaign-to-campaign variance in jet-range fraction percentage roughly halved between campaigns 1 to 4 and campaigns 9 to 12.
Variables that did not improve were driven by factors outside model control: feedstock quality variation between batches and end-of-life catalyst behavior, which grows less predictable with more regeneration cycles. We expected aging effects to be captured better as campaign count grew; they were captured partially. The model outperforms manual operation on early and mid-life deactivation; late-life catalyst behavior remains difficult to predict reliably. We are not saying the model controls these factors; we are saying it compensates for some catalyst deactivation.
What the Improvement Pattern Says About the Model Architecture
Across 12 campaigns, the largest gains came from setting inter-stage quench duty correctly. Quench flow between the MTO and oligomerization sections controls the temperature entering the oligomerization bed and the olefin concentration, because it affects MTO conversion before the inter-stage. Those factors interact strongly with oligomerization reactor behavior. Manual operation generally used a conservative quench setting that was not optimal for the actual feed condition. The model learned to adjust quench duty dynamically from the upstream MTO state, and that adjustment accounts for a disproportionate share of the energy improvement.
We are not saying the quench-duty interaction was unknown. Lars Henriksen had identified its importance in process design, and methanol-to-olefin literature discusses it. We are saying the model converted that conceptual understanding into numerical recommendations that followed actual process state instead of a nominal design point. The practical difference is applying known process behavior consistently during operation.
Where the Limits Are
Twelve campaigns provide a useful dataset, but not a large one by industrial process optimization standards. The model is reasonably confident in the operating region we access routinely, and its uncertainty bounds widen when a recommendation would leave that region. That uncertainty is genuine. We have not explored some parts of the operating space, and they may contain better operating points than those we have found.
Moving from pilot scale to a larger reactor will require recalibration. Pilot dynamics do not perfectly represent a larger fixed-bed reactor, especially for axial heat distribution and inter-bed mixing. We will use the pilot model as a starting prior, not a direct transfer, and update it with the first campaigns at larger scale.
Campaign 13 will be the first extended run with a new catalyst formulation evaluated by Lars Henriksen and the process team. We will treat it as partial re-exploration, broadening the operating envelope to characterize the catalyst rather than immediately narrowing to the model's current optimum. The exploitation and exploration trade-off remains whenever the process changes. For a new pilot operator, the 12-campaign pattern suggests measured exploration followed by model updates as catalyst and process conditions change.