A refinery does not look like a place waiting for a software update. Pipes sweat. Furnaces roar. A minor adjustment can change the mix of products coming out hours later, while feed quality, energy prices and equipment limits keep moving underneath. The operator at the console is responsible for all of it. Now imagine a neural network asking for the keys.
That is the business Imubit chose. The Houston company builds models of refineries, chemical plants, cement kilns and other continuous-process facilities. Its software learns from years of operating history, simulates possible moves and searches for a more profitable combination of throughput, yield, energy use and constraints. In recommendation mode, a person decides what to do. In closed loop, Imubit’s controller writes setpoints into the plant’s existing distributed control system.
It is the difference between a weather forecast and a thermostat. Plenty of industrial AI vendors forecast. Imubit wants to adjust the temperature.
The product is a loop, not a chatbot
Imubit’s architecture has two main pieces. The cloud-based Industrial AI Platform is the workshop: teams build, test, simulate and monitor models. The Deep Learning Process Control application runs on premises, inside the operational-technology environment, and connects to historians, laboratory data and the control system. The company says the production controller does not require a two-way exchange with its cloud database.
The umbrella product is called Optimizing Brain. Underneath sits what Imubit calls a Foundation Process Model - a reusable model of one customer’s plant, refreshed with new data. The same asset can support a reinforcement-learning controller, operator training, planning checks, catalyst-degradation analysis, a dynamic simulator and performance dashboards. That reuse is the clever part. A costly modeling project becomes infrastructure rather than a single-purpose experiment.
The buyers are not generic knowledge workers. They are process engineers, advanced process control specialists, planners, console operators and plant executives. Public customer material names Marathon Petroleum, CITGO, HF Sinclair, Big West Oil, Oxbow, Ash Grove Cement, Preem, Delek, Chevron and others. In early 2025 Imubit said seven of the 10 largest U.S. refiners used its technology across more than 90 applications. Its newer materials put the count above 100.
What failed first: the old modeling ritual
The first thing Imubit attacks is not an AI competitor. It is the traditional step test. To build an advanced process control model, engineers may spend roughly two weeks asking operators to make deliberate moves, then watching how the unit responds. The exercise consumes engineering time, can push production away from its economic best and creates a predictable control-room argument: the engineer wants revealing data; the operator wants a quiet, safe shift.
Imubit’s answer is to use the history the plant already created. Years of changing feed, fouling, weather and operating regimes provide a broader sample than a short test. Neural models estimate nonlinear relationships, while reinforcement learning practices against the model before a controller acts on the real unit. It does not make physics optional. Engineers still choose variables, encode constraints, inspect relationships and validate behavior. It changes where the experiments happen.
What changed minds was evidence at awkward edge cases. Marathon engineer Chris Harrison said the company was not looking to replace its existing APC estate. It wanted help with problems those tools found difficult. A coker application reduced suboptimal cycles and associated giveaway by a reported 25 percent. Big West used its model as a training game, which made operators compete over better strategies. HF Sinclair compared model relationships with planning assumptions and found useful discrepancies. The AI arrived as an augmentation, then earned a larger job.
A hydrocracker case study offers the neatest economic claim: full value in under six months, 0.6 percent more throughput beyond the incumbent APC volume and a $20,000-per-day increase in the economic objective. Oxbow reported a closed-loop kiln implementation in six months, with 1 to 3 percent more yield and 15 to 30 percent less natural gas. These are selected, vendor-published cases, not a promise that every plant gets the same result. Still, in industries measured by barrels and energy intensity, a small percentage is an adult-sized number.
What did it cost?
PRICE
NOT FOUND
Imubit does not publish list pricing. Its reseller agreement describes an annual recurring subscription for each deployed site, plus packages and services defined in an order form. The honest buyer’s calculation is therefore not seats times dollars. It is expected margin or energy value minus implementation, integration, validation, training and ongoing model care.
The company says some applications pay back in less than 90 days. Buyers should ask for the baseline, confidence interval, uptime, excluded production periods and who pays for the work required to keep the model current. A refinery may happily spend seven figures on software that reliably creates eight figures of margin. It should be less happy with a beautiful counterfactual.
Imubit’s business appears to mix recurring enterprise software with implementation and domain services. The subscription is organized by deployed site rather than a simple consumer-style seat. That fits the product: one installation touches a facility’s data, models, workflows and controls. A partner channel is now part of the expansion strategy. Preem is deploying the system on the FCC at its Lysekil refinery. Ketjen is combining catalyst expertise with Imubit’s live platform. Evonik began more cautiously in Singapore in 2026, using an open-loop pilot to evaluate catalyst-life insight and operating tradeoffs before considering further automation.
The moat is permission
Imubit competes with familiar control and optimization stacks from Honeywell, AspenTech, Yokogawa, Siemens and Schneider Electric, plus newer industrial analytics vendors. But its daily rival is the bundle that already works: local APC controllers, first-principles models, spreadsheets and an experienced operator’s intuition.
The differentiation is a data-first, nonlinear model that can span multiple units and move from insight to governed execution. Legacy APC is often excellent on a well-defined local problem. Imubit is most interesting when the relationships are messy, the optimum crosses unit boundaries or economics change faster than a static model can be rebuilt. The company wisely presents its layer as compatible with existing controls rather than insisting the customer rip them out.
That compatibility does not remove the human problem. A model can be accurate and still sit unused if an operator cannot see why it moved. Imubit has responded with gain visualizations, what-if tools, operator training and staged deployment. The Evonik pilot is the cleanest expression of the playbook: begin with recommendations, establish a baseline, validate under live conditions and let safety governance determine whether the loop ever closes.
The copyable playbook
Founders can steal five things. First, sell in the customer’s native unit. Dollars per day and barrels per hour beat an abstract accuracy score. Second, use the installed base as a distribution surface: Imubit sits above existing control systems. Third, turn implementation knowledge into product checks. Fourth, make one expensive core asset useful across several workflows. Fifth, treat adoption as a feature. The simulator and explainability layer are not marketing garnish; they are how the controller earns runtime.
Customers can copy the sequence even without buying Imubit. Choose one constrained, valuable process. Write down the economic objective and safety boundaries. Audit the data before selecting a model. Start in shadow or recommendation mode. Compare decisions against experienced operators. Measure against a defensible baseline. Automate only when the organization can explain when to disengage.
When it will not work
Imubit’s own readiness guidance is refreshingly specific. It recommends at least six months of process history, roughly one-minute recording intervals, enough laboratory samples, complete records for process values and controller states, reliable time synchronization and healthy base-layer controls. No algorithm can recover a valve position that was never stored or command a loop that lives permanently in manual.
- Thin, compressed or unrepresentative operating history
- Broken instruments or weak base regulatory control
- No economic target worth optimizing at plant scale
- Operators excluded until the commissioning date
- A simple process already handled well by existing APC
- No OT security, change-control or fallback discipline
There is another limit: model drift is a maintenance problem, not a launch-day problem. Feed changes, catalyst ages, equipment fouls and markets shift. Imubit’s claim of a continuously refreshed model is only valuable if the plant has people and process to review it. Closed loop also concentrates responsibility. The software must respect hard constraints, fail safely and leave operators able to take over.
That is why the company’s trajectory matters. It began with a controller and has expanded toward a shared industrial modeling platform, workforce training and partner-delivered applications. The move is commercially logical, but it also admits something important: the algorithm is only one organ. Data, control infrastructure, economics and human judgment make the body run.
Imubit’s proposition is neither magic nor modest. It asks a plant to turn its accumulated history into a colleague that can practice for thousands of simulated years, explain what it learned and eventually take a shift at the controls. The buyer should remain skeptical. The operator should, too. The compelling part is that Imubit built a product around earning the right to change their minds.