By Alex Sorokin, Business Consulting Engineer, Imubit
Speakers at the 2025 World Petrochemical Conference returned again and again to one theme. The chemicals industry faces a period of uncertainty driven by geopolitics, trade conflicts and supply chain volatility. Plant operators cannot control those factors. They can control how well they run the assets they already have, and many speakers pointed to AI as the way to do that. What the conference did not settle was how AI makes a chemical plant more resilient to that uncertainty.
Where AI programs stall
AI covers a wide set of tools, applications and approaches. That breadth makes it hard to know where to start, how to show value from a first project, and how to keep capturing that value once the project team moves on.
In our experience the AI programs that hold up are the ones where cross-functional teams see the benefit for themselves. When planning, operations and engineering work from the same model of the plant, they can settle questions together that each group used to answer alone. At Oxbow's petroleum coke calcining sites, operators, engineers and leadership could all see the results, and that gave them a basis to expand into other improvements (read the Oxbow story).
Across work with plants in volatile markets, three practices keep coming up.
1. Move from single-unit to plant-wide optimization
Operating points in different areas of the plant have to shift together to meet common, plant-wide goals. That means the right feeds, catalysts and process conditions, inside process and environmental limits.
Plant-wide automation has been hard to achieve with traditional methods because of the number of constraints, scenarios and people involved. An optimizer has to understand how the interactions between units affect plant-wide goals and constraints. If one unit shuts down, or takes an unscheduled catalyst change, how should the other units run to re-optimize the plant?
In one deployment, the controller adjusts the feed rates and APC targets of several units to compensate for run-limiting equipment fouling, then readjusts them when the equipment is cleaned. In another, the models challenged long-established control strategies on individual units and found value at the plant level.
The lesson has held across customers, regions and industries. Optimizing the value chain needs more coordination and flexibility than a series of disconnected single-unit solutions can provide.
2. Run existing assets to their real constraints
The aim is to reduce giveaway, downtime and time spent at sub-optimal rates by running to the real underlying limits as feeds vary, equipment fouls and catalysts degrade.
At the low point of a market cycle, plants are under the most pressure to turn a profit and have the least capital for equipment upgrades. Existing assets have to be pushed further. Even the best operators run to conservative targets to avoid deactivating catalyst, degrading equipment or violating limits they cannot see clearly. Existing automation narrows the gap, but value stays on the table when the real limits are poorly understood, nonlinear or subject to long time dynamics.
These are problems a learned model of the process is suited to. On semi-batch delayed coking units, the controller works to maximize the value of each batch as it runs. At a site with two parallel mixed-feed steam crackers, it sets feed rates against the actual coking limits and downstream constraints instead of the values the site had historically assumed.
3. Put the models in the hands of process experts
Organizational adoption is a major bottleneck for automation and digitalization. It has been true of APC, where many controllers drift out of tuning and maintenance after commissioning, and plants are right to expect the same risk with AI.
The answer is AI that people across disciplines can use. In the Imubit Platform, process engineers build Deep Learning Process Control® (DLPC®) models in DLPC® Builder with no Python and no machine-learning background. They review what the models have learned and compare what-if scenarios, so they can see the result of a change before they make it. Some sites use the models in operator training, as a safe place to test moves and compete against the optimizer. Others feed what the models learn into established workflows, from scheduling APC gains to updating LP vectors.
Plants rely on technology more than ever, and that technology needs willing and capable users. Sustained value comes from AI that is built from the start to be explainable and usable by the people who run the plant.
Imubit applies Closed-Loop Physical AI across refining, chemicals and polymers, cement, and mining and minerals. The Imubit Platform maps plan-to-plant gaps to the recurring decisions behind them, models the causality of each unit from its own operating history, and takes the critical decisions to autonomous execution, inside limits set by plant engineers, on top of the APC and DCS already in place. Read more about chemicals and polymers.
