Production plans at oil refineries do not always survive contact with the plant floor. Changes in feedstock quality, operating rates or equipment conditions can cause units to perform differently from planning models, affecting yields, product quality and energy consumption and forcing engineers to make manual adjustments during operations.
Industrial artificial intelligence models aim to narrow that gap by combining engineering knowledge with actual operating data, according to Hussein Zein, Emerson’s vice president for Saudi Arabia and Bahrain.
The approach does not replace the conventional models that refineries have relied on for decades. Instead, it enhances them to better represent the nonlinear behavior of units under different feedstocks and operating conditions.
“Conventional models often rely on linear or semi-linear representations of unit behavior,” Zein told Asharq Al-Awsat.
That forces planners to work with simplified assumptions that may lose accuracy as market conditions, feedstock characteristics or operating levels change.
Hybrid models allow operators to test more scenarios, adapt production decisions to changing conditions and reduce the time engineers spend manually adjusting models.
Where conventional planning falls short
Refineries use planning models to turn demand forecasts, feedstock availability, operating constraints and product specifications into executable production plans.
Linear models are easier to use in optimization, but they cannot always capture the complex interactions inside refinery units.
Reactors, separators and blending processes do not necessarily respond linearly to changes in feedstock quality, flow rates or catalyst conditions. When a plan is transferred from the model to the plant, operators may find that yields, product specifications or operating constraints differ from forecasts.
Zein said the gap appears when planning models fail to represent how units behave under actual operating conditions.
The result may be extra engineering work to adjust the plan during execution or operations running below the intended level of efficiency.
The effects can spread beyond one unit. Each unit’s output influences feedstock blending, production targets, margin forecasts, energy consumption and coordination among interconnected units across the refinery.
Predictive accuracy of up to 98.5%
Emerson and Aramco said hybrid models achieved predictive accuracy of up to 98.5% in selected refining units.
The figure reflects how closely the model’s forecasts matched actual operating performance across different feedstocks and operating conditions.
Zein said the figure was the highest level achieved in specific units — continuous catalyst regeneration units and catalytic reformers known as Platformers — and was not an average across all units or operating sites.
Teams are now working to apply the same approach to hydrocracking units and test its performance there.
Even small gaps between forecasts and actual performance can affect yields, energy consumption and production planning. They can also influence feedstock blending, production targets and margin estimates.
Zein said predictive accuracy of 98.5% in some catalytic reforming units reduces the area where margins are often lost: the difference between planned performance and actual results.
When a plan reflects plant behavior more closely, operating teams can follow blending strategies, select feedstocks with greater confidence and pursue optimization opportunities without costly adjustments during execution.
The data provided did not include a specific figure for increases in yields or margins. It linked higher predictive accuracy to narrowing the gap between planning and execution and improving operational decision-making.
Engineering meets operating data
Hybrid models begin with a first-principles foundation that includes reaction kinetics, thermodynamics and mass and energy balances.
AI then uses operational data to calibrate that foundation, allowing the model to reflect the behavior of the actual unit rather than only the theoretical process.
The approach combines the strengths of physics-based and data-driven models.
A purely physics-based model can be difficult to calibrate when trying to capture every changing detail in a real operating environment.
A model based only on data, meanwhile, may produce recommendations that conflict with physical laws or fail when it encounters conditions absent from its training data.
Zein said the hybrid approach provides “rigor and practical accuracy in a single model.”
Engineering knowledge defines the physical limits, while operating data adapts the model to the characteristics and conditions of each unit.
This is particularly important in complex facilities with equipment from different generations and systems that were not built around a unified data architecture.
It also allows refineries to use existing data while relying on engineering principles to compensate for gaps that purely data-driven models may struggle to address.
Humans remain responsible
More accurate models do not mean handing every decision to automated systems.
Zein distinguished between decisions requiring repeated calculations across a large number of possibilities and those involving safety or unusual operating conditions.
AI can support or partially automate feedstock-blending optimization, production planning across several periods and routine model maintenance, while operators and engineers retain oversight.
Decisions involving personnel safety, environmental risks or unusual operating conditions should remain under direct human control.
The aim, Zein said, is for “AI to expand what engineers can achieve without transferring responsibility for critical judgments away from the people who understand the plant.”
When an AI recommendation conflicts with the judgment of an experienced engineer, the disagreement should not be treated as a contest between humans and machines.
It may reveal a factor missing from the model or information known to the engineer that has not yet been translated into rules or data.
“A conflict between an AI recommendation and engineering judgment is a valuable signal, not a problem resolved by choosing one side,” Zein said.
When an engineer’s concern exposes a genuine gap, the model should be updated. When the system identifies an opportunity that was not previously clear, the disagreement becomes part of a process of review and learning.
Models must change with the refinery
Refinery conditions change constantly as catalysts lose activity, equipment deteriorates and feedstock characteristics shift.
Static models may gradually lose accuracy unless they are updated to reflect those changes.
Hybrid models maintain accuracy through two elements.
The first is an engineering foundation representing fixed physical laws, including thermodynamics and material balances.
The second is an AI-based component that can be recalibrated as new operating data becomes available.
This allows the models to track equipment degradation, declining catalyst efficiency and changes in feedstock properties.
It can also reduce the burden of model maintenance compared with conventional methods and give planners greater confidence that forecasts remain linked to the unit’s actual condition.
But limited data and aging equipment are not the only barriers to wider deployment.
Zein said the most common challenge is integrating models, data and workflows across planning, engineering and operations departments.
A facility may have high-quality data and modern equipment, but a project can remain confined to a single unit if it is not incorporated into daily decision-making and the organization’s operating structure.
Continuous corrosion monitoring
Industrial data can also be used to monitor corrosion, a task traditionally based on scheduled inspections or intervention after a problem emerges.
Continuous monitoring offers an updated view of equipment conditions through wall-thickness sensors, wireless communications and real-time data analysis.
Instead of asking only whether a scheduled inspection is due, maintenance teams can assess whether a particular asset requires intervention before it fails.
Zein said the conventional approach could delay production and detect a problem only after it occurred, without always providing a continuous view of how it developed.
Continuous monitoring supports maintenance decisions based on the actual condition of equipment rather than relying entirely on a fixed schedule.
It can also reduce the need for some manual measurements that may be hazardous or physically demanding for workers, but it does not eliminate human inspections.
“Human judgment and verification will remain essential,” Zein said.
Continuous data instead helps specialists focus their time on the assets most in need of intervention.
Safeguards before automation
Cybersecurity, governance and safety requirements depend on how a model is used and how directly it can affect operations.
A system that advises production planners does not require the same controls as one involved in control decisions.
Security requirements include secure architecture, access management and monitoring under established industry standards.
Governance covers model ownership, approval procedures for deployments and updates, and audit records that document changes and decisions.
Safety controls include testing models against operating limits, retaining human oversight for high-consequence decisions and providing fallback procedures when a model operates outside the range for which it was calibrated.
The appropriate controls depend on the application. An advisory system requires different safeguards from a model that can directly influence production or maintenance.
Scaling beyond one unit
Another challenge arises when a company tries to expand a successful application from one unit to an entire facility or across several sites.
Zein said the main obstacles are usually not technical. They emerge when organizations attempt to copy the same model without accounting for differences among units.
Each unit has its own feedstocks, operating range and history. Models must therefore be adapted and recalibrated rather than copied unchanged.
Scaling also requires shared data architecture, standardized modeling practices and organizational processes capable of supporting maintenance and updates across sites.
That is how industrial AI can move from a stand-alone application in one unit to a broader tool for planning and operations, while preserving engineering oversight, security controls and human responsibility for critical decisions.
