By Gil Cohen, CEO of Imubit
Imubit brought together refining professionals from companies across Europe for two Transcend events, in Athens and Munich. The sessions were frank conversations about the operating challenges the industry faces today and about how refiners are adopting AI in their plants. The themes were clear: safety and reliability, economic optimization and workforce reskilling. Attendees also showed a strong interest in new approaches.
In one session, "From coffee to crude: democratizing model building," Imubit business consultant Dennis Rohe showed attendees how accessible it can be to build a model from raw data and derive decisions from it, whether the process is an espresso machine or a conversion unit.
What are your top concerns?
While meeting refining leaders at the events, I asked them what their top concerns were.
The first answer, short term and day to day, was safety. The long-term answer was survival. Will my company be around in 10 years, given an unpredictable macroeconomic environment and the competitive position of the company and the plant?
Tracing back from that question of survival, several themes come up in conversation, and recent surveys show them too.
- Reliability. Unplanned outages are a plague, especially in aging Western refineries. Most are not formally reported, which makes industry-wide data hard to assemble. McKinsey has written about the case for refinery reliability.
- Economic optimization. Once the plant runs safely and reliably, the priority shifts to optimization, and to dealing with multi-year margin pressure and shifting crude supply.
- Workforce reskilling. Valued, but not prioritized. Most executives say they value reskilling and training, yet because it feels less urgent than safety, reliability and optimization, comparatively few companies are investing in retraining, as one industry survey found.
Why AI is not yet part of daily operations
Adopting AI matters more and more to the longevity of every business. Yet in DNV's 2024 energy industry research, 15% of oil and gas respondents said their organizations use AI in live, day-to-day operations, and 3% reported highly integrated or advanced use. Refining leaders gave several reasons why AI has not yet made it into daily refinery operations:
- data privacy and cybersecurity
- the reliability and trustworthiness of AI models
- how hard it is to understand what a model is telling you
- the bandwidth of plant engineers and operators, for whom keeping the plant running is a full-time job
- initiative overload, where every problem solved with AI becomes a multi-year program with multidisciplinary support, governance structures and change management
What kind of AI can earn a refiner's trust?
Large language models (LLMs) are changing every industry and helping businesses of every size accomplish more with the staff they have. Many roles are being redefined so that AI automates tasks built on set workflows and processing, while people focus on work that needs a deep understanding of the environment and the ability to come up with ideas the AI has not seen before.
LLMs solve a wide range of problems, but these particular problems need a different form of AI alongside them. To trust an AI model with plant decisions, refiners need it to have certain properties.
- It is trained on the plant's own dataset, or on a known, documented and trusted dataset.
- Because the environment is safety-critical, plant engineers control every source of data that goes into training, and every data point within each source.
- The process dynamics the model has captured are transparent to the user, and the user can control them if they choose.
- Model building is open to everyone at the plant: anyone can evaluate, simulate, update and build models.
- It captures the real dynamics of the plant from plant data and shares that insight with all technical staff, instead of relying on human-written documentation, tribal knowledge and biased assumptions.
- It serves as one model for a given area of the plant, useful to every technical discipline that works there.
- It is one initiative that solves several problems, so engineers and operators learn fewer new tools, and management carries less of the burden of new initiatives, governance and change management.
Deep Learning Process Control® (DLPC®) is built around these requirements. Each model learns a unit's nonlinear dynamics from that plant's own operating history, guided by first principles. Process engineers build and evaluate the models themselves in DLPC® Builder and DLPC® Explorer, with no Python and no machine-learning background, and can see the result of a change before they make it. In closed loop, the controller runs on-premise and writes to the existing APC or DCS, inside limits set by plant engineers.
Where does your plant stand?
If these concerns sound familiar, it may be time to look at where your operations stand today. Three questions show it: whether your plan-to-plant gaps are traced to the recurring decisions behind them, whether those decisions are modeled on cause or on correlation, and how autonomously the high-value decisions execute. The answers show where reliability, optimization and workforce goals can be addressed with one trusted initiative.
