Mark Venables spoke to Dennis Rohe, Business Consulting Team Leader at Imubit, to explore whether closed-loop AI technology can transform process control. While the technology promises dynamic adaptability and unprecedented efficiency, questions remain about its scalability, transparency, and ability to deliver measurable results across diverse industrial environments.
Closed-loop AI technology has been hailed as a game-changer for process control, but does it truly deliver on its ambitious promises? Imubit’s Optimizing Brain claims to redefine operational efficiency by integrating deep learning with plant systems to execute strategies autonomously and continuously. This innovation has even spurred ARC Advisory Group to create a new category, ‘Closed Loop AI Optimisation’, acknowledging its potential to disrupt traditional approaches. Yet, questions remain about its scalability, reliability, and long-term value.
Imubit positions its Optimising Brain as a revolutionary departure from traditional model-based systems, which often falter under the dynamic conditions of refineries. By building AI models rooted in historical data and domain expertise, the company promises both marginal gains and transformational results. However, critics might question whether this data-first approach can consistently outpace hybrid systems that combine AI with proven first-principles methods.
The technology’s claimed ability to simulate millions of scenarios and compress decades of experience into actionable insights sounds impressive, but is it realistic? While early adopters like Marathon Petroleum Corporation report significant improvements, including a $1.9 million annual value increase, these case studies are not universally representative. This article will critically examine the claims, assess whether closed-loop AI lives up to the hype, and explore whether it genuinely outperforms conventional methods in practice.
Is closed-loop AI genuinely transformative?
Dennis Rohe, Business Consulting Team Leader at Imubit, highlights two key distinctions that set their approach apart from traditional optimisation tools like Model Predictive Control (MPC) and Advanced Process Control (APC). “Firstly, while we are not offering a traditional MPC system, we rely on a dynamic process model based on historical process data,” he explains. “This is enhanced with engineering knowledge and a sprinkling of first principles to ensure a sound solution. The focus is on using the customer’s data to build a system that works for their specific processes.”
The second distinction lies in the use of reinforcement learning. “Our control mechanism goes beyond algorithms; it learns from millions of iterations and adapts dynamically, almost like a human operator,” Rohe says. Unlike MPC, which relies on static models, reinforcement learning offers the potential to respond more effectively to changing conditions.
While these claims are compelling, they raise questions about scalability and real-world performance. Can a system train on historical data truly emulate the decision-making of an experienced operator across a range of complex and unpredictable industrial environments?
Many companies claim their models are dynamic, but Rohe argues that Imubit’s approach is fundamentally different. “Usually, other companies achieve the dynamic aspect by performing step testing,” he says. “This involves moving a variable and observing the response, but it is time-consuming and disruptive. By contrast, we leverage the millions of data points plants have collected over the years. The AI can find patterns and responses using deep learning, even to incrementally small movements, without requiring a step test.”
This method reduces disruption and allows the system to extract dynamics directly from historical data. However, the effectiveness of this approach depends on the quality and richness of the data available. If a plant’s historical data is incomplete or inconsistent, does this limit the AI’s ability to deliver the promised insights?
Rohe is confident that deep learning can deliver results that traditional methods cannot. “By observing the data, the AI identifies patterns a human would likely miss. It creates an intelligent system capable of adapting to conditions not anticipated during the initial setup,” he says.
Going beyond incremental gains
A common criticism of AI in industrial control is that it merely offers incremental improvements. Rohe disputes this, stating that Imubit’s solution goes further. “Our dynamic process model identifies relationships and nuggets in the data that traditional tools, like MPC and APC, often miss,” he explains. “If I have a changing relationship, I could probably control far more smoothly and efficiently against constraints than if I assumed a single static relationship.”
He emphasises the transformative potential of data visibility. “Avoiding premature assumptions allows the data to reveal the plant’s full potential. This enables operators to target the right areas for optimisation, unlocking opportunities that might otherwise remain hidden,” he adds. These claims suggest significant advantages, but do they hold up under scrutiny? While uncovering hidden insights is undoubtedly valuable, the practical impact depends on whether these insights translate into measurable gains. Can Imubit consistently deliver transformational results across diverse and complex industrial setups?
Imubit positions its solution as a complement to existing systems rather than a replacement. “We believe in letting each technology excel in its specific role,” Rohe explains. “For instance, DCS handles regulatory control, while APC optimises against constraints. Our solution sits on top of this, addressing more complex trade-offs that APC cannot resolve.”
This layered approach minimises disruption and preserves the value of legacy systems. “If a customer is happy with their APC system, we integrate our solution to unlock additional value,” Rohe says. “We handle the very complicated trade-offs that an APC, with its priority order, cannot solve.” Yet, integrating advanced AI into an existing system is challenging. Adding a sophisticated layer to the ageing infrastructure could create interoperability issues or overwhelm operators with new, complex data. As Rohe points out, success depends on ensuring the solution remains user-friendly and aligned with the operator’s needs.
Addressing AI’s Achilles’ Heel
Model drift, the tendency for AI models to lose accuracy as conditions change, is a well-known challenge. Rohe describes how Imubit addresses this: “We monitor our models and the process in parallel to ensure there is still a match. If deviations occur, we initiate retraining offline to align the AI with current conditions.”
He stresses the importance of keeping the AI’s learning process under control. “We do not let the AI continuously adjust itself in real time because it could learn something that deviates from first principles or customer requirements. Instead, we validate changes offline before reinstalling them at the plant,” he explains. This cautious approach prioritises safety and reliability but may limit adaptability in dynamic environments. Is the trade-off worth it? And does the reliance on offline retraining mean the system could lag rapidly evolving conditions?
The use of deep learning introduces another challenge: transparency. Rohe acknowledges that the black-box nature of deep learning can create mistrust among engineers. “Deep learning has this image of being a black box, but we want to open that box and validate its findings against process knowledge,” he says. This effort to demystify AI is critical for building trust, but transparency alone may not be enough. Even with validated outputs, will operators feel confident relying on a system they may not fully understand?
The data-first debate
Imubit adopts a data-first approach, rejecting hybrid models that combine first principles with AI. “Hybrid models tell the AI what the first-principles model considers truth and then fit the process data to that framework,” Rohe explains. “This constrains the AI and devalues the data by discarding insights that deviate too much from the predefined model.”
In contrast, Imubit’s data-first philosophy allows AI to fully explore the data’s potential before engineers validate the findings. “By starting with the data, we avoid prematurely imposing constraints. Engineers then use their expertise to judge what makes sense and refine the model,” he says. This approach maximises flexibility but may challenge organisations accustomed to relying on first principles. Can a purely data-driven strategy consistently outperform hybrid models that combine the strengths of AI and engineering expertise?
Rohe acknowledges that poor data quality can undermine the success of AI systems. “Your data is valuable, and AI can find value in it, but bad data leads to bad predictions,” he warns. Imubit works with clients to consolidate and standardise their data, but this process can be resource-intensive.
“Good data governance is critical,” Rohe adds. “Having a well-organised data historian ensures success and accelerates the time to value.” However, the cost and complexity of data preparation may deter smaller organisations or those with limited technical expertise.
Imubit emphasises the importance of post-implementation support. “We monitor the system to ensure it continues delivering value and retrain models as needed,” Rohe explains. “Many customers may not have the expertise to evaluate whether a model is still performing well, so we take on that responsibility.”
Is the market ready?
Closed-loop AI is still early, representing a small fraction of the process optimisation market. “The industry often misunderstands what AI entails,” Rohe says. “Many solutions labelled as AI are rudimentary, and customers need education to differentiate between them.” Imubit’s focus on advanced AI and measurable results positions it as a leader, but broader adoption depends on overcoming misconceptions and addressing practical barriers.
Closed-loop AI technology has the potential to revolutionise process control, but its transformative promise must be critically evaluated. Imubit’s Optimizing Brain offers compelling advantages, from dynamic adaptability to data-driven optimisation, but it also faces scalability, transparency, and real-world applicability challenges.
For the technology to fulfil its potential, it must consistently deliver measurable results, earn the trust of operators, and address the complexities of implementation and data quality. As the process industry grapples with these questions, one thing is clear: the journey toward true transformation requires innovation and critical scrutiny.
