PTQ Q3 2025 Issue

An Optimising Brain for your plant

Imubit industrial AI platform

Dynamic process simulator Reinforcement learning controller Performance dashboards Imubit academy

Operator optimisation training Planning tool augmentation Process degradation insights

Imubit Deep Learning Process Control (DLPC)

Imubit value sustainment

Imubit workforce transformation

Data-rst decisions Modelling democratisation One model collaboration Retain new generation

Economic engineering Systems integrations Proactive monitoring and training Continuous model updates

One view, one team

Figure 1 Architecture of the closed loop AI optimisation solution implemented sitewide at Big West

new technology. There is the familiar and comfortable sta- tus quo way of working and the need to understand any technology that could impact productivity, quality, reliabil- ity, or safety. For some, it can be difficult to overcome the perception of AI as a ‘black box’. Traditional, typically linear models are well accepted in refining, thanks to their well-understood and easy-to-explain nature. But the class of deep neural networks capable of solving the industry’s toughest, most nonlinear problems are understood by a much smaller subset of the population. To overcome this complexity, AI solutions must build in explainability functionality, such as providing a sandbox for simulation and validation of pro- cess relationships. The solution to overcome both of these change objections is conceptually simple. Build trust. Build it early and build it often. For AI-based closed-loop optimisation in particular, build the deepest trust with the people in the organisation who have the ability to turn off the technology – operations. Ensuring transparency, involving all affected people, and providing thorough training on the technology and explain- ability features throughout the project lifecycle is crucial. Without a well-planned change management strategy designed around building trust in technology, workforce scepticism may hinder or delay adoption. Early buy-in was accomplished by first implementing a relatively simple model to debottleneck the debutaniser tower. The functionality of the model was easy to compre- hend for the Big West team, who were newcomers to the technology. They intuitively understood the moves that the AI controller was making, and the results were immediate. Within a day or two, tower operation was significantly improved, leaving room to increase throughput in the tower. This built immediate confidence in the technology. w Scalability Big West envisioned sitewide implementation of AIO technology, which presented a need for foundational IT infrastructure and standard processes for application development and deployment. It also meant there would

be variability in model complexity across units. One of the advantages seen in partnering with a technology provider for this venture was the ability to leverage the experience and best practices developed over technology deploy- ments at companies of various sizes to ensure consistent governance across multiple projects in accordance with site change management practices. This proved essential for sustainable scale-up. By addressing these challenges, this refiner was able to move forward with implementing a solution that would simultaneously boost profitability and establish a for - ward-looking operational culture – one equipped to handle evolving market demands and technological advancements. Solution to process and people challenges The challenges this refinery in Utah was looking to solve were centred around nonlinear, complex refining systems that, in some cases, overlapped several processing units. AIO technology (see Figure 1 ) was developed with these challenges in mind and has been proven, with more than 90 applications in the refining industry. The AIO model is developed in the following sequence:  Inferentials Built from years of operating and lab sample data, infer- entials serve as continuous indicators of otherwise slow sampled variables. They may also be used to represent a parameter that is not instrumented but can be readily cal- culated from other correlated variables. Beyond their use in the development of the closed-loop optimiser, these inferentials are being leveraged in other ways, such as determining when products can be switched to on-spec rundown tanks during start-ups, saving hours of slopping valuable products. v Dynamic process simulator Using historical data and advanced modelling techniques, a deep neural network is created to provide a high-fidelity representation of the process. This process model serves as the training environment for the controller to acquire the

PTQ Q3 2025