Subscribe
Share
Share
Embed
In this episode of the Hydrocarbon Engineering Podcast, we’re joined by Simon Rogers, Chief Technology Officer at KBC (A Yokogawa Company). Simon explains how physics-based models, AI, and contextualised enterprise data are converging to improve performance and decision making in the downstream sector.
We discuss:
We discuss:
- What has changed in the industry that makes process twins and industrial AI possible today.
- Why process twins are a prerequisite for industrial AI.
- Organisational and cultural challenges facing refiners when implementing process twins and industrial AI.
- Practical steps that refiners should take today to prepare for increasingly autonomous operations.
- And much more.
This episode of the Hydrocarbon Engineering Podcast is sponsored by Sherwin-Williams Protective & Marine.
Sherwin-Williams Protective & Marine delivers world-class industry subject matter expertise, unparalleled technical and specification service, and unmatched regional commercial team support to customers around the globe, including in the energy market. The company’s broad portfolio of high-performance coatings and systems – including protective liquid and powder coatings, as well as fire protection coatings – excel at combating corrosion and help customers achieve smarter, time-tested asset protection. For more information, visit protective.sherwin.com.
Creators and Guests
What is Hydrocarbon Engineering Podcast?
The Hydrocarbon Engineering podcast: a podcast series for professionals in the downstream refining, petrochemical and gas processing industries.
Hello and thanks for joining us for another episode of the Hydrocarbon Engineering Podcast. Today we're exploring how physics based models, AI and contextualized enterprise data are converging to improve performance and decision making in the downstream sector. Our guest is Simon Rogers, Chief Technology Officer at KBC, a Yokogawa company and our conversation follows on from a recent article that Simon wrote for the September 2026 issue of Hydrocarbon Engineering Magazine titled Foundations for Industrial AI. Now if you have a subscription to Hydrocarbon Engineering magazine you can access Simon's article by logging into your account over at our website and if you're not a subscriber we'd like to offer you a free trial today simply visit hydrocarbonshineering.com forward / magazine and follow the simple on screen instructions. So let's talk AI with Simon.
Advert:This episode of the Hydrocarbon Engineering Podcast is sponsored by Sherwin Williams Protective and Marine. Sherwin Williams Protective and Marine delivers world class industry subject matter expertise, unparalleled technical and specification service, and unmatched regional commercial team support to customers around the globe, including in the energy market. The company's broad portfolio of high performance coatings and systems, including protected liquid and powder coatings as well as fire protection coatings, excel at combating corrosion and help customers achieve smarter, time tested asset protection. For more information, visit protective.sherwin.com.
Callum O'Reilly:So welcome, Simon. How are you today?
Simon Rogers:I'm good, thanks Callum, thanks very much for having me on today.
Callum O'Reilly:Great stuff, okay Simon, so this is your first time joining us, we always like to begin with some introductions. So can you just tell us a little bit about yourself and your role at KPC?
Simon Rogers:So I'm a chemical engineer by training. I'm currently the Chief Technology Officer at KBC. So I'm responsible for all of our software development and product management. Actually this is the third time I've been at KBC. I started my career at KBC in 1986 after finishing at Imperial College.
Simon Rogers:And yeah, it's been very nice to be here today, look forward to chatting with you.
Callum O'Reilly:Okay Simon, thanks so much for that introduction. As you know there's a lot of excitement surrounding AI and its potential to revolutionize industrial operations. But what has changed recently for the industry that makes Process Twins and industrial AI possible today?
Simon Rogers:We've been working with Process Twin and some form of Industrial AI for many years, so it has been possible, but I think what's making it much easier today, first of all cloud computing, so we have much more powerful computers available, on almost unlimited computing power. We have better data platforms now and then on the AI side of obviously the latest is generative AI and that has significant applications in the context of process twins and digital twins.
Callum O'Reilly:And why are process twins a prerequisite for industrial AI?
Simon Rogers:Process twins are required in the process industries for two main reasons: one is to improve the quality of the data. Obviously AI requires quality Within the refining industry and the process industries generally there is significant uncertainty about any process measurement and you can correct that by making sure that overall there is a material and energy balance across the process. So you can certainly use simulation to improve the quality of the data. You can also use simulation to calculate things which you can't measure. In particular things like stream qualities, product qualities, feed qualities.
Simon Rogers:These are often measured only in the lab. There may be some limited online analyzers but they're not widespread. But you can calculate those values from process simulation. For instance, if you're doing lab samples, may be doing that once a week. So that's a relatively small amount of data when you're looking over a year, you've only got 50, 52 samples.
Simon Rogers:So machine learning requires much more data than that. So I think improving the quality of the data that you do measure and then adding data that you cannot measure.
Callum O'Reilly:And can you explain why it's essential for AI to be grounded in engineering physics?
Simon Rogers:In our industry, obviously safety is the number one priority. It's very important that any recommendations or any action that AI takes is trusted and does not cause any unsafe situation. In order to do that it needs to obey the laws of physics and a good example is that when you train machine learning or any AI algorithm you use the data that you've seen in the past but of course there are many new situations that may occur or even just new feedstocks, products, different grades of product that you need to make, changing the catalyst. There are a lot of things which change in the process which you may never have seen before and it's impossible to get an accurate machine learning model for those cases. Machine learning doesn't extrapolate, it interpolates, it's okay in the region that you've got data but outside of that region it's not reliable, and of course it really does need to be reliable in hazardous industries.
Callum O'Reilly:And Simon, why is maintaining synchronization between engineering models and plant operations still such a challenge?
Simon Rogers:Yeah, that's a good question. Has been forty years I've been trying to work on that challenge. One of them is the data, that because of the uncertainty in the measurements, the inaccuracy in the measurements and sometimes failures in the sensors, there's always a little bit of a question as to whether there's a problem with the model or a problem with the data, so that's one challenge. And then historically it's taken a lot of engineering time to review and compare the model with the data and then to calibrate the model, tune the model. Now with modern machine learning techniques, a lot of that can be automated.
Simon Rogers:We've introduced a new solution which we call Process Twin Pro, which looks at all of the data sets that you have over the last year or so, chooses the best quality sets and then uses those to automatically calibrate and tune the model, so the model stays automatically matching the plan. So it's now much less time consuming to ensure that your models are matching the plan and then you can use modern machine learning techniques like anomaly detection to identify problems with the data, problems with the model and identify when you need to regenerate or retune and recalibrate your simulation models.
Callum O'Reilly:So Simon, can you outline the business case for Process Twin and Industrial AI? Because as you know many refinery leaders struggle to justify digital investments. So can you talk us through where refiners are seeing the greatest value? So for example, we talking about energy efficiency, yield improvements, planning accuracy, or reliability? And in order to justify these investments, how should executives measure success?
Callum O'Reilly:What outcomes should they expect?
Simon Rogers:Absolutely, the important thing whether you're talking about simulation, process twins, AI, any application that you implement, the main focus has to be the outcome. And generally the outcomes are those that you've mentioned, so yield and energy would be probably the most impacted outcome, so increasing the yield of valuable products by reducing the amount of product quality giveaway, reducing the energy to produce the same amount of product and increasing the amount of the high value products versus the low value products. So ultimately you can use a cheaper feedstock, cheaper crude oil for instance in the refining business. But it's not only about yield and energy, the process twin can also provide valuable information to improve the safety and the reliability of the plant. All of those outcomes can be impacted by process twins and when I talk about process twin, in most cases they are hybrid between a first principles simulation model and machine learning, so you get the best of both worlds.
Callum O'Reilly:And your article introduces the concept of the digital thread which connects engineering models, plant data, planning, maintenance and enterprise systems. So why is this connectivity becoming essential for AI enabled refinery operations?
Simon Rogers:That's always been a dream, so I work mostly in the refining industry historically and people have been trying to implement a refinery information system for decades and with limited success, some success, they've always been very difficult to implement and very difficult to maintain. And the goal of a refinery information system historically is to provide a single window on the operation of the refinery. So, if you're an operator, you have all the information you need to operate the plant, if you're an engineer, you have all the information you need to do your engineering job, and same for a planner or maintenance person. And the reality is today that you often have to go to many different applications to get the information that you need to do your job. So you spend a large percentage of your time trying just collecting information and of course that's a waste of time and also frustrating for the engineers, they'd much rather have everything at their fingertips so they can focus on the high value and more interesting activity or task.
Simon Rogers:Now with generative AI, you can see the opportunity to create those information systems, that single window onto the plant much more easily than you did in the past. But in order to do that, to use things like Microsoft Copilot, M365 Copilot, You need to be able to give the language models of the AI the correct context. Now in the plant, the way to do that is probably to use some sort of knowledge graph, so that you connect all of the different systems and allow the language model, the AI, to collect the right information from the lab system, the process historian, from your maintenance systems or planning systems. And often those systems will refer to the same piece of information differently, so you've got to be able to connect them so that the language model knows that the tag name in SAP for instance from your new maintenance system is X and the same piece of equipment is referred to as Y in your process historian or in your P and IDs or in your engineering data. So organizing all of that data into a digital thread so that you can provide the right context to the language model and ultimately these language models will start to move towards AgenTik AI, that will be used in AgenTik AI pipelines to automate work processes and of course it's very important that they can access the tools that they need, these AI agents, and access the information that they need from all of those disparate data sources.
Simon Rogers:So some sort of ontology or knowledge graph is required in order to organize all of the information in the different databases that you have within a refinery or in any process plant.
Callum O'Reilly:And what factors contributed most to the success of the implementation that's highlighted in your article? And what advice would you offer others that are starting a similar journey?
Simon Rogers:Yeah, that's a good question. I mean obviously the first thing is to choose a high value use case, so the particular example that's mentioned in the article is related to planning. So within a refinery people use linear programs to plan the refinery, that means to select what crudes they're going to buy and to optimize the operation of the refinery to produce the products that they need to produce. And those decisions that they're making using those linear programs are hundreds of millions of dollars in the case of buying the cargo of crude. So it's very important that the models that you use to make those decisions are as accurate as possible.
Simon Rogers:So in this particular case, we're using our Process Twin Pro system to monitor how well the planning model matches the first principle simulation model and matches the actual operation of the unit. What we found in this particular case in one example was that when the hydrocracker in this plant was in high conversion mode, the accuracy of the LP model was very poor. So we were able to improve the accuracy of the identify that and then improve the accuracy of the model during those high conversion periods and that means that the accuracy of the LP overall, the refinery model that they are using to buy the crude and to optimize the operation of the refinery is improved and in that case that contributed to millions of dollars a year, 2 or $3,000,000 a year of extra profit. So high value use cases are the most important criteria.
Callum O'Reilly:So let's move beyond the technology for a moment, and I was wondering what organisational or perhaps cultural challenges do refiners face when implementing process twins and industrial AI?
Simon Rogers:The challenge there, it depends in region to region, but in the West in particular, there's been a reduction in the number of people that are working in refineries because of the margin pressures. So that means that people have less time, and also possibly less experience. There's also significant turnover of staff as people retire, new people come into the industry. So using these tools has been something that requires significant amount of training and practice that you need to continue to use them on a fairly regular basis otherwise it can be difficult. With AI now it's much easier to create natural language interfaces rather than having to have a user specific graphical user interface for the simulation tool.
Simon Rogers:I think then the second with AI in particular is trust, that people have to trust. Everybody who's heard of generative AI has heard of hallucinations and you know they're therefore a bit nervous about implementing those systems and following the recommendations because in the end the person that's taking the action, the operator, engineer, whatever, they're the ones that are responsible and accountable and they have to be able to trust whatever tools they use, whether that's simulation models or whether it's AI. And then I think the last one is that there's also sensitivity about sharing the data. So again, there are issues related to governance and security of the data. Sometimes that's because the company values its data, doesn't want to share that data, sometimes they have agreements, non disclosure agreements and other agreements with licenses and others that mean that they can't share the data.
Simon Rogers:So that's another challenge related to the, which is not technical specifically, it's more commercial. And I think that in general there's also a concern in some cases that people's jobs are going to be threatened by AI and of course in my experience there's plenty of work to do, there's plenty of opportunities for instance a process engineer on a refinery to add more value, if you can take away some of the mundane work, you'll get I don't think that means that the process engineer's job goes away, not in any time soon anyway, they would just be able to work more effectively and deliver more value which should improve their job retention prospects. So given all of
Callum O'Reilly:these challenges that we've talked about Simon, what practical steps should refiners take today to prepare for increasingly autonomous operations? So where should it start and what common mistake should they be looking to avoid?
Simon Rogers:One thing I've seen over the years in the recent use of machine learning before generative AI, a lot of our clients have been trying to implement machine learning systems. I think that when they've tried to do that with a pure data approach without trying to incorporate physics, they've failed most of those systems have failed. There are cases where you can get value from machine learning, pure machine learning. But in general I would recommend a hybrid approach. So use a combination of first principles physics and machine learning.
Simon Rogers:Make your first principle models data driven, so connect your models, probably hybrid models, so physics and machine learning, connect them to the plant so they're constantly learning from the plant, constantly adapting to the actual operation of the plant and then integrate the results of that simulation which is running in near real time with all of your other data using that digital thread, using a knowledge graph or an ontology and then provide access to that data to everybody and every application that might be able to use the output of simulation. I see the Process Twin as the brain of an autonomous plant, because it's impossible to completely automate a plant without a model of how it operates. That model needs to incorporate physical laws and respect physical laws. So connect your models to data, make sure that they automatically adapt to the actual operation of the plant, making them data driven and then making them data centric, making the application data centric so that the value is coming from the data and not from the application itself. You need to make that data available to all of the applications, all the people that could get benefit from the data to give you that single window, that single view on the plant.
Simon Rogers:Initially that will be mainly for humans but in the medium term we'll start to see AI agents and other workflow automation applications using that data more autonomously.
Callum O'Reilly:So you mentioned the medium term there Simon, so how do you see the role of refinery engineers evolving over let's say the next five to ten years?
Simon Rogers:Yeah, that's a good question, think this is the $20,000,000 question at the moment in pretty much every industry. I perhaps can answer it from my own perspective, from KBC's perspective. We develop software, obviously software engineering is one of the most impacted areas of work by generative AI, at least initially. And my perspective is that if we could double the productivity of our software engineers and the chemical engineers that develop our software, my take on that would be to double the amount of development that we do. To create twice as much value with the same number of people, not to half the number of people and create the same amount of value.
Simon Rogers:And I think the same is true in a refinery mostly. There's a lot of work that could be done, a lot of value that could be created by engineers and operators and others on the plant, which they don't have time to do. So I think that what you'll see is that the number of people on the plant at least initially won't fall dramatically, it might over time because obviously the vision ultimately is to have a fully autonomous plant. In the medium term I would expect that the refineries would just become more efficient, less emissions, higher margin, of course they're competing with each other, so to some extent that's a zero sum game, but that's the opportunity, is to use the technology to increase the performance of the refinery and perhaps particularly with the net zero challenge that we have, the greenhouse gas emission challenge, reducing the energy and emissions of the plants, making them more sustainable is a great opportunity. And then also as part of that move towards a circular economy, we're going to be transitioning from fuels to chemicals and obviously biofeedstocks and green energy.
Simon Rogers:That requires significant amount of change in the refining sector and in other industries and that's going to require time from process engineers and going
Callum O'Reilly:to require them to have better tools to be able to make those changes quickly because we don't have that much time to reduce the emissions of carbon dioxide and other greenhouse gases. So finally, if you could leave refiners with one key message about preparing for the next generation of industrial AI, what would it be?
Simon Rogers:I think the main thing is to incorporate physics. So don't use a pure data driven approach, machine learning or generative AI which is a form of machine learning. It requires scientific rigor, so the right approach is a combination of first principles and machine learning. Together I believe that they can enable autonomous operation. When in Yokogawa had introduced this concept of industrial automation to IA to IA about ten years ago and at the time I was thinking well that's a long way away, it's a nice vision but it's a long way away.
Simon Rogers:But with the progress in the last three or four years in generative AI and AgenTik AI more recently, you can start to see that this could become a reality possibly even in my lifetime which is something I could never imagine. But start with those principles and combine physics and machine learning, that's the takeaway.
Callum O'Reilly:Great, Simon, thank you so much for joining us today. Really interesting to learn a little bit more about some of the foundations for industrial AI in the downstream sector, and it will be really fascinating to see how some of these technologies that we've discussed continue to shape the next generation of refining and petrochemical operations. Thanks again for your insights, really appreciate it.
Simon Rogers:Thanks so much Callum, it's been a pleasure to speak to you today, and if anyone would like to continue the conversation please contact me on LinkedIn. I always like to learn, there's huge amount to learn, the things are changing so fast at the moment and it would be great to continue the conversation with anyone that's interested.
Callum O'Reilly:That was Simon Rogers from KBC and we're really grateful for his insights into the rapidly evolving world of industrial AI. As a reminder, if you're a subscriber to Hydrocarbon Engineering you can read Simon's article from our September issue by logging into your account over at our website, hydrocarbonshineering.com. Thanks for listening and we'll be back with another episode soon.
Advert:This episode of the Hydrocarbon Engineering Podcast is sponsored by Sherwin Williams Protective and Marine. Sherwin Williams Protective and Marine delivers world class industry subject matter expertise, unparalleled technical and specification service, and unmatched regional commercial team support to customers around the globe, including in the energy market. The company's broad portfolio of high performance coatings and systems, including protected liquid and powder coatings as well as fire protection coatings, excel at combating corrosion and help customers achieve smarter, time tested asset protection. For more information, visit protective.sherwin.com.