Podcast: AI Technology Autonomously Optimizes Complex Chemical Processes

To discuss the winning innovation — Factorial Kernel Dynamic Policy Programming, or FKDPP, a reinforcement learning AI developed by Yokogawa and the NARA Institute of Science and Technology and applied by Yokogawa to process industries — as the first reinforcement learning AI to autonomously control complex chemical processes, FKDPP complements manual and conventional control methods like PID and advanced process control. It allows plants to balance objectives such as product quality and energy use without labeled data or lengthy trial runs, building robust models in just 30 learning trials and reducing implementation time. One of our judges noted FKDPP's rapid learning capability, as it is at the heart of an AI algorithm that brings a new level of autonomous optimization to chemical operations, particularly ones with hard-to-control dynamics.

To better understand this award-winning product, I want to introduce Karthik Gopalakrishnan, part of the digital transformation, smart manufacturing, artificial intelligence, cybersecurity and industrial automation team at Yokogawa. Congratulations to you, Karthik, and your entire team at Yokogawa.

Karthik: Thank you, Traci.

Traci: Let's start out. I want to get a little bit of background on you and learn a little bit more about your affiliation with the technology. Can you tell us a little bit about what you do?

Karthik: Thank you, Traci, and thank you for this opportunity to speak to you about our technology. Super excited about this. I'm Karthik Gopalakrishnan, a solution consultant with Yokogawa. I've been with the company and the industry close to 21 years now. I started off at the DCS application all the way back in 2005 and moved up the ladder. The last five years, I've led the digital transformation initiatives, predominantly catered toward our external customers in North America. One of our flagship solutions, I can call it, is the control AI, which is FKDPP-based technology, and we are super excited to bring it to North America and give our customers this opportunity to evaluate it.

Traci: Well, we are super excited to have you on board as one of our Vaaler winners, and I am looking forward to learning a little bit more about it. I know I read the application, obviously, but talking to you about it is going to give me a little bit of better insight. So let's go ahead and get into that. Can you explain this programming algorithm's rapid learning and how it can be incorporated into existing process controls?

Karthik: Sure. Our AI control technology, as you already stated, is based on FKDPP, which predominantly stands for Factorial Kernel Dynamic Policy Programming algorithm. Moving on, I am going to use the abbreviation, and I hope you and your listeners are OK with that. So FKDPP is a reinforcement learning algorithm, and the key of FKDPP is how fast and how quick it can learn — as one of the judges already stated. The key of that technology is to make sure there is no overdependency on simulator accuracy. Mainly, we can use this technology predominantly based on the historical behavior of the processes and make sure we can control the final element.

Just to give a little more detail about this technology itself: The reinforcement learning agent, which in this case is the controller, interacts with the environment — in this case, the process or the equipment that it wants to control. By this interaction, it learns how to control the system by a mechanism of trial and error. So predominantly, the whole goal here is to make sure it knows the pattern and it knows how to react based on the behavior of its parameters.

Also, one more thing I want to definitely mention here is the application or the algorithm itself works more like a reward-and-punishment approach. Basically, if the solution or if that particular application is going toward the desired need, then it is rewarded, so that agent is actually getting rewarded so it knows it's doing the right thing. But if it is going away from the control or the desired outcome, then it is punished so that it also knows that it is not the right approach. So that's predominantly the whole gist about how this technology works.

Traci: Tell me, how does it get punished?

Karthik: So it is actually coded into the technology itself. At the end of the day, what we are telling this particular application is you have a boundary; you cannot cross beyond those boundaries. So there are some good guardrails defined for the control parameters. As it is going away from the designed outcome, we are giving a specific coded factor to it, so it knows that if that count is increasing, it is absolutely learning itself that it is not something where it is desired to go, and it is automatically course-correcting itself. So that is how we look at it as a punishment.

Traci: Interesting. And tell me, how does it differ from existing supervisory controller platforms?

Karthik: Well, first thing I want to state is, for sure, everything has its own place in the control world, right? If you specifically look at FKDPP: When it comes to our traditional PID or even the advanced process control, our PID predominantly is straightforward. At the end of the day, it's a feedback system. It looks at the error; it knows what the error is, so it'll do a course correction and take care of the correction. But it is really hard to control fast and rapidly changing processes. That's always been the problem. Now, when it comes to APC, there is a little bit of intelligent optimization added on top of it, but it still relies on accurate models and periodic human oversight.

This is where the difference comes in. FKDPP predominantly represents the next generation. It leverages AI to achieve what traditional technologies cannot do, especially in environments with high variability and nonlinear or even unconventional models. It is really hard to just use APC and PID to achieve the desired outcome, and that is where the difference comes in when it comes to FKDPP. One thing I also want to mention is, as stated, it's always a complementary technology. You still have a PID and APC in a process plant, but FKDPP comes in very uniquely, especially for things which typical PID or APC is not solving or cannot solve.

Traci: Now, you're saying it's complementary. Are there integration challenges that need to be considered?

Karthik: There are always integration challenges in our world, to be frank, Traci. Having said that, the way we have designed the technology is actually very, very light footprint. So we have developed the model on an offsite system. Basically, the models are developed and the agents are trained. And when it comes to the control system itself — obviously, as you mentioned, my background is a little bit of cybersecurity also — I'm not obviously very comfortable opening up all the way my control system into a cloud application. So we have developed an agent which is a very, very light footprint, which can install on a simple application, and all it needs is an OPC communication with the DCS, and that will take care of it. So far, we haven't faced a lot of challenges with all the POCs and the projects we are going through. That said, I will never say you'll not face challenges, but so far, so good.

Traci: Well, we trust you when you say you won't face challenges. If you're telling us there's no challenges, we don't have to worry, right? Now, you've kind of been talking about this already, but how did you develop this technology?

Karthik: So as you probably stated earlier also, Yokogawa collaborated with the Japanese graduate university, the NARA Institute of Science and Technology. The technology was originally used for robotics. Reinforcement learning is predominantly popular for robotics technology. So Professor Matsuhara, who is the lead of Robotics Learning Laboratory, developed with his students a reinforcement learning algorithm called KDPP, which obviously is the Kernel Dynamic Policy Programming, that was able to train robots with many degrees of freedom in a short term. And that was actually the key for us, because especially in the controls world, there's so many different parameters which move and change very quickly. So we wanted to use that technology so we can control those particular parameters as quickly as possible, and that definitely was the key for us. So what we did is we worked with the NARA Institute of Science and Technology and codeveloped this Factorial Kernel DPP technology, which is the algorithm that is predominantly catered for industrial-specific use. So that's how we developed that.

Traci: And can we talk a little bit about deployment time, standard deployment time? Obviously, there's going to be anomalies and outliers out there, but is it weeks? Is it months? Does it take more than that?

Karthik: Well, although I'm known for being super optimistic within my company and my peers, I want to keep that away from me. But the reality is — I'll try to give a long answer here on purpose, mainly because I want to explain the complexity. Yes, it will go as a proof of concept. The key here is to make sure, no matter what technology we use — and especially with AI technology — we want to make sure it's super safe for implementing that into a control environment, completely tested and vetted and validated. And that is what we want to strive toward. So typically, although modeling and engineering time could be a little faster, we are looking at something close to one to two years maximum, obviously based on various different complexities. And we have seen customers trying to push it faster, which may or may not happen, but we want to make sure it is safely and securely developed for that particular use case. So we are looking at something between one and two years.

Traci: And it takes that time to understand and to get those rewards and those punishments to make sure that everything is moving the way it should be. So that's understandable.

Karthik: Absolutely, yeah.

Traci: Can you discuss some of the real-world applications and your field tests that you mentioned in the documentation you gave to us, and how this could be applied in the chemical industry, how all of this works?

Karthik: Sure. Let's start with probably our first success story, I would say, which is the Ube material field test. The application was initially started as a POC. We were obviously in constant touch with them. We sort of knew exactly the pain point they were going through, so we wanted to support them and see what are the latest and greatest technologies we have under our umbrella that are going to solve the problem. So one of the main pain points they had was a distillation column, and the supply of waste heat and steam needed to be tightly controlled to efficiently separate and extract high-purity butadiene. So that was a real challenge they had. They were not able to extract the high-purity butadiene, mainly because the fluctuation in the liquid level within the distillation column was negatively impacting the quality of the extracted butadiene, which mainly is caused because the temperature variation and the boiling point and the extraction point are very close to each other.

So what it really meant was the operator had to closely watch that particular process and constantly change the control valves and the set points for that particular controller. If I'm not wrong, I think they recorded that every 10 to 15 minutes, the operator had to adjust the set point of that controller. And that system, that process, was running for years even before we went there, which was fine when they were doing it, but what ended up happening is the operators started getting eventually overloaded. There's so much operator action which had to go into controlling this facility. So they said, "OK, hey, can you really help us here?"

So what we really did is use this technology. We understood that there is a limitation, especially on the DCS side, which the typical PID could not do. So we developed this FKDPP and took in all the variables, including the ambient temperature, which is the environmental factors there, and developed it using FKDPP. So at the end of the day, the outcome was there was year-round stability, even with external temperature — as I told you, hot or cold. And in Japan, yeah, it's a beautiful country, but it can get a little hot also sometimes. It was almost 40 degrees Celsius there.

The other value addition was that we were able to optimize energy and also reduce time and cost while maintaining quality, enabling almost up to a 40% cut in CO2. And that was actually calculated and approved by the customer itself. And I would say one of the key wins also was we were able to lighten workload and improve safety, which is one of the factors there. And again, now the customer is so impressed that they've been running this for close to three years now. The initial POC was 35 days, and then they liked the technology. We implemented it as a full-blown project. Now they've been running it for three years, and they started looking at expanding it to different parts of the site as well. So that's one of the use cases I want to talk about.

The second one is very, very fairly hot news. Not sure if you followed the news, but we had a very good success story with Aramco. And the uniqueness of this opportunity and the POC was where we were able to use multiple RL agents so that they were complementing each other and controlling the facility. So that was the uniqueness. Again, as I told you, it's very, very fresh, hot in the press. I think, if I'm not wrong, two weeks ago we came out with a success story. So definitely more to follow on it as we start discussing more opportunities on those lines with customers like Aramco.

The third one — although I want to say it's my favorite, maybe it's not fair to say — but it's an example for a food and beverage company where we actually did a POC for a craft beer case where the beer manufacturer, based in Japan, had an issue, especially monitoring and managing the temperature and the time it took to ferment the beer. So they typically were looking at almost 336 hours to ferment the beer. But with using the application we developed, we were able to reduce the fermentation time from 336 to 240, which is almost close to 30% reduction, which was a big thing. So you all get a lot more beer a lot faster here.

Traci: Well, I can see why that's one of your favorites. It would be mine, too. Let's talk a little bit about AI. The chemical industry is fairly conservative, and AI is still fairly new across the sector itself. How is this being received? And is there a hesitancy toward adoption? You have all of these proofs of concept happening and all of this great news happening, but is there hesitancy because of the lack of knowledge, maybe?

Karthik: Short answer is yes, but there's — to expand more — for good reason. Again, I'm not the one who's sitting in a customer position and writing the checks, because at the end of the day, they need to make sure their facilities are operated safely without any interruption. And the last thing they want is any experimentation of any technology, which is going to happen no matter what — it's AI or non-AI, no matter what. Having said that, if I can dial back the clock probably a couple of years ago, there was a lot of push toward using AI more on the advisory layer. So, yes, take all my data, import it to a data lake and then show me all the insights, show me recommendations. Is there any operator action you would recommend?

One of the classic examples is preventive maintenance. So that is a great way of using supervised learning, which is an AI methodology or part of the AI umbrella, to see the patterns, if there's any anomalies for any asset issues and, more important, giving recommendations for the maintenance. So again, at the end of the day, it's not a closed-loop system; it's an advisory layer; it's an open-loop system. So customers started feeling more and more comfortable. So obviously, multiple vendors, including Yokogawa, have started rolling out solutions and technologies like that.

And now we are into the closed-loop systems where we are using technologies like reinforcement learning and making sure it is definitely working within the system safely and securely to make sure we are doing something an operator was not able to do for all this time. So more and more use cases, more and more value propositions are definitely going to help. It's always toward the right direction. So I'm a big proponent of latest and greatest technology. There's always value there, but we have to do it safely and securely, and that's where we are right now.

Traci: Absolutely. Safety and security are the key of the game. How far do you think we are from true autonomous control, and what role do you think that this technology will play?

Karthik: Well, it really depends how we define autonomous control. In general, autonomous control is something which is eventually going to be part of our life. The way I would really see this, Traci, is yes, there will eventually be solutions and technologies like FKDPP, and Yokogawa would definitely be one of the leading companies in developing technologies like this where it can start taking actions autonomously and start looking at some pieces of autonomous operations. I would also be sure to mention that there will always be some element of human intervention. What it may say is, "Hey, yes, you're doing your actions autonomously, which is perfect, but there's always human validation some of the time." So there is going to be some pieces of it.

Also, I also want to mention, as more and more use cases — and again, I'll go back to the example I mentioned about Aramco — we actually had an agentic AI approach here where we had multiple AI agents working independently. So now we are starting to look at more and more situations and scenarios where more AI solutions and AI use cases will start talking to each other and start being a little more autonomous. At the end of the day, you still need a human in the loop to make sure those AI are not going haywire and going toward AI hallucination, which is a real thing. So it really depends how you define autonomous operation, but most importantly, is it making humans more safe using technologies like AI? We are absolutely there, and we are going to get there more and more.

Traci: Good points made, and the hallucinations need to stay in line for sure. Anything you want to add that we didn't touch on?

Karthik: For sure, I didn't think about it, but let me throw some out. I think in general we covered most of the points, but maybe some things I would definitely add, Traci, is this solution, this technology, is here to stay. It's not a bubble, at least in the automation world, I would call it, because we are always super cautious on using latest and greatest technologies. Yes, there is a big push toward AI, especially in the IT space, and a lot of use cases coming up. But the beauty about the industry we are in — yes, we want to adopt new technologies, but only if we need to. And most importantly, is it safe, and we want to make sure there's good return on investment for it. So technology is here to stay. We're really scratching the surface here, so I would expect more and more use cases to come out and, more importantly, more and more value propositions, especially using technologies and solutions like AI in this space. So that's where we are, yeah.

Traci: Well, any technology that brings beer to the market faster is always good in my book. Karthik, I want to thank you for the insight that you've given us today. It was a great conversation. It's just something that, as you said, technology is here to stay, and it's always good to learn what's happening and how we can further help our readers and audiences and listeners adapt to that new technology and apply it best to their situations. Listeners, if you want to stay on top of innovation in the chemical industry, subscribe to this free podcast via your favorite podcast platform to learn best practices and keen insights. You can also visit us at ChemicalProcessing.com for more tools and resources aimed at helping you achieve success. On behalf of Karthik and the team at Yokogawa that helped propel Factorial Kernel Dynamic Policy Programming to the 2025 Vaaler Awards status, I'm Traci, and this is Chemical Processing's Distilled podcast. Thanks for listening.