Aspen Technology’s Heiko Claussen on AI’s Role in Reducing Energy Consumption

Hello, and welcome to the latest edition of the Techstrong AI video series. I'm your host, Mike Vizard, and today we're talking with Heto Clawson, who CTO for Hesp been techer of technology demand, and how you wanna describe that. And we're talking about, well, everybody kinda knows that AI eats up a lot of energy at this point, but it turns out maybe AI can help us maybe make that whole energy grid more efficient.

Hi, welcome. The show. Great to be here.

Walk us through exactly what could be done here with ai, because I think people have it in their heads that the grid, um, maybe is more efficient than it is, but what is the opportunity right now? Yeah, so I feel that, uh, you know, there's a lot of, um, news around large language models and that, uh, particularly the foundational models are in focus here. When we are talking about AI driving energy consumption, it's a little bit of an oversimplification.

So, uh, it's certainly true that foundation model training is consuming a lot of energy. We are seeing a significant spike there, but also, uh, processing of large data sets. Now, you, you have your weather data images from volumes like subsurface imaging, medical imaging, uh, videos.

This could, uh, uses a lot of compute. Um, also if you really scale it like, uh, you know, image generation with, uh, and social media, uh, you know, networks from millions of users. So this is kind of what I think, uh, drives this, um, this, uh, assumption that AI is in itself, uh, very energy hungry.

That, in my opinion, is far from true, uh, that it is a net driver of energy, uh, usage. In, in other areas, particularly in the industrial SP space. AI is mainly used for optimization of operations.

So it's really there to make things more, uh, efficient, lower, uh, emissions and, uh, ways. So we are also having significant, uh, you know, less bandwidth in, in certain use cases. You have, uh, sensors, um, sampling rates that are like maybe once per second.

Uh, you know, the, the plant is running in an optimal, uh, environment. So it's, uh, you know, lots of the redundant information. It's, it's not always new, not always varying.

So it's, it's a little bit of a different problem. And, uh, as I said, so AI is really a tool for optimization in that field, and there are many use cases around that. Do you think that the ultimate gains in efficiency and from that optimization will far outweigh whatever the total consumption is for creating and training the AI in the first place?

Definitely for these use, use cases, uh, definitely because the models are also not that large. So, um, you know, if you, uh, see like on large language model that's trained with the, like, world knowledge of, of language kind of, uh, these are very, very large models with a lot of parameters. If you're talking about, uh, models that, uh, have maybe 20, 30, 40 inputs, uh, you know, the orders of magnitude, uh, less training data required, uh, and, and, uh, therefore they don't consume much, uh, energy to be created.

And when they're used to optimize equipment, particularly in the industrial space, like, uh, it asset intensive, uh, industries, you can significantly, uh, improve efficiencies. Imagine, uh, you know, you have like on first principle driven model, so using physics, um, to control a particular asset, there's always a simulation reality gap. So, uh, the actual asset in the field is never exactly representing, uh, all the aspects that you're assimilating.

Uh, so to close that you can use data from the field and, uh, and that then en, uh, enables a better representation and a closer control of that asset. Uh, if you, uh, reduce only by less than a percent, you have a huge impact in energy consumption and emission. And we have examples where our customers are using, for example, our hybrid models that combine these first principle knowledge with AI to improve, uh, yields of, for example, fluid catalytic crackers, which are really, really large assets, and then, uh, and are planned by about 10% increasing the yield.

Yeah. Um, as we kind of think all that through to the AI models themselves, this is the first generation thereof. Do you think over time though, just become more energy efficient in general as we get better at building and deploying them?

Yeah. So again, if we're just focusing on the large language models and foundation models, there's a significant trend towards, towards, um, uh, small, uh, language models. So, uh, there's a realization that, uh, you don't have to encapsulate all content for a particular application.

So if you are, uh, covering a more narrow space, again, you need less complex models. You can, uh, use less, you're more data efficient. Uh, also the inference is going to be, be more efficient.

This is a very important realization that, by the way, also something, uh, that we are building on when a set combining first principle knowledge and domain knowledge with AI knowledge is exactly that. So these first principles, they, they give you a, basically a playing field where it's physically possible that a plant or that your answer is going to lie in. So you don't have to learn the whole world.

You only have to learn to represent the physical, meaningful area of your particular domain, of your particular, uh, problem. And therefore, you're much more data efficient. It's going to be much more, uh, fast to train to adapt this model.

You're going to get better answers because you have a, a, a larger set of data for this particular, um, uh, problem. So a a lot of promises by combining this, uh, domain knowledge with, uh, with ai, How far down the path are we on this journey? Is it your sense that the people who build, uh, I don't know, or airplane engines and industrial control systems are, uh, mature in their use of ai?

Is it still early days? I mean, how long before we kinda get to this, uh, next wave of technology that does make our energy consumption overly more efficient? So, uh, AI has been around for a very long time.

We, for example, uh, have been using deep learning already more than 20 years ago in our products. And the same as industries, as you said, uh, engines for planes and others. So there's been, uh, machine learning, e-learning other a aspects of the AI already, uh, used for, for a longer period of time.

The essence is really, uh, and the, and the optimization, you know, how can we get the most out of these, these assets? How can we predict failures, uh, et cetera. So there are many, many use cases.

They're may be not as prominent to the, to the public. I feel that particularly on the large language model space, people really interacted and, and it was really magical to many people how this machine comes up with these answers that, uh, are unexpected. But, uh, in, uh, these models are in essence not new and are, in my opinion, key to solve these challenging problems that lay ahead.

There's, uh, many trends, like there's not enough, uh, expert, uh, experts there to scale use cases for necessary for energy transitions. You know, many times use cases become more and more complex. They become, um, uh, they, they, uh, become more, more, uh, specialized.

So scaling, in order to scale, you need this automation. You need to push down costs, drive efficiencies, uh, enable flexibility to adapt, uh, to, to these new use cases. And this is where AI is really key to unlock, uh, this capability.

Mm-Hmm. Is it your sense that each individual company will try to solve this on their own? Or will they kind of form many consortiums where they're all collaborating and pulling together their own research?

'cause it kind of benefits everybody equally? Well, I think, uh, the, it's in, in the end, uh, at collaboration is key. Uh, we are also collaborating, uh, with, uh, you know, our partners, like, for example, AWS Microsoft, there's no, uh, no value for the customer to try to reinvent everything.

So building up top of, uh, uh, collaboration really accelerates, uh, value to the customer. Uh, that being said, uh, the different, um, you know, companies then provide their secret sauce with their particularly good to and build, uh, basically and enrich this offering for, for the customer. Yeah.

We, for example, come with, uh, more than 40 years of domain expertise in, in our spaces and, uh, can significantly, uh, bring this to be here. Uh, if you, for example, uh, talking about industrial ai, so bringing these more generalized state of the art models and make them robust for our industries, uh, which is really key that they're trusted and adopted in these, uh, in the safety critical environments. So we went from, to your point, being shocked and odd last year by Gen ai.

And I can't help but feel it. Maybe we're already on the cusp of taking all of this for granted soon. So how quickly before we just kind of live in a world where we assume that AI has taken care of all these things?

Yeah, I, I think it's just the beginning. I think, uh, it's, uh, in a way something that, as I said, speaks to people because it's our language. But, uh, also here, generative AI is a much broader term than just in a large language, uh, space.

So I think that you'll see more and more the patient of other generative processes. In a way, it's, um, in a change of thinking. Um, you know, before, if you thinking for example, about the design process before you had an, uh, an expert that was tasked based on KPIs to create a specific design, and, uh, based on the expertise, then they, they found their way.

And compared to the KPIs, maybe had to iterate and simulate, and eventually the KPIs were fulfilled. And, uh, and, uh, you know, every, that was submitted as, as the end result. Nowadays with these generative models, you can provide these KPIs to a machine, and it can generate many or multiple options in the space.

And therefore, the expert can then, uh, can then be more aware of a broader set of, uh, opportunities of possible solutions and trade them off with maybe non-functional requirements that are not defined in the KPIs. So in a way, you, you have more scale, more speed, uh, and more agility, um, and are not becoming as dependent on a specific expert that may be, uh, is retiring out of that industry shortly. So ultimately, do you think that the, the management of a lot of these, uh, processes, whether they're in, I don't know, oil fields or wherever they might be today, there's a lot of specialists and they require a, a significant amount of skill.

Are we just gonna democratize all that to the point now where the, the folks who are managing all this, you know, maybe they don't need a master's degree to manage the refinery? Uh, I, I think that the, maybe the, um, uh, it's not that, that people will not be as educated going forward in the future. I think that there will be the different set of skillset and, uh, we just have to drive efficiencies to scale.

There's so much more we can do, uh, and, uh, if we really want to create this, uh, you know, sustainable future for us, then, then we really have to enable this scale. And for that, there's such an automation as through AI is, is in my opinion, necessary. It's not that, uh, uh, you know, these, uh, context systems will be run, uh, by not as well educated people in the future.

It's really just a different skillset. So what's your best advice to the folks who build these kinds of systems out there? What should they be thinking about today?

I mean, is it really just a matter of experimenting, or are there things they should be doing that are a little more advanced? I think there's a lot that, uh, uh, the teams can embrace already. It's, it's not new technology.

It's been in the market for a long, uh, period of time. Uh, traditional industries around asset intensive industries particularly are, you know, uh, cautious, particularly when they see what large language models are doing. You've heard aspects around, uh, uh, you know, uh, uh, that these systems are probalistic and, uh, maybe, uh, not as robust in the, in the replies.

So this, this realization that you're not just apply any AI that, uh, you find in the internet to your particular application is very important. But, uh, not to oversimplify and feel that, uh, you know, their AI is always like that. You can enrich it with these guardrails.

You can, uh, you know, uh, use non-black box solutions. It's very key in our industries that operators stay in control. Imagine you have a big plan and something happens out of maybe externally that that changes the situation.

And you have to react quickly. You have to be enabled to make these important decisions. So these aspects are really key in order to scale, uh, uh, AI in an, in an, in a reliable, robust way in into the industry.

And I think there's a lot of opportunity to be had. All right, folks, you heard it here. Well, AI is here.

In fact, it's already being used. But the question now is, well, how are we gonna supervise it? Hey, how come?

Thanks for being on the show. Great to be here. Thanks.

All right. And thank you all for watching the latest episode of Techstrong AI video series. You can find this and others on our website.

By all means, check them out. Until then, we'll see you next.