AI Chatbots Can't Optimise Industrial Plants. But Physics-Informed AI Can

For decades, the world's most energy-intensive industrial plants—ammonia facilities, refineries, and petrochemical complexes—have operated on the same fundamental principles: experienced engineers, tightly controlled processes, and hard-won operational data that rarely left the facility walls. 

But, like every industry, they are not immune to AI-led disruption. AI can cut operational expenses and emissions through real-time data analysis, predictive maintenance, and automated process adjustments.

Engineering giant KBR and AI company Applied Computing are attempting something similar with their physics-informed AI platform called INSITE 3.0. Powered by Applied Computing's Orbital AI model, the platform embeds foundational energy models into KBR’s process engineering expertise to optimise plant operations, maximise yields, and reduce emissions. 

But while conceptualising the platform, the companies were clear that they did not want to build another AI chatbot.

"Large language chat prediction is useless for an engineer who has to sign off stress cases or make safety-critical decisions around technologies that are really important," Andy Webster, Senior Director of Digital - Sustainable Technology Solutions at KBR, which offers science, technology and engineering solutions in energy and industrial sectors, tells AIM. 

The concern is not abstract. Webster recounts a customer at a significant company presenting work to its board that contained errors because it had "flown through language prediction with no understanding of what those words actually mean in our business." In a world where plants contain explosive gases, operate at extreme temperatures, and serve as critical national infrastructure, an unreliable recision rate can turn into a disaster. "There's no way you can turn a valve on a live plant at 70% sure," Webster quips.

Dan Jeavons, President of Applied Computing, frames this as the central bottleneck the partnership was built to solve. 

"The biggest challenge is bringing together two worlds that are very different. On the one hand, the world of AI, which is moving very quickly, but is often relatively generic... and then trying to integrate that with the world of energy, which is very specific. In general life, AI can be 80% and it's still going to be very useful. That doesn't work [in industries]," he adds.

The Data Problem Inside Plants

Building an AI system grounded in physics—rather than in statistical language prediction—demands a different kind of training data, and obtaining it proved to be one of the project's greatest hurdles.

The challenge begins with the nature of industrial plant operations themselves. These facilities are engineered for stability, not variability. As Webster explains, the data that emerges from day-to-day refinery operations tends to be extraordinarily repetitive: "The goal of operators is not to move." When the same process variables oscillate within narrow bands for months on end, the resulting dataset tells an AI model very little about where things can go wrong or what happens under unusual conditions.

To fill in the gaps, KBR had to combine real operational data with engineering simulation—using its decades of design expertise to synthetically generate the edge cases and failure modes that real plants rarely produce. "We had to use the real world piece to make sure we knew what's really happening, but we had to use our side as designers and think about what's the safe operation to give it the fuller context," Webster explains.

Beyond the technical challenge lay a legal one. The energy industry's operational data is treated as a matter of national security in many territories, and both companies had to construct careful legal frameworks before any information could flow to the AI system. 

"We have shown nothing to the AI company," Webster states plainly. "We've kept a very heavy wall until we're confident we have the legal right to use this data." That caution slowed development, but neither company was willing to compromise. "We can't be naïve with our data exposure."

What the Platform Actually Does

At its core, INSITE 3.0 works by encoding the physical laws and engineering relationships that govern plant behaviour, then learning from both operational and simulated data to model how variables interact within those constraints. Because the platform understands the physics, it can reason about conditions it has never directly observed.

"Once you understand the interaction of those variables and the physical laws that constrain them, if you add a new parameter to it or a change of condition, you can predict how those variables will move," Jeavons explains. "Uncertainty and volatility don't necessarily scare us—and in fact it's good for business, because the customers don't have things that can deal with it in this way."

The practical application, at least initially, centres on energy optimisation. By identifying inefficiencies in how a plant runs, the platform can help operators extract more product for the same energy input, or reduce energy consumption while maintaining output levels. Given that some of these facilities rank among the most energy-intensive operations on earth, even marginal gains translate to substantial reductions in carbon emissions. 

"If you can do that by 1%, 2%, 3%, the savings in terms of CO2 [emissions] are absolutely enormous," Jeavons notes.

The Road from Ammonia to Capital Projects

The partnership moved at an unusual speed. 

KBR and Applied Computing met in August 2025, signed their first contract in November, and within six months had brought INSITE 3.0 to market for ammonia customers—already with a paying client in hand. In parallel, they built a data connectivity platform and an engineering simulator, the latter becoming essential once the limitations of purely operational data became apparent.

But ammonia optimisation is only the beginning. The companies have set their sights on transforming how large capital projects are engineered. Today, taking a new technology such as an ammonia plant from initial concept to completed engineering design typically takes around three years. The vision is to use AI to compress that timeline dramatically, allowing customers to respond to market conditions faster and at lower cost.

Beyond existing technologies, the platform's ability to generalise from known components—pumps, heat exchangers, reactors—to new configurations opens the door to accelerating the development of entirely novel technologies, from carbon capture to sustainable aviation fuels to plastics recycling. The logic is straightforward: if the AI already understands how a pump behaves under thousands of different conditions across dozens of existing plants, it can make confident predictions about how that same pump will perform in a technology it has never encountered before.

Earning Trust

For all the technological ambition, both companies are acutely aware that the final barrier to adoption is human. The engineers and operators who run these facilities have, in many cases, spent their entire careers within a single plant. They are not looking to be disrupted; they want to be supported.

"These guys are really vested in their plants," says Ron Populin, Technical Director at KBR. "Presenting them with a change that's understandable—that's really the key. These aren't whiz-bang operations. These are people who are running very complex plants, and they're very calm and very stable in the way that they do it."

The approach, then, is as much about communication as it is about capability. It is, in Webster's phrase, about being a "translator"—bridging the world of rapidly evolving AI and the world of people who would quite reasonably prefer their plant not to explode.

The ambition, ultimately, is not to replace those engineers. It is to equip them with a system that can tell them not just what is happening in their plant right now, but what might happen next—and what they should do about it.