My father was driving a cab in New York City when he started OLI Systems in 1971. Between fares, he worked on a single belief: that chemistry was knowable from first principles. That if you were obsessive enough about the science, you could predict what a system would do before it did it. Fifty-four years later, about 400 companies use what he built to make decisions in refineries, mines, and chemical plants where getting the chemistry wrong shows up fast.
I came to this company eight years ago from a hedge fund. Twelve years watching e-commerce disrupt retail taught me one pattern: powerful new tools enter industries faster than the foundations can absorb them. The companies that survived understood what was irreplaceable about them. Most didn’t figure that out until it was too late.
That’s the question I keep turning over as AI moves into industrial operations. I’m not skeptical of AI. The companies furthest along are doing work I find impressive. The problem isn’t capability. It’s what happens when statistical models operate without a physical understanding of the systems they are optimizing. In conversations with senior process leaders and in our DOE work on critical materials, I keep seeing the same gap. Not in AI capability. In what sits underneath it.
In industrial chemistry, the underlying data is far thinner than most people realize.
What AI is inheriting
There’s a knowledge transfer happening in industrial operations right now that doesn’t show up in any technology roadmap.
The engineers who built and ran these processes accumulated something over their careers that isn’t in the data. They knew which crude blends behaved unexpectedly. They knew where the models ran out. They knew that a particular combination of conditions would produce a corrosion mechanism nothing in the historical record would flag, because they’d seen it once, fifteen years ago, in a different plant. What they understood, often instinctively, was when the system had moved into chemical territory the statistical model had never seen.
They are retiring. What AI is inheriting is the data they generated, not the instincts they developed.
“The concern isn’t that AI is incapable. It’s that AI is confident”
The concern isn’t that AI is incapable. It’s that AI is confident. A statistical model trained on historical operating data will tell you everything is fine right up until conditions shift outside the envelope it was trained on. It doesn’t know what it doesn’t know. The engineer who built the process often did.
I was talking with some of our engineers recently about a refinery case that stuck with me. A plant had switched to a new crude blend. Within days, corrosion in the overhead lines accelerated beyond anything the operating team had seen. The plant’s statistical model saw nothing wrong. Every parameter read within spec. The optimizer kept running, statistically confident, physically headed toward an unplanned shutdown. What the model couldn’t see was how chloride distribution, acid formation, and neutralizer chemistry had shifted under the new feed. It pattern-matched against conditions that no longer existed.
When they ran a physics-based validation against the new feed composition, the problem became visible. The thermodynamic layer flagged the chemical conditions that made the corrosion mechanism possible. An electrochemical corrosion model built on top of it predicted the rate and severity. The team had time to intervene before the plant failed. The avoided downtime and equipment damage came to roughly $11 million.
The statistical model wasn’t wrong. It was ungrounded.
When AI operates without physical validation, it gets things wrong with confidence, at scale. Not through obvious failure. Through well-formatted recommendations that violate the underlying laws of physics.
What “grounded” actually means
The simplest way to put it: a statistical model needs to have seen conditions like yours before. A physics-based model doesn’t.
You feed it the actual composition, temperature, and pressure of your system and it calculates what will happen. What precipitates. What corrodes. What phases form. What reactions dominate. It solves from the underlying physics of how ions, molecules, and phases behave. It doesn’t need to have seen the conditions before.
Theory alone isn’t enough. What makes these models useful in practice is 54 years of validation against experimental data across thousands of chemical species, from ambient water treatment to extreme mining and energy brines. The physics provides the framework. The accumulated data provides the calibration.
Because that thermodynamic foundation spans multiple industries, insights learned in one domain transfer into another. Work done in oil and gas brines informs critical minerals processing, compressing the data deficit that first-of-kind projects usually face.
When I arrived at OLI eight years ago, I assumed the data work could be modernized. That with the right tools and the right approach, we could systematize what had been built and move faster. It took time, and a lot of conversations with our chief scientist, to understand what I was actually looking at. The data isn’t waiting to be downloaded or cleaned up. In many cases it had to be built, measurement by measurement, across decades of experimental work in conditions no simplified lab system captures. What looked like a data modernization problem turned out to be something closer to institutional knowledge encoded one hard case at a time.
“What looked like a data modernization problem turned out to be something closer to institutional knowledge encoded one hard case at a time”
Open-source data tends to come from simplified lab systems with limited species and controlled conditions. Real industrial brines involve mixed chemistries, extreme concentrations, and interactions that those datasets simply don’t capture. That is the fundamental advantage of a physics-based model over a statistical one. The underlying physics doesn’t change when you move from a simple system to a complex one. Models calibrated on controlled experiments can predict behavior in conditions that are far messier than anything a lab was designed to test.
To be clear about where this applies: it’s strongest in equilibrium thermodynamics and aqueous chemistry, the domains where robust, industry-validated models exist. In stable processes with consistent feedstocks and good instrumentation, statistical models work well and physics grounding may not add much. The problem starts when conditions change.
Physics-based models were historically used at the design stage. Set the operating windows, select materials, size the equipment, then hand the system to statistical controls.
What we’re building toward is physics in the operational layer: a thermodynamic engine sitting between an AI optimizer and the control system, checking proposed setpoints against physical reality before the system acts. Valid states pass. States that violate chemistry get flagged before anything moves.
That change in approach matters most where AI is moving fastest: complex, shifting chemistry where the historical record is thin and the cost of a confident mistake is high.
Those conditions describe much of the emerging critical minerals industry.
Where this is being tested at scale
The U.S. and its allies are trying to build domestic supply chains for lithium, rare earths, and cobalt. The pressure to move fast is real. The chemistry is difficult. Most of these projects are first-of-kind, which means they’re data-poor by definition. That’s not a gap you close with more sampling. It’s a structural condition.
That’s the worst environment to deploy ungrounded AI. It’s also exactly where it’s being deployed fastest.
The DOE recently launched the Genesis Mission, an initiative to build an AI-driven discovery and scale-up pipeline for critical materials. The collaborators include Argonne National Lab, Idaho National Lab, Oak Ridge, leading AI platforms, and OLI. We’re in that room for a simple reason: when AI proposes a process, someone has to check whether the chemistry is real. For a project of such national importance, that check isn’t optional. In many of these systems, the datasets needed to train AI models simply don’t exist yet.
We’ve been part of this ecosystem since 2013 as a founding partner in DOE’s Critical Materials Institute, working with leading institutions to turn research into thermodynamic models that pilot plants and industry can actually use. The goal was always the same: encode chemistry knowledge in forms that transfer, so a pilot plant doesn’t have to rediscover what a national lab already characterized.
China’s advantage in these chemistries isn’t just scale. It’s thirty years of accumulated operational learning, built through decades of failed projects, process adjustments, and hard-won knowledge of how these systems actually behave. You don’t compress that by running faster. You compress it by encoding the chemistry in models that carry that knowledge forward, and by not burning capital on processes that violate the underlying science before you find out at scale.
What I keep coming back to
The companies furthest along in AI adoption aren’t asking whether to use it. They’re asking how to go fast without losing the ground beneath them.
That’s not a technology question. It’s a knowledge question. What do you actually understand about your process? Where does your model run out? What did the engineers who built this know that never made it into the data?
Those questions were always important. AI makes them urgent, because the cost of getting them wrong now compounds at speed.
My father believed you could predict chemistry from first principles if you were obsessive enough about getting the science right. Fifty-four years of validation later, that belief has held across conditions nobody anticipated when he was coding up solutions in FORTRAN between cab fares.
AI will be everywhere in industrial operations. The question is where it remains grounded in physical reality.
At what point are you willing to let statistical confidence outrun physical certainty?