Honeywell, Nvidia chart new path for industrial autonomy

The process industries have made steady progress toward more advanced process control over the year. However, at this year's Honeywell Users Group gathering in Phoenix, a more fundamental shift was on the table. In a joint session, Rahul Negi, director of AI and industrial autonomous operations, Honeywell Process Automation, and Ahsan Yousufzai (pictured), global head of business development and strategy, Nvidia, made the case for physical AI—artificial intelligence grounded in the laws of physics, chemistry and thermodynamics—as a significant leap forward for the industry.

Before making the case for the new technology, Negi explained industry to a packed room of Honeywell’s partners and customers that four structural challenges are converging simultaneously. The first is workforce attrition. "We have seen in the energy industry that many workers in every advanced economy are near retirement age, and nearly every new worker entering the field is below 25," Negi said.

The experience gap created is driving up human-error-related upsets, unplanned shutdowns and production losses. Second, more than half of process industry assets are managed by teams working on equipment older than 20 years, with aging infrastructure feeding unscheduled downtime. Third, capacity utilization across the Americas is running near 90 percent, the point at which traditional optimization approaches have largely exhausted their incremental gains.

"Small step changes brought us to where we are," Negi said, "but they are not going to get us that next step forward."

Finally, market volatility now demands a speed of response that human-led planning cycles simply cannot match.

Operational journey toward adaptive operations

Negi placed physical AI within a 50-year evolution: from manual operations in the 1970s, through DCS and basic automation, to intelligent operations enabled by analytics and digital connectivity. Remote operations further consolidated expertise across distributed assets, but each step led to information silos.

"We created a lot of isolated solutions," Negi said, "and how do you now achieve that next move toward adaptive operations?"

Adaptive operations—and ultimately autonomous ones—are the destination. Physical AI is the enabling technology for the transition.

What exactly is physical AI?

For process engineers accustomed to data-driven models, the term needs definition. "Physical AI has something called a deterministic core," Negi explained. "The laws of physics, chemistry and thermodynamics, as well as mass and energy balances, are known to us, and the AI learns within those boundaries."

The concept is different than the AI most people are currently familiar with, which tends toward large-language models (LLMs) and agentic AI, that have no inherent understanding of why a distillation column cannot violate component mass balances or why a compressor has hard operating limits, for example. With physical AI, design limits are respected, reliability constraints are honored and decades of engineering standards become guardrails embedded in AI.

Meanwhile, the practical benefits are significant. AI operates within a physically consistent envelope so it can anticipate upsets before alarms are set off rather than reacting afterward. It adapts in real-time to feed composition changes, temperature excursions and pressure disturbances. It also coordinates across the control and optimization stack.

"The AI can talk to other AI agents and make things happen in a coordinated way," Negi said. "That brings us the capability to move from reactive to predictive to prescriptive operations."

From conversational to agentic to physical AI

During his presentation, Nvidia's Yousufzai explained why the moment for physical AI is now. The first AI era—conversational AI—produced tools useful for knowledge work but largely tangential to control room needs.

"In the industrial space, we do workflows," Yousufzai said. "You just cannot change a setpoint without looking at your mass balance system and understanding what the next target should be."

The second era brought agentic AI, or systems capable of executing multistep workflows, reasoning across data sources and taking actions rather than merely generating responses. That capability is directly applicable to industrial decision chains.

The third era is physical AI, and it’s where process operations become the primary use case. Existing plant instrumentation such as acoustic sensors, vibration monitors, camera, and analytical instruments that are "already in your refinery but hardly used," as Yousufzai put it, become the perceptual layer for an AI that can maintain a real-time, physics-consistent model of plant state and act on it.

Nvidia's three-computer framework

Yousufzai outlined the platform architecture Nvidia developed to enable physical AI deployment across industries, built around three sequential steps.

Training is where existing plant data—historian archives, lab results, equipment logs—build the underlying models. Imperfect data is not a blocking issue, Yousufzai said. "There will always be some noise in the data. You let the compute figure it out and refine continuously."

Simulation is where Nvidia's roots in physics-fidelity rendering become directly relevant. "To bring one autonomous car to the road, you require 80 million miles of testing," Yousufzai explained. "But you bring those scenarios into a completely virtual environment and that model is so accurate you can simulate any scenario."

For process operations, this means testing AI-recommended control actions against a high-fidelity dynamic simulation of the unit under every conceivable upset before those actions ever reach the real plant. Reinforcement learning agents improve through repeated simulation cycles, developing expertise across scenarios that may occur only once a decade in real-world operations.

Inference is deployment at the point of operation. It doesn’t require cloud connectivity. "These models can run completely isolated, in a container, in a very secure way," Yousufzai said. "You do not need internet connectivity. You can run all of that intelligence on any device near to you."

For an industry where cybersecurity and operational isolation are non-negotiable, this architecture is a foundational requirement.

Honeywell, Nvidia collaboration

The most concrete signal for the engineers in the room was Yousufzai's description of the active partnership between Nvidia and Honeywell. "Our journey with Honeywell started a while ago—this is already work in motion," he said.

The collaboration pairs Nvidia's simulation and AI infrastructure with Honeywell's deep process control and safety domain expertise. Work already underway in smart building technologies is now being extended explicitly into oil and gas. The ambition is enterprise-scale deployment.

For Negi, the partnership also directly addresses the trust deficit that has historically blocked AI adoption in safety-critical operations. "The solution to the challenge of trust is visibility and explainability," he said. "Physical AI provides that because it operates within known physics boundaries. The operator can see why the system is recommending what it is recommending."

The message from both men was clear: workforce and asset challenges facing the industry are structural, while the efficiency ceiling reached by conventional automation is real. The convergence of physics-constrained AI, high-fidelity simulation, secure edge inference and partnerships between domain experts and AI platform leaders is opening capabilities that did not exist two years ago.