Nothing in the past decade has created the same levels of excitement, ideation, and trepidation that we have seen with the rise of artificial intelligence (AI). In reality, there have been very few technologies in the history of the world that have risen in popularity as fast as AI. As a result, not only is everyone talking about the AI future, but we’re already seeing a shocking number of new AI technologies entering the marketplace on any given day.
Yet, while AI brings many exciting possibilities, as Claudio Fayad and Krishnan Kumaran explore in their recent article in Control Engineering, not every AI technology is suitable for process manufacturing, or industrial applications in general. Rather, operational technology (OT) teams must be selective in their application of AI solutions, opting for fit-for-purpose industrial AI software.
Industrial AI
One of the key problems of generic AI engines is that they are incredibly expansive. The typical GPT models the public is used to working with are as prepared to answer a question on dog obedience training as they are to answer one on a hydrocracking unit. This leads to two critical issues. First, those models need incredible data capacity and scalability, meaning that they must be hosted in the cloud. Due to security constraints and latency issues, cloud connectivity is often a non-starter for OT teams.
More importantly, however, those massive AI models are often unreliable. They pull data from so many places—all contained within a black box—that it’s hard to have confidence in their results.
Industrial AI is different. It is intentionally constrained to content necessary for its task. Claudio and Krishnan explain,
“Immutable first principles constraints— based in detailed physics and chemical data built into the hybrid models—guide and instruct both the way Industrial AI models are trained, and the potential results those models can generate. Those guardrails create a safety zone where the AI can operate, ensuring the AI does not simulate scenarios or build models the OT team would not want to run because they are impossible, dangerous, or expensive. The world is too big to explore every possibility, so Industrial AI’s exploration space is limited to realistic, safe scenarios.”
Limitless models are a great tool when the results are low stakes. In the case of manufacturing, however, teams need more confidence in their guidance, and that’s where industrial AI is the right tool for the job.
AI-driven agents
When a team has access to constrained, fit-for-purpose AI models, they can seed those models with data specific to their environment for a more customized experience. When that happens, it becomes possible to deploy site-specific and mission-specific agents within the AI to perform specific tasks.
“Today, Industrial AI agents use specific and relevant data sources to solve a wide variety of problems. A common example is seen in the reliability space, where enterprise-level reliability solutions are seamlessly integrated with predictive and prescriptive asset health software to create comprehensive asset health solutions. These systems rely on fit-for-purpose agents built on extensive failure mode and effects analysis databases to make advanced analytics more intuitive.”
For example, Emerson’s AMS Optics works in tandem with AI agents built into Aspen Mtell® to help reliability teams more effectively monitor asset health and performance, predict asset failures, and prescribe next steps. The AI tools in AMS Optics with Mtell can use pattern recognition based on contextualized historical data not only to warn operators and technicians of impending failures, but to give them step-by-step guidance to remedy that failure and identify root causes.
But AI agents are not only useful for reliability. Engineers are also starting to see benefits to project design and implementation as Emerson incorporates AI technologies into their software,
“AI agents can also be leveraged for process optimization. As teams engineer processes, Industrial AI helps them see multiple alternatives and factor in multiple criteria to develop a range of possibilities to select the best design more easily. With the help of Industrial AI, hybrid models have become more robust and compatible with existing equipment and plant designs. As those models are further refined, they can become AI agents of their own, helping test and refine operations as use-case-specific AI models built for a more granular purpose, such as those designed for heat exchangers or distillation columns.”
And that’s just the beginning. As Emerson continues to pursue Project Beyond toward an enterprise operations platform, engineers are bringing AI software to the control layer where it will be run on software-defined control technologies. All of these technologies will be game changers, and all of them will be bounded in the first principles design inherent to industrial AI.
