Large language models (LLM) and generative AI (genAI) are proving increasingly strategic for enterprises across all industries. However, there will be certain impacts of this revolution on the modeling frameworks of industrial companies in the next decade.
Will AI ever run plants autonomously?
Let’s start with the ultimate question: when will artificial intelligence (AI) take over and run plants autonomously? Looking far into the future, we can imagine AI constructing a plant and operating it autonomously without human intervention. The concept of a robot plant will likely happen 100 years from now. Could it be 50, or even 30 years from now? It’s hard to predict.
We can, however, grasp where technology might take us over the next decade. We predict that plant decision-making and modeling will be based on “controllable AI models,” LLMs and human domain experts, who will continue to make critical decisions.
Traditional plant modeling framework
A complex manufacturing plant includes five to seven major, technical decision-making domains, each built on technical expertise acquired over long careers. Examples include process engineering, maintenance, advanced process control (APC), control systems, process safety, control room operations, and planning and economics.
In each domain, information is received from the plant, and processed by a carefully engineered scheme of tailored models and human expert reasoning—to generate decisions—fed into other decision-making domains or plant processing equipment. Each domain uses a different modeling technique to describe overlapping plant areas, representing different “versions of the truth.” The reason for these different, often conflicting, versions of the truth is the computational limitations of modeling technology. Most process manufacturing plants operated for decades before personal computers were introduced.
Developed in a data-limited era, each model was tailored to solve a specific problem within a decision-making domain. Each domain developed its own set of first principles and mathematical assumptions to build parametric models supporting their automated and manual decisions. For example, planning and economics teams use a linear programming (LP) model to optimize complex, nonlinear plants. Applying linear solutions to nonlinear problems involves a delicate system of assumptions and procedures honed over decades to approximate, validate, maintain and adapt these models.
Over decades, the user community of plant technical staff from each domain built their careers around the modeling technologies that served them, honing their skills and developing methodologies around their modeling technology. Modeling technology providers delivered solutions tailored to their specific domain users. The effect of these supplier and consumer behaviors on today’s technology landscape is a series of domain-specific toolsets that are difficult to interconnect and don’t automatically communicate. All communication between the decision-making domains is done manually by the plant’s technical staff.
Traditional modeling encumbers genAI
Modern genAI is based on LLMs trained on large portions of human-generated data, and increasingly used throughout enterprises as co-pilots to support human decision-making. To get ahead of the curve, most companies push their limits of autonomous decision-making, exploring how their future operations will look with AI-assisted workers, and further down the road, as AI agents working alongside humans.
LLMs exhibit promising, human-like autonomous capabilities. They can acquire human collective knowledge and interact with the world in a way that mimics human interaction. However, LLMs are also fundamentally flawed in terms of reliability, safety, stability and explainability. Research is being applied to solve LLM flaws, and until these issues are solved, we don’t expect plant operating companies will let LLMs make autonomous decisions in mission-critical environments.
The probability of an industrial plant allowing AI to make decisions made by human technical experts is extremely unlikely in the next one or two decades. Before enabling AI automation, any AI decision must be transparent, explainable and verifiable by a human expert. AI operating on top of traditional, siloed decision-making domains must respect established habits and methodologies, mimicking human analysis and decision-making. AI on top of traditional modeling will provide value to each decision-making domain by capturing knowledge, organizing human decisions and discussions, and formalizing communication between teams.
However, applying AI within the confines of the traditional, plant-modeling framework will fail to unlock the power and value that modern AI can provide by pushing a manufacturing plant to its true global optimum. This approach will burden AI with the same assumptions, biases and limitations that plague traditional approaches. A truly optimized plant means that every action taken will likely best serve the plant's business goals within its constraints, given all the available raw plant data at the time of the action. That will require a holistically different plant modeling framework.
Next generation plant modeling via controllable AI
Increasing computing capacity and market needs give rise to a new type of AI, which we’ll refer to as “controllable AI.” Controllable AI models aren’t genAI or LLMs, nor are they hybrid models—traditional approaches with AI bolted on. Controllable AI models are intended to empower domain experts to optimize plants more closely with the plant’s business goals, within its constraints, and provide domain experts with more observability and controllability into the AI models.
Get your subscription to Control's tri-weekly newsletter.
Controllable AI models are constructed using an architecture that’s known to the user, and trained only on datasets known to the user. The user has full decision-making power over the precise dataset used for model training, including data sources and data points within each source.
Once trained, controllable AI models are well-defined deterministic, mathematical functions or algorithms. Key properties of the model are mathematically or statistically observable in a way that’s easily understood by users. They include certain engineering selections that letusers steer the model training to express, for example, a set of desired relationships between variables.
These traits of controllable AI give engineers more controllability and observability, which meet the harsh requirements of safety-critical environments such as industrial, process plants.
The 2035 equation
Human, domain experts plus controllable AI plus genAI equals improved plant performance.
For genAI to unlock its increasing strengths and opportunities, it must be unburdened by siloed human assumptions and biases developed over past decades by each domain’s experts. GenAI must generate decisions as close to the optimum as possible based on current plant data and one shared,verifiable version of the truth.
In the next decade, effective and safe use of genAI will be accomplished by integrating genAI with controllable AI. Using controllable AI, genAI serving all domains will develop the same understanding of the plant via shared, controllable AI models. This understanding of the plant will be validated, verified and approved by domain experts through controllable AI. GenAI will explain its decisions and recommendations using a sequence of reasoning, also known as a “chain of thought,” based on controllable AI models, which are shared and understood by all decision-making domains.
To return to our originally posed questions, we don’t expect any decision-making domain to vanish in favor of genAI, at least not in the next decade. Each domain will employ genAI that boosts the productivity of its human experts, and generates high-frequency, automated decisions and recommendations for manual decisions made between domains. Controllable AI will serve as a bridge, protocol or platform between the decision-making domains, where the shared model presents one lens of reality.
This integration between controllable AI and genAI will not only empower the plant workforce and enable high-performing collaborative teams, it will unlock the true power of modern AI in process plants.
