In industrial plants, automation means control loops in which a sensor, a controller, and an actuator keep a process at a set point, and most of that work is done by programmable logic controllers (PLCs), distributed control systems (DCS), and supervisory control and data acquisition (SCADA) systems that follow fixed, tested algorithms. Learned AI models mostly sit beside that layer, but one peer-reviewed 2022 study placed a neural network inside a fast control loop on a research tokamak, running at 10 kHz.
Key takeaways
- NIST describes a control loop as sensors, controllers, and actuators, with cycle times from milliseconds to minutes.
- PLCs, DCS, and SCADA are three different architectures, split by size, location, and how much a human is involved.
- NIST states that in most safety-critical operations human oversight of a dangerous process is essential.
- In the 2022 Nature study, a reinforcement-learning controller was trained in a simulator, compiled for 10 kHz control, and tested on the TCV tokamak, while a traditional controller handled plasma formation.
- A claim that a plant is "AI-automated" should name the loop, the model, and who supervises it.
The control loop is the basic unit
The National Institute of Standards and Technology (NIST) describes operational technology in its September 2023 guide SP 800-82 Revision 3, as programmable systems and devices that interact with the physical environment. Inside such a system, a control loop uses a sensor to measure a physical property, a controller to compare the measurement with a target set point, and an actuator, such as a valve, breaker, switch, or motor, to change the process. The guide recognizes three configurations: open loop, where output follows established settings; closed loop, where output feeds back to the input; and manual mode, where humans control the system entirely.
Loops can be nested, with the set point of one loop coming from another. NIST says supervisory and lower-level loops run continuously with cycle times in the order of milliseconds to minutes. NIST says a PLC has user-programmable memory for functions such as input and output control, logic, timing, counting, and three-mode proportional-integral-derivative (PID) control.
Three architectures, compared
The table paraphrases the NIST descriptions of each topology. NIST adds that real systems may be hybrids that blur these lines.
| Architecture | Typical scope | Examples NIST gives | Human role |
|---|---|---|---|
| PLC-based | Local discrete process, often without a central server | Automobile assembly lines | Minimal during operation; closed-loop control |
| DCS | One site, with a supervisory level over local controllers | Oil refineries, chemical manufacturing, pharmaceutical processing | Operators supervise; loops hold set points automatically |
| SCADA | Dispersed assets, with central data collection | Water distribution, pipelines, electrical transmission, rail | Automatic control or operator commands, in near real time |
A PLC can serve as the local controller inside a DCS or SCADA design, so the three are layers as much as alternatives. In NIST's PLC example, engineers reach the controller through a programming interface on an engineering workstation, and process data is stored in a data historian. A historian is also where analytics, including machine learning, would draw its data.
Why timing and safety limit where learned models go
NIST lists the factors that shape an OT design. Control timing is one: some processes need speed, consistency, and synchronization that humans cannot reliably meet, and some computation has to happen close to sensors and actuators to avoid communication delay. Safety is another: systems must detect unsafe conditions and trigger actions that bring them back to safe ones, and in most safety-critical operations human oversight and control of a dangerous process are described as essential. The guide also notes that simple controllers and preset algorithms can handle many control functions, while more complex systems need human operators to confirm that actions serve the larger objectives.
A reasonable reading of these requirements is a division of labor. Models that forecast demand, flag wear, or schedule production can run on historian data with seconds or hours to spare, and their output reaches a person or a higher-level planner. A model that commands an actuator directly must meet the same timing, testing, and fail-safe expectations as the controller it replaces.
One documented case inside the loop: the tokamak controller
Jonas Degrave and colleagues reported in Nature on February 16, 2022 a magnetic controller for tokamak plasmas designed through deep reinforcement learning. The abstract states that the architecture learns to command the full set of control coils and produced a diverse set of plasma configurations on the Tokamak à Configuration Variable (TCV), including elongated shapes, negative triangularity, "snowflake" configurations, and sustained "droplets" with two separate plasmas in the vessel.
The paper shows how a learned controller was fitted into industrial-style practice. The policies were learned through interaction with a tokamak simulator, then bundled with the experiment targets into an executable built with a compiler tailored to real-time control at 10 kHz and loaded into the TCV control framework. Each experiment began with standard plasma-formation procedures in which a traditional controller maintained plasma location and current. The result concerns one experimental machine, so it does not show that learned controllers are ready for a refinery or a production line, where certification and failure costs differ.
Questions for any claim of an AI-run plant
- Which loop does the model control, and what is its cycle time?
- Does the model write set points to a PLC or DCS, or does it only advise an operator?
- What happens when the model fails or its inputs drift: is there a fallback controller?
- Who is the accountable person on shift, and what can that person override?
- Was the model tested only in simulation, on one site, or across several plants?
FAQ
Is every automated factory using AI?
No. NIST's description of PLC-based control centers on logic, timing, counting, and PID control, none of which involves a learned model. Many plants automate heavily with no machine learning.
What is the difference between SCADA and a DCS?
NIST says SCADA controls dispersed assets where central data collection matters as much as control, such as pipelines and utilities. A DCS controls production systems at a single location, such as a refinery, with a supervisory level over local controllers.
Can a neural network replace a PLC?
The tokamak study replaced a hand-designed magnetic controller on one research device, and it still relied on a traditional controller during plasma formation. It does not show that PLCs in general can be replaced.
Does automation remove the human operator?
NIST says some PLC systems provide closed-loop control with minimal human involvement, yet for most safety-critical operations it calls human oversight essential. How much a person does depends on the architecture and the risk.
Related reading
- How augmenting automation affects employment
- Ten AI technologies used across industries
- AI methods inside data science work
- What the Deloitte AI Institute publishes
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Guide to Operational Technology (OT) Security, NIST SP 800-82 Rev. 3, National Institute of Standards and Technology, September 2023
- Magnetic control of tokamak plasmas through deep reinforcement learning, Degrave J et al., Nature 602, 2022