Closed-Loop Control (CLC)

The defining characteristic of a closed-loop system is the continuous measurement of the Process Variable (PV) via physical sensors. This feedback signal creates a data pathway from the process output back to the controller input.

• Sensor Types: Thermocouples, strain gauges, laser micrometers, and accelerometers convert physical phenomena into electrical signals.
• Sampling Rate: The frequency of measurement must satisfy the Nyquist-Shannon sampling theorem to accurately capture process dynamics without aliasing.
• Signal Conditioning: Raw sensor signals require amplification, filtering, and analog-to-digital conversion before the controller can interpret them.
• In-Situ Metrology provides feedback without removing the workpiece, while Virtual Metrology uses machine learning models to predict quality characteristics from equipment sensor data when direct physical measurement is impractical.

The controller continuously computes the error signal by subtracting the measured Process Variable (PV) from the desired Setpoint (SP). This error represents the instantaneous deviation from the target operating condition.

• Error Formula: e(t) = SP(t) - PV(t) where a positive error indicates the process is below target and a negative error indicates it is above.
• Deadband: A configurable tolerance zone around the setpoint where no corrective action is taken, preventing actuator hunting and excessive wear from responding to negligible noise.
• Error Classification: Modern systems distinguish between random error (common-cause variation addressed by Statistical Process Control) and systematic error (special-cause drift requiring root cause analysis).
• The comparison function is the logical bridge between the desired state and the actual state, enabling the system to self-assess its performance without human interpretation.

The controller executes a deterministic algorithm that translates the error signal into a Manipulated Variable (MV) command. This computation defines the system's dynamic response to disturbances and setpoint changes.

• PID Control: The Proportional term responds to present error, the Integral term eliminates steady-state offset by accumulating past error, and the Derivative term anticipates future error by reacting to the rate of change.
• Model Predictive Control (MPC) uses a dynamic process model to predict future behavior over a finite horizon and solves an optimization problem to compute the optimal control trajectory while respecting actuator and safety constraints.
• Gain Scheduling automatically adjusts controller parameters based on the current operating point, maintaining stability across non-linear process regions.
• Run-to-Run (R2R) Control modifies recipe parameters between discrete processing runs based on post-process metrology, making it ideal for semiconductor wafer fabrication where in-situ measurement is impossible.

The computed control signal drives a physical actuator that manipulates the process input, closing the loop by influencing the Process Variable. The actuator and the physical process together form the plant in control theory.

• Actuator Types: Variable frequency drives adjust motor speed, proportional valves regulate fluid flow, heating elements control temperature, and servo motors position robotic joints.
• Actuator Saturation: Physical limits on valve travel, motor torque, or heater power create non-linear behavior when the controller demands outputs beyond the actuator's capability, requiring anti-windup compensation in the integral term.
• Process Dynamics: The plant exhibits characteristic behaviors including dead time (transport delay before the actuator's effect reaches the sensor), time constants (the speed of response), and process gain (the magnitude of output change per unit input change).
• Disturbance Rejection: A primary measure of closed-loop performance is how effectively the system attenuates external disturbances—variations in raw material, ambient temperature, or upstream pressure—that would otherwise degrade product quality.

Closed-loop control fundamentally relies on negative feedback, where the corrective action opposes the detected error. This self-correcting architecture is what distinguishes CLC from open-loop systems and enables automatic disturbance rejection.

• Negative Feedback Principle: If the Process Variable rises above the setpoint, the controller reduces the Manipulated Variable to bring it back down. This opposition creates a restoring force toward equilibrium.
• Stability Margins: Gain margin and phase margin quantify how much additional gain or phase lag the system can tolerate before becoming unstable and oscillating uncontrollably.
• Bode Plots and Nyquist Criteria: Frequency-domain analysis tools used during controller design to ensure the closed-loop transfer function has no poles in the right-half plane.
• Positive feedback, in contrast, amplifies deviations and drives the system toward saturation or failure—a condition rigorously avoided in industrial control design.

While the core closed-loop controller maintains a given setpoint, a higher-level supervisory system continuously optimizes that setpoint to maximize business objectives such as throughput, yield, energy efficiency, or profit per unit.

• Advanced Process Control (APC) layers sit above regulatory PID loops, using multi-variable models to coordinate dozens of interdependent control loops and push the process toward economic optima without violating constraints.
• Golden Batch Profiles: Stored time-series trajectories from historically optimal production runs serve as dynamic setpoint references, enabling the system to replicate ideal conditions rather than static targets.
• Bayesian Optimization sequentially explores the process parameter space, building a probabilistic surrogate model to find global optima with minimal expensive physical experiments.
• Digital Twin integration allows the supervisory layer to simulate setpoint changes against a high-fidelity virtual replica before committing them to the physical plant, de-risking optimization moves.

The foundational regulatory control algorithm at the heart of most closed loops. PID tuning adjusts three gain parameters—proportional (present error), integral (accumulated past error), and derivative (predicted future error)—to achieve a stable, responsive correction signal.

• Proportional: Reacts to current deviation from setpoint
• Integral: Eliminates steady-state offset over time
• Derivative: Dampens overshoot by anticipating error trajectory

Poorly tuned PID loops cause oscillation, sluggish response, or instability, directly degrading product quality. Modern auto-tuning methods use relay feedback or model-based optimization.

A complementary strategy that measures disturbances directly and applies corrective action before they impact the process variable. Unlike feedback, which must wait for an error to develop, feedforward is preemptive.

• Requires an accurate disturbance-to-output model
• Typically paired with a feedback trim loop to correct model inaccuracies
• Excels at rejecting measurable load disturbances like incoming material temperature or pressure fluctuations

Example: A heat exchanger measuring incoming fluid temperature and immediately adjusting steam flow, rather than waiting for the outlet temperature to deviate.

A predictive technique that estimates quality characteristics using equipment sensor data and machine learning models, eliminating the delay and cost of physical measurement.

• Replaces or supplements in-situ metrology for wafer-level process control
• Enables wafer-to-wafer or run-to-run control without breaking vacuum
• Uses Gaussian Process Regression or neural networks to provide prediction confidence intervals

Critical for semiconductor fabrication, where measuring every wafer is physically impossible and closed-loop correction must occur at production speed.

A high-fidelity, real-time synchronized virtual replica of a physical asset or process. The digital twin consumes live sensor streams to mirror actual state, enabling what-if simulation, anomaly detection, and control optimization without risking production.

• Physics-based models capture first-principles behavior
• Data-driven models learn from historical operational patterns
• Enables setpoint optimization by simulating thousands of control scenarios in parallel

The twin serves as the dynamic process model for MPC and the sandbox for testing new control strategies before deployment.

The execution of trained machine learning models directly on factory-floor hardware—gateways, industrial PCs, or embedded accelerators—rather than in the cloud. This is non-negotiable for closed-loop control where latency must be deterministic.

• Sub-millisecond inference on dedicated NPUs or FPGAs
• Eliminates network jitter and bandwidth bottlenecks
• Enables air-gapped operation for security-sensitive processes

Example: A vision-based defect classifier running on an edge device triggers a reject actuator within 10ms, closing the quality loop before the next part arrives.