Closed-loop process control
📍 Where we are: Part III · Sensing and Control — Chapter 7 of 19. The last chapter asked whether defects can be seen during a build; this one asks whether the machine can act on what it sees, quickly enough to matter.
A thermostat works because the room changes slowly: it reads the temperature, switches the heater, and the room responds over minutes. Now picture a thermostat for a spot of molten metal the size of a pinhead that changes in a thousandth of a second, in a room you can only see through a keyhole. That is closed-loop control of additive manufacturing. Everything depends on how fast you can read, how much of the truth you can read, and what knob you are allowed to turn.
What this chapter covers
We separate monitoring (recording) from control (acting), explain the three constraints that make AM control hard, and walk through what has actually been demonstrated: feedback on laser power in powder bed fusion, melt-pool control in directed energy deposition (DED), feed-forward compensation from simulation, and commercial offerings. Each claim is labelled lab-scale, production or contested.
Problem
Fixed process parameters are blind to local conditions. In laser powder bed fusion the same laser power is applied to a long solid region and to a small feature where short scan vectors let heat accumulate. Wang et al. note that a constant parameter set over a whole print ignores heat accumulation and feature size, so short raster lines on small features over-melt [1]. NIST researchers make the same point from the other side: a fixed power per scan vector ignores thermal variation, which can cause lack of fusion in underheated areas and keyhole porosity in overheated ones [2]. A controller that measures the melt pool and adjusts power would, in principle, keep the process inside its window (Melt-pool physics) without an engineer pre-computing every case. The need for real-time control of powder bed fusion was set out in a NIST measurement-science report [3], and a review of in-situ monitoring treats automated defect detection and process control as critical fields for future smart PBF systems [4].
State of the art
Feedback on laser power in powder bed fusion (lab-scale)
Wang et al. built a customised PBF-LB/M testbed in which a thermal-emission sensor drives a controller that regulates laser power at 2 kHz to hold a target melt-pool temperature; microscopy and 3D scanning showed improved accuracy on small features [1]. That is a deliberately simple, fast loop on a research testbed, and a good illustration of how far the demonstrations are from a production architecture.
A 2025 NIST publication regulates melt-pool area in line, using a co-axial camera and a PID loop. On nickel superalloy 625 plates, the laser power is adjusted within about 118 microseconds of a measurement being acquired, and the authors report that feedback meaningfully improved process stability [2]. Control within the same melt pool, not the next layer, is the point.
Layer-to-layer control (simulation)
If the loop cannot act within a track, it can act between layers. Liao-McPherson et al. propose a trajectory-optimisation controller using a linear parameter-varying model that changes laser power for the next layer to track a target melt-pool depth. The controller was validated in a high-fidelity particle-hydrodynamics simulation of a 2-D process, not on a machine; the authors present computational cost only as an outlook toward real systems [5]. Their motivation for feedback is the uncertainty in the process, which a model alone cannot anticipate [5].
DED: where control is further along
DED has a bigger, slower, more accessible melt pool, so feedback is easier. In a study on an Optomec LENS MR7, a co-axial two-wavelength imaging pyrometer controlled peak melt-pool temperature at 1650 ± 50 °C by modulating laser power; closed-loop parts had porosity of 0.036-0.043%, against 0.032-0.068% for open-loop parts [6]. Notice what improved: mainly the spread, not the mean. A 2025 study held melt-pool height using a long-wave infrared camera and a temperature-difference feedback signal, reporting geometric error against the digital model under 4% for multi-bead, multi-layer structures [7]. Feedback control of melt-pool temperature and deposition height in direct metal deposition was already demonstrated in 2011 [8]. Model-based feed-forward height control, which computes a laser power trajectory beforehand and runs it without sensing, was also demonstrated on a LENS MR-7 [9] (a single small study; we found no review that covers feed-forward DED control). A 2025 review compares the requirements and performance of state-of-the-art DED online controllers [10], and a 2026 paper reports closed-loop control from plasma-plume emission [11]. Evidence level for DED temperature and height control: lab-scale, with a smaller gap to production than PBF.
Commercial systems
Vendor-described closed-loop features exist. EOS markets Smart Fusion, which uses live optical-tomography data to adjust laser power during a build so that thermal conditions stay within limits set in pre-processing; a 2025 review of overhangs reports that it reduces the minimum overhang angle to 20° [12]. That is the clearest commercial example we found; we could not find independent validation numbers, so we treat it as production, vendor-reported. Renishaw's InfiniAM is described as a monitoring suite with support for user plug-ins rather than native closed-loop control [13].
Why it is hard
- Speed. Keyhole pore events occur on submillisecond time scales (In-situ monitoring and metrology). Loops at 2 kHz or with a latency of about 118 µs are an engineering achievement; but any loop slower than the melt-pool dynamics only corrects what has already happened.
- Observability. The controller sees an integrated surface signal (temperature proxy, melt-pool area, plume light), while the defect lives below the surface and depends on keyhole depth, which is not directly measured. Controlling a proxy is not the same as controlling the defect.
- Actuation. Laser power can be changed quickly. Scan speed, hatch spacing and gas flow cannot be changed as fast, and changing power alone moves the process along a line in the process map rather than steering it in any direction.
- Geometry blindness. Controllers that regulate a single melt-pool signal do not use part geometry. A 2026 study proposes a geometry-conditioned architecture in which an ontology sets the controller targets for overhangs, evaluated on a calibrated analytical melt-pool surrogate rather than on a machine [14]. Treat geometry-aware control as lab-scale at best.
- Qualification. A machine that changes its own parameters during a build creates a part whose process history differs from the one that was qualified (Qualification and certification).
A closed-loop laser-power controller: the loop needs fast sensing, an observable signal and an actuator; feed-forward from simulation sets the starting point.
Original diagram by the authors, created with AI assistance.
Open questions
- Can a loop control the defect (keyhole depth, pore generation) rather than a surface proxy? Operando X-ray and acoustic work gives labels; closing the loop on them is not yet shown outside the lab.
- How should feed-forward from simulation (Simulation and ICME) be combined with feedback so that each covers the other's weakness?
- How do controllers tune themselves across alloys and machines? A Bayesian-optimisation approach to in-situ controller auto-tuning has been reported in 2025 [15]; we cite it for the approach, not for its results, which we did not verify.
- How should controlled builds be certified when the machine itself varies parameters during the build?
- Can geometry-aware, multi-actuator control become standard, or does it stay one-signal, one-knob?
Industry impact
Feedback on laser power has been shown on research testbeds, and a vendor offers a version in production for overhangs. DED control is closer to practice because of its slower dynamics. For most users the practical value today is feed-forward compensation and monitoring-driven inspection, not autonomous correction of defects. The gap between monitoring and control is largely a gap in validated, certifiable decision logic; the next chapters show that materials and powder variability add noise that no controller currently absorbs (Alloys for AM and the powder problem).
Why it matters
Control is the logical end of the monitoring story: if sensors are trustworthy, the machine should use them. Reviews of monitoring and process control in metal laser AM cover this progression [16]. If they are not, control amplifies errors. The field has demonstrated the loop at laboratory scale but has not yet shown that it improves certified part quality across machines.
In the real world
Public examples include the NIST testbed result and the DED pyrometer studies above, and commercial features such as EOS Smart Fusion on overhangs. Practical users should ask vendors for independent test data before assuming a feature that "adjusts the laser" improves part properties.
Key terms
- Closed-loop control — measuring the output and adjusting an input to keep it near a target.
- Feed-forward control — adjusting inputs using a model of the process, without measuring the result.
- PID controller — a standard feedback rule combining the present error, its accumulated sum and its rate of change.
- Observability — whether the quantity you care about can be inferred from what you can measure.
- Actuator — the part of the machine that can be changed on command, such as laser power.
- Layer-to-layer control — adjusting parameters between layers rather than inside one track.
- Melt-pool temperature proxy — a surface emission signal standing in for the true temperature field.
- Latency — the delay between measuring and acting.
Where this leads
Even a perfect controller works on whatever powder and alloy it is given. Alloys for AM and the powder problem turns to the material itself.