In-situ monitoring and metrology
📍 Where we are: Part III · Sensing and Control — Chapter 6 of 19. The previous chapters explained how defects form and how models and machine learning try to predict them; this chapter asks whether we can see them while the part is being built.
Imagine baking a very long loaf of bread inside an oven with no window, and only cutting it open at the end to find out whether the middle is raw. In-situ monitoring is the attempt to fit that oven with a window, a thermometer and a microphone. The hard part is not collecting the readings. It is knowing which flicker of light or click of sound means "raw dough here", and proving it by cutting the loaf open afterwards.
What this chapter covers
We follow the signal chain from the melt pool to a defect call: which sensors exist, what each can and cannot resolve, how signals are linked to defects, how those links are validated, and why the industry still treats most monitoring as a record rather than a verdict. The emphasis is on laser powder bed fusion (PBF-LB/M), where most of the evidence sits. Claims are labelled lab-scale, production or contested.
Problem
A metal part is made of tens of thousands of layers and millions of melt tracks. Defects such as lack of fusion, keyhole pores and cracks (see Defects, residual stress and distortion) are typically found only afterwards, by X-ray computed tomography (CT) or by destructive testing. Both are slow and expensive, and CT struggles with dense or large parts. Foundational reviews catalogue the defect types, the process signatures that can be sensed and the detection approaches for powder bed fusion [1][2], and an updated review stresses a persistent gap between methods demonstrated in the literature and industrial implementation [3]. An in-situ signal that reliably flags a flaw would cut inspection cost and, ultimately, support qualification (Qualification and certification). A NIST report written to support regulation of AM nuclear components reviews in-process monitoring and in-process non-destructive evaluation (NDE) and reports that in-process NDE methods are at lower technology readiness levels than post-process ones [4].
State of the art
The sensor families
| Sensor | What it sees | Typical role | Evidence |
|---|---|---|---|
| Co-axial photodiode / pyrometer | Integrated melt-pool light along the laser path; Renishaw's MeltView, for example, records visible (300-700 nm) and infrared (700-1700 nm) emission [5] | Melt-pool stability, power fluctuation | Production (as a logging tool) |
| Co-axial or off-axis camera, thermography | Melt-pool shape, hotspots, cooling | Heat accumulation at overhangs, cold spots at track intersections linked to lack of fusion [6]; melt-pool, plume and spatter features from a high-speed vision system were classified with SVM and CNN to identify build-quality level [7] | Production logging; defect calls lab-scale |
| Layerwise optical tomography (OT) | Time-integrated light per layer from a camera | Layer-by-layer anomaly maps; with a self-organizing map and a U-Net, OT images from an EOS M290 were used to predict lack-of-fusion and keyhole porosity, checked against CT [8] | Production (as a map); defect prediction lab-scale |
| Acoustic emission (contact or airborne microphones) | Sound from keyhole oscillation and collapse | Regime and keyhole-pore detection | Lab-scale |
| Interferometric / OCT-type depth sensing | Surface and keyhole depth with 1-20 µm axial resolution [5] | Depth metrology | Mostly lab-scale |
| Powder-bed imaging (off-axis) | Recoating defects, short feeds | Process stability | Production |
A 2024 review of in-situ porosity detection concludes that optical monitoring mainly finds large anomalies of the powder bed, thermal methods can detect defects down to about 100 µm and keyholes, and multi-sensor combinations are advantageous [9].
Ground truth from X-ray: where signals get their meaning
A signal means nothing until it is tied to a physical event. The strongest evidence comes from operando synchrotron X-ray imaging, first shown for laser powder bed fusion by Zhao et al., who resolved melt-pool dynamics and keyhole-pore formation in Ti-6Al-4V [10]. It films the melt pool and keyhole beneath the surface while sensors record in parallel. Ren et al. combined high-speed X-ray imaging with thermal imaging on Ti-6Al-4V, labelled short-window scalograms as "pore" or "non-pore" from the X-ray frames, and trained a convolutional neural network that detected stochastic keyhole-pore events with submillisecond resolution [11]. For acoustics, Hamidi Nasab et al. synchronised X-ray and sound recordings and found that acoustic amplitude rises when the process moves from conduction to keyhole mode, which they link to changing laser absorptivity [12]. Later work reports high-frequency acoustic components re-appearing with deep melt pools at the onset of keyhole-induced porosity [13], and sub-millisecond pore detection from combined sound and light sensors [14]. An airborne-microphone study validated against ex-situ CT predicts keyhole pores per unit scan length with a lightweight network [15].
All of this is lab-scale evidence: single tracks or thin walls, one or a few alloys, a synchrotron beamline and a custom rig, not a production machine on a full build.
The production reality
Commercial PBF-LB/M machines ship with photodiode and camera systems, and OT and melt-pool monitoring are used to log and compare builds. Their dominant use is statistical process control and flagging builds or layers for closer inspection, not replacing CT. That distinction matters, because it is where the "contested" label applies.
Why it is hard
- Resolution gap. A pore of tens of micrometres is far smaller than the pixel footprint of a layerwise camera, and sub-surface flaws are buried under later layers. In one study an infrared camera missed defects smaller than 600 µm, and defect area measured by IR and CT differed by about 60% [16].
- Registration and false indications. Signals must be aligned in space and time with CT voxels. Some flagged areas are remelted by later layers and are never actual defects, producing false positives [16].
- Proving detection. Snow et al. built a monitoring-based flaw detector and compared its calls with post-build CT using probability-of-detection (POD) and probability-of-false-alarm curves. Their system found flaws in the range 200-1000 µm, and they argued that standard machine-learning metrics such as accuracy are inadequate for judging flaw detection; they proposed the a90/95 size (flaw size detected 90% of the time at a 95% lower confidence bound) [17]. Notably, this is the language of NDE qualification, not of data science.
- Data volume. A single melt-pool camera at kilohertz rates produces a data stream of the order of a gigabyte per second (illustrative, not measured: one megapixel, 8 bits per pixel, 1000 frames per second), and a build runs for hours. The literature we surveyed discussed real-time image processing as a challenge but offered few hard numbers on storage, so we treat the severity as unquantified.
- Transfer. Models trained on one alloy, machine, scan strategy and gas-flow condition do not necessarily transfer (Machine learning for additive manufacturing).
Four sensor families feed a defect call; validation against operando X-ray (lab-scale) and post-build CT is the weak link.
Original diagram by the authors, created with AI assistance.
Open questions
- Can a monitoring system reach a statistically defensible a90/95 for flaws at the size that matters for fatigue (tens of micrometres), not the 200-1000 µm reported in the cited study?
- Which sensor combinations transfer across machines and alloys without re-labelling with X-ray or CT? This remains a contested area.
- How should in-situ evidence be accepted in qualification: as a complement to CT, or as a substitute for part of it? (Qualification and certification)
- What are the right data formats and compression for kilohertz melt-pool streams, so that they can live in a digital thread (Digital thread, data and the supply chain)?
- Can acoustic keyhole detection, shown on single tracks and thin walls, hold up on full builds with spatter, plume and machine noise?
Industry impact
Monitoring is already in production as logging and anomaly flagging: it shortens root-cause analysis when a build fails and lets operators compare builds. It has not replaced volumetric inspection for critical parts. The commercial case for monitoring is strongest where it reduces scrap or inspection scope; the regulatory case depends on the validation questions above. The step from watching to acting is the subject of the next chapter, Closed-loop process control.
Why it matters
If a sensor can be trusted, a build can be stopped early, a part can be flagged for targeted inspection, and a qualification package can rest on process evidence rather than on sampling alone. If it cannot, monitoring merely adds data nobody can act on. The difference lies almost entirely in the validation, which is slow, expensive and unglamorous work.
In the real world
Machine vendors offer monitoring suites such as Renishaw's InfiniAM (LaserView and MeltView modules) and EOS's Smart Monitoring [5][18]. Public research relies on synchrotron facilities, such as the beamlines used in the X-ray studies cited above, to supply the ground truth that production machines cannot.
Key terms
- Operando X-ray imaging — filming the melt pool inside the metal with a high-speed synchrotron X-ray beam while the process runs.
- Co-axial monitoring — a sensor that looks along the same optical path as the laser, so it always sees the melt pool.
- Optical tomography (OT) — a camera that integrates the light of each layer to form a layer-by-layer map of anomalies.
- Acoustic emission — sound produced by the melt pool, used as a signature of keyhole regimes.
- Probability of detection (POD) — the chance a flaw of a given size is detected; the standard way NDE methods are qualified.
- a90/95 — the flaw size detected 90% of the time at a 95% lower confidence bound.
- Registration — aligning in-situ data with CT or part coordinates in space and in time.
- False indication — a flagged location with no real defect, for example one erased by remelting.
- Technology readiness level (TRL) — a nine-step scale of maturity from idea to proven system.
Where this leads
Seeing a defect is not the same as preventing it. Closed-loop process control asks when a sensor signal can be fed back fast enough to change the process.