Watching Every Layer: A Robust Framework for In-Situ Geometric Inspection of LPBF Lattice Structures

How edge-based active contours, smart deviation metrics, and industrial powder-bed cameras can catch defects layer by layer — before the build is finished

Published: July 2026

Source: Additive Manufacturing, Volume 124 (2026), Article 105228

Keywords: In-situ monitoring, lattice structures, geometric deviation detection, active contours, powder-bed imaging, LPBF

Part qualification is the silent cost driver in metal additive manufacturing. For complex geometries — particularly lattice structures with their intricate internal surfaces — X-ray computed tomography (CT) is often the only viable inspection method. And it is slow, expensive, and poorly suited to large or dense components. The result: qualification can account for a substantial fraction of total production cost, eroding the economic case for AM in exactly the applications where its geometric freedom is most valuable.

But what if you could inspect the part while it is being built — layer by layer — using the cameras already mounted inside commercial LPBF machines? This is the promise of in-situ powder-bed imaging, and a team from Politecnico di Milano and Siemens AG has just taken it a major step closer to industrial reality.

At the heart of any in-situ geometric inspection system lies image segmentation — the algorithmic task of identifying which pixels in a powder-bed image correspond to solidified metal and which correspond to loose powder. The research team adopted active contours, a class of algorithms that iteratively refine an initial boundary (seeded from the nominal CAD geometry of each layer) by minimising an energy functional that balances two competing forces:

• Region-based terms, which drive the contour toward areas of homogeneous pixel intensity, working best when solidified regions have consistent brightness distinct from the powder background.

• Edge-based terms, which drive the contour toward sharp intensity gradients — the boundaries between bright fused metal and dark powder — working best when those edges are crisp and well-defined.

A parameter w (0 ≤ w ≤ 1) balances these two forces, with w = 1 representing purely region-based segmentation and w = 0 representing purely edge-based. The kernel radius rkernel controls the spatial neighbourhood over which region-based statistics are computed.

Previous studies had largely used default parameter values (w = 0.5, rkernel = 5) without systematic calibration. The team asked a simple question with profound implications: what if those defaults are suboptimal — and what if the optimal choice makes the system robust to the very sources of variability that plague industrial environments?

To answer this, they manufactured gyroid lattice structures in AlSi10Mg across two builds with deliberately different lighting conditions — Build 1 used dark-field illumination (a directional light source beside the camera, known to be favourable for segmentation), while Build 2 used conventional diffused LED strips (more common in production but less forgiving for image processing). Each build contained structures with two wall thicknesses: 0.8 mm (thin, delicate) and 2 mm (thick, robust).

A full grid search over w (0 to 1, step 0.05) and rkernel (1 to 12 pixels, step 1) — calibrated against X-ray CT ground truth — delivered a clear verdict: w = 0, i.e., purely edge-based active contours, was optimal under every condition tested.

Once the solidified region is reconstructed for each layer, the next challenge is quantifying how much it deviates from the nominal geometry — and doing so in a way that reliably separates true defects from natural process variation.

The team compared two fundamentally different deviation metrics:

• ΔArea(z): The simple difference between the reconstructed area and the nominal area in layer z. Robust to misalignment errors, but blind to shape distortions or translations that preserve total area.

• Dtot(z): The symmetric difference — the area of the union minus the area of the intersection between nominal and reconstructed regions. Sensitive to any deviation affecting overlap, including shape changes and translations, but more vulnerable to residual alignment errors.

Critically, Dtot(z) can be directionally decomposed into Dexcess(z) (local oversizing, excess material) and Ddeficit(z) (local undersizing, lack of material). This decomposition proved decisive.

To validate the framework, the team introduced controlled defects of varying severity. In Build 2, the rubber recoater blade was deliberately damaged with two indentations: a high-severity notch (4 mm wide, 1 mm deep) and a low-severity notch (1 mm wide, 1 mm deep). These produced visible anomalies in the powder bed — local ridges of excess powder that translated into geometric superelevation (swelling) on the affected parts.

The detection performance was evaluated using a statistical process control framework: B-spline weighted least-squares fitting of 1-D deviation profiles for each unit cell, with control limits set at k = 3 standard deviations from the calibrated baseline. The results, summarised in Table 1 below, demonstrate the power of metric selection:

Table 1: In-situ defect detection performance. Dexcess(z) achieves 100% true positive rate for high-severity defects across both wall thicknesses, with false positive rates ≤ 2%. The directional metric dramatically outperforms the non-directional Dtot(z).

The key findings leap off the table. The directional metric Dexcess(z) achieved 100% true positive detection for all high-severity defects at p = 5 consecutive-layer alarm threshold, while maintaining false positive rates at or below 2%. In contrast, the non-directional Dtot(z) — which lumps excess and deficit together — missed over 80% of high-severity defects in the thin-wall case. The low-severity defect in thick walls proved challenging for all metrics, highlighting the detection limit for subtle anomalies buried within larger feature cross-sections.

The study distils its findings into actionable recommendations for manufacturers looking to deploy in-situ geometric inspection:

• Calibrate your active contours. A one-time calibration using a small set of representative parts and X-ray CT ground truth can dramatically improve reconstruction accuracy. The optimal parameters for this study (w = 0, edge-only) proved robust across lighting conditions and part geometries.

• Choose your metric based on the defect you are hunting. If you expect local oversizing (e.g., from recoater damage, powder spreading anomalies), use Dexcess(z). For undersizing (e.g., balling, lack of fusion), use Ddeficit(z). For general monitoring, ΔArea(z) provides robustness to alignment errors.

• Set realistic alarm thresholds. Requiring p = 5 consecutive control-limit violations before raising an alarm eliminated nearly all false positives without significantly compromising true positive detection for high-severity defects.

• Account for location effects. Parts at different positions in the build chamber reflect light differently toward the camera. Location-dependent calibration of control limits can reduce false alarms without sacrificing sensitivity.

What makes this work particularly significant is its grounding in industrial reality. The powder-bed camera used (an IDS UI-5490SE-MGL with a 10.55 Mpix sensor) is already integrated into commercial LPBF systems. The active contour algorithm is computationally tractable for layer-by-layer execution. The statistical process control framework is familiar to manufacturing quality engineers. And the calibration methodology — while requiring X-ray CT for ground truth — only needs to be performed once per material–geometry–machine combination.

The remaining challenges are clear: extending the framework beyond regular lattices with repeating unit cells to arbitrary free-form geometries; characterising detection probability as a function of defect size through formal Probability of Detection (PoD) analysis; and further reducing location-dependent variability through improved imaging hardware, such as blade-mounted cameras that maintain a constant, close distance to the powder bed.

Reference: Grasso, M., Ertelthalner-Nikolaev, D., Dallmann, M., & Colosimo, B.M. (2026). In-situ qualification of lattice structures in L-PBF: A framework for robust geometrical deviation detection. Additive Manufacturing, 124, 105228.