Seeing Through the Glow: How Dual-Wavelength Pyrometry and Deep Learning Combine to Predict Porosity Across Any LPBF Process Regime

Published: August 2026

Source: Additive Manufacturing, Volume 127 (2026), Article 105282

Keywords: LPBF, in-situ monitoring, two-wavelength pyrometry, melt pool, pore defects, deep learning, real-time quality assurance

Every engineer who has pointed a thermal camera at a laser powder bed fusion (LPBF) melt pool knows the frustration. The molten metal glows brilliantly — thousands of kelvin — but converting that glow into an accurate temperature requires knowing the material's emissivity. And emissivity in LPBF is a moving target: it depends on wavelength, temperature, surface roughness, oxidation state, and whether you are looking at solid metal, liquid metal, or powder. Get it wrong, and your temperature measurement can be off by hundreds of degrees.

This is not merely an academic inconvenience. Pores — the internal voids that compromise fatigue life, tensile strength, and part qualification — are born in the melt pool. Their formation is governed by the very thermodynamics that emissivity uncertainty obscures. Keyhole collapse, lack of fusion, gas entrapment: every pore formation mechanism leaves a thermal signature. But if your thermal sensor cannot distinguish a 2200 K melt pool from a 2600 K one, that signature is lost in the noise.

A team at the University of Pittsburgh led by Prof. Xiayun Zhao, in collaboration with Georgia Tech, has developed and patented a fundamentally different approach. Their Single-Camera Two-Wavelength Imaging Pyrometry (STWIP) system acquires simultaneous images at two carefully selected wavelengths (550 nm and 620 nm) using a single high-speed camera. By taking the ratio of the signals at these two wavelengths, emissivity cancels out — assuming it does not vary dramatically between the two narrow bands, which is a well-justified approximation for most LPBF metals.

The numbers are impressive: acquisition rates up to 100,000 fps (22,500 fps demonstrated in sustained multi-layer builds), spatial resolution of 20 μm, and full-field melt pool temperature and morphology measurements — all while integrated coaxially into a commercial EOS M290 system. Unlike photodiodes that provide only single-point intensity, STWIP delivers spatially resolved 2D temperature maps. Unlike conventional thermal cameras, it does not need emissivity calibration.

The experimental campaign was ambitious: 16 Inconel 718 fatigue specimen blocks printed with systematically varied laser power, scan speed, and hatch spacing to cover three critical process regimes — lack of fusion, transition, and keyhole. The resulting dataset comprised 2,461,740 individual melt pool observations, each enriched with extracted temperature, area, width, length, and spatial coordinates. Ex-situ X-ray computed tomography served as ground truth for pore metrics.

If STWIP provides the eyes, deep learning provides the brain. The team developed a context-aware framework that transforms raw melt pool time-series into three complementary representations:

• Temporal features: 1D time-series of melt pool temperature, area, and morphology, fed into LSTM and transformer architectures that capture the sequential dynamics of laser scanning.

• Spatial features: 2D maps of melt pool properties across each layer, processed by convolutional neural networks (CNNs) that learn spatial correlations between adjacent scan tracks.

• Spatiotemporal scalograms: Continuous wavelet transforms (CWT) of melt pool signals converted into time-frequency images — scalograms — that simultaneously encode both temporal dynamics and frequency-domain signatures of process instability. These are fed into vision transformer (ViT) models.

Critically, the framework is process-agnostic. It does not require nominal process parameters (laser power, scan speed) as inputs, nor does it assume the process is operating in a known regime. It learns directly from the physics captured by STWIP — the actual melt pool thermodynamics and morphology — making it robust to the stochastic regime transitions and unknown process deviations that characterize real manufacturing environments.

The results represent a significant advance over the qualitative or binary classification approaches that dominate the literature:

Table 1: STWIP + deep learning framework performance. These results are particularly strong given the extreme class imbalance (pores occupy <1% of volume) and the process variability across 16 builds with intentionally varied parameters.

These numbers deserve context. Pores occupy a tiny fraction of the total part volume — typically well under 1%. The extreme class imbalance makes this a profoundly challenging machine learning problem. Achieving 85% precision and 84% recall under these conditions, and maintaining performance on entirely unseen builds, is a genuine achievement. The 72% accuracy for porosity ratio prediction, while lower than the other metrics, still represents a functional quantitative estimate where most prior work offers only binary yes/no detection.

The term 'process-agnostic' in this work carries a specific and important meaning. It does not claim to work across different AM processes (LPBF vs. DED vs. binder jetting) or different materials. Rather, it means the framework operates without knowledge of nominal process parameters or predefined regime labels. It does not need to know the laser power was 285 W or the scan speed was 0.96 m/s. It does not need to be told this is a 'keyhole regime' build. It learns directly from the physics.

This is important because in real production, nominal parameters are often poor proxies for actual melt pool conditions. A machine set to 300 W may deliver 290 W. Powder quality, gas flow, and build plate temperature all drift. The framework's ability to characterize pores from melt pool signatures alone — rather than from what the machine was told to do — is what makes it robust to production variability.

Several features of this work position it for industrial translation. STWIP's coaxial integration means it sees what the laser sees, eliminating the registration errors and optical distortions that plague off-axis monitoring. The adaptive wavelength selection capability means the system can be tuned for different materials — aluminium, titanium, steel, nickel alloys — by selecting wavelength bands optimized for each material's emission spectrum. And the spatiotemporal learning architecture, by processing melt pool data at the voxel level with context from adjacent layers and scan tracks, captures the remelting and thermal accumulation effects that simpler models miss.

Challenges remain. The current X-ray CT ground truth has a detection limit of ~25 μm, meaning the smallest gas pores are not characterized. The 72% porosity ratio accuracy, while useful, needs improvement for applications where tight porosity specifications (e.g., <0.1% for aerospace) are critical. And the computational pipeline — from 22,500 fps data acquisition through wavelet transformation through deep learning inference — must be hardened for real-time deployment at production throughput.

But the trajectory is clear. This is one of the most complete demonstrations to date of quantitative, process-agnostic pore characterization from in-situ monitoring data, combining a genuinely innovative sensor with a thoughtfully designed learning framework. For an industry where part qualification remains the dominant cost driver, that combination is exactly what the future looks like.

Reference: Zhang, H., Caputo, A.N., Neu, R.W., Vallabh, C.K.P., To, A.C., Alam, M.J., & Zhao, X. (2026). Process-agnostic multi-metric nondestructive characterization of porosity in additive manufacturing via specialized single-camera two-wavelength pyrometry. Additive Manufacturing, 127, 105282.