One Camera, Two Measurements: How a Simple RGB Setup Finally Captures True LPBF Cooling Rates

Published: August 2026

Source: Additive Manufacturing, Volume 127 (2026), Article 105278

Keywords: LPBF, cooling rate, ratiometric imaging, melt pool monitoring, solidification, cell spacing, process monitoring

Of all the physical quantities that govern laser powder bed fusion (LPBF), solidification cooling rate is among the most consequential — and the least measured. Cooling rate determines the microstructure that forms: cell spacing, phase constitution, grain morphology. It controls defect susceptibility: shrinkage porosity becomes endemic when cooling is too slow. It governs residual stress accumulation. And it is the critical link between process parameters and final part properties.

Yet, remarkably, almost no one measures it directly. The vast majority of LPBF cooling rate 'data' in the literature comes from analytical models (Rosenthal, Eagar-Tsai), computational fluid dynamics simulations, or ex-situ inference from post-solidification microstructural features like cell spacing. Each approach has value, but each carries assumptions that can produce errors exceeding 100% relative to true values. The industry has been flying largely blind on one of its most important physical parameters.

A team at Carnegie Mellon University — Weeks, Myers, Wassermann, Quirarte, Beuth, Narra, Singh, and Malen — has devised a strikingly elegant solution. Their setup uses a single off-the-shelf RGB color camera and a blue laser diode illumination source. That's it. No exotic sensors, no synchrotron beamtime, no six-figure equipment budget.

The genius is in how they partition the camera's native color channels. The red and green channels capture the melt pool's thermal emission at two distinct wavelength bands — enabling ratiometric (two-color) temperature measurement that cancels out the unknown emissivity. The blue channel, meanwhile, captures the reflected blue laser light, which illuminates the melt pool and surrounding powder bed — providing a crisp geometric image of the melt pool boundary. A UNet machine learning model, trained on hand-labeled images, segments the illuminated frames to extract melt pool length and width.

By tracking the evolution of temperature and melt pool geometry over time, and combining them to calculate the temperature gradient along the melt pool tail (dT/dx), the team computes the travel-direction solidification cooling rate: Ẋ · dT/dx. The method is simultaneous — both data streams come from the same camera, at the same time, perfectly spatially registered. It is low-cost. And it is experimentally validated against the ultimate ground truth: the actual microstructure that forms.

Single-track scans on 316L stainless steel across six power–velocity combinations yielded cooling rates between 170,000 and 3,000,000 K/s — a range spanning more than an order of magnitude. The cooling rate showed a clear inverse relationship with linear energy density (LED = P/v): as more energy is deposited, the melt pool becomes larger, convective flows homogenize temperature, latent heat buffers solidification, and cooling slows.

The team benchmarked their results against multiple baselines. The widely used Eagar-Tsai analytical conduction model over-predicted cooling rates by an average of 119%. Prior in-situ measurements using illuminated imaging alone (without experimental temperature data) over-predicted by 134%. OpenFOAM and FLOW-3D multiphysics simulations fell within the experimental uncertainty range for powder-layer conditions, validating both the measurement technique and the simulation frameworks.

Perhaps the most compelling validation came from the microstructure. The team cross-sectioned each single-track scan, etched the specimens, and measured cellular arm spacing via SEM. Cell spacing and cooling rate are linked by a well-established power-law relationship (λ = A · Ṫⁿ), where the exponent n typically falls between −0.28 and −0.40 for steels.

Fitting their experimental data, the team obtained n = −0.27 and A = 16,500 nm·(s/K)ⁿ — an exponent closely matching literature values and a pre-factor marginally lower than previous reports. The R² of 0.958 represents a remarkably strong correlation between in-situ cooling rate measurement and ex-situ microstructural ground truth. This is the first time such a correlation has been demonstrated with experimentally measured (not modeled) LPBF cooling rates.

The implication is significant: previous studies that inferred cooling rates from cell spacing using literature correlations may have systematically overestimated the cooling rates experienced in LPBF. For a given cell spacing, the true cooling rate in 316L SS L-PBF appears to be somewhat lower than previously inferred — a finding with implications for microstructure prediction models across the field.

High-speed imaging at melt-pool resolution generates enormous data volumes — too much for continuous, part-scale monitoring. But the team discovered something that could change the game: simple correlations between individual camera channel signals and cooling rate. The total red channel signal (sum of red pixel values in each frame) showed an R² of 0.84 with cooling rate. Melt pool length vs. red channel signal showed R² = 0.72.

These correlations open a practical pathway. A simple, low-bandwidth photodiode sensitive to the same wavelength bands as the camera could be calibrated using the hybrid imaging setup and then deployed for continuous monitoring over entire builds — capturing cooling rate trends across every layer without the data bottleneck of full-frame high-speed video. Calibration would need to be material-specific and performed under production-like conditions, but the concept is proven.

This work matters because it bridges a critical measurement gap with a practical, low-cost tool. The experimental setup — a color camera, a blue laser diode, a hot mirror to separate thermal from illumination wavelengths — costs a fraction of a synchrotron beamline or even a research-grade infrared camera. The technique can be retrofitted to existing machines. The data processing pipeline, while currently offline, is architecturally suited for real-time implementation.

The implications extend across the LPBF value chain. Process developers gain an experimental tool to map cooling rate vs. parameters for any material. Simulation engineers gain validation data against which to benchmark their multiphysics models. Quality engineers gain a potential in-line monitoring metric that correlates with the most important microstructural features. And the broader community gains a demonstrated methodology that can be replicated, adapted, and improved.

Limitations are honestly acknowledged. The technique currently resolves only the tail of the melt pool — the hottest regions saturate the camera sensor at practical exposure times. Cell spacing variation within a single melt pool cross-section exceeds the positional variation along a track, limiting the spatial precision of cooling rate–microstructure correlations. And the transition from single-track experiments to multi-layer, multi-hatch production builds will require overcoming challenges in data management, illumination uniformity, and segmentation robustness.

But these are engineering challenges, not fundamental barriers. The principle is established: you can measure true LPBF cooling rates with a color camera, a blue laser, and the right algorithms.

Reference: Weeks, C.M., Myers, A.J., Wassermann, N.A., Quirarte, G., Beuth, J.L., Narra, S.P., Singh, S., & Malen, J.A. (2026). Combined thermal and illuminated imaging for cooling rate measurements in laser powder bed fusion. Additive Manufacturing, 127, 105278.