ABSTR ACT In situ melt pool monitoring is a set of technologies widely deployed on industrial, metals-based laser powder bed fusion (LPBF) additive manufacturing (AM) systems. This study investigates the use of a calibrated tungsten ribbon lamp as a reference standard to calibrate a photodetector based, on-axis melt pool monitoring system. Calibration demonstrates two functions: (a) enable a reference for measuring and ensuring system repeatability, and (b) enable reference to physical temperature values based on the measured photodetector signals. The second function is explored in this paper. A regression-based model is derived based on bichromatic Planck thermometry theory. The calibrated tungsten lamp is then placed within a LPBF system, and resulting photodetector signals are measured at different lamp temperature set points to calibrate the model. Finally, several additional characterization tests and their results are presented verifying the temporal response of the lamp, measurement noise as a function of sampling time, and spectroscopic measurements of the LPBF optics and their potential effect on temperature calibration. A framework is also developed to normalize temperature readings across the build plate to remove location-dependent optical artifacts. KEYWORDS: additive manufacturing, in situ process monitoring, coaxial melt pool monitoring Introduction Process Monitoring in Additive Manufacturing In situ process monitoring in additive manufacturing (AM) utilizes various sensor modalities to detect variations in the AM process, with the aim of associating these variations with defects in finished components (Everton et al. 2016 McCann et al. 2021 Grasso and Colosimo 2017 Spears and Gold 2016). Nondestructive evaluation methods are not robust to geometry. Some become less reliable when components have intricate surfaces, and others struggle as the total amount of material in a component’s cross section increases (Chauveau 2018). Ultimately, a goal of AM process monitoring is to detect component defects in arbitrary geometries. One common technique is coaxial melt pool monitoring, in which single point or image-based photodetectors are optically aligned with the laser. One or more wavelength range is selected for measurement. Process monitoring is critical in AM due to the presence of stochastic defects as well as complex but predictable process deviations. For example, if the same component is fabricated multiple times, defects may occur in similar locations each time due to complex thermal behavior. In the case of stochas- tic defects, it is entirely impossible to predict a priori where defects will form with reasonable reliability if the same com- ponent is manufactured multiple times, defects may occur in entirely different locations with different frequencies (Snow et al. 2020). In the case of predictable process deviations, these may be modeled a priori, but process monitoring is desired to confirm that the process occurred as expected. In the case of stochastic defects, process monitoring or post-process inspec- tion is essential to validating the components. Due to the complexity of the relationship between process monitoring data and final part quality, researchers have turned to machine learning and deep learning for component certifi- cation. These methods are able to produce rich representations of process monitoring data (Meng et al. 2020). However, these models, as interpolators, are not necessarily generalizable. It is critical that process monitoring signals are consistent across machines and across the build plate of a given machine for developed machine learning models to function. For example, a convolutional neural network (CNN)—commonly used to process AM process data—is comprised of shift-equivariant layers under the assumption that the identity of process signals THERMAL CALIBRATION OF RATIOMETRIC, ON-AXIS MELT POOL MONITORING PHOTODETECTOR SYSTEM USING TUNGSTEN STRIP LAMP BY BRETT DIEHL*†, ALBERTO CASTRO*, LARS JAQUEMETTON*, AND DARREN BECKETT* * Sigma Labs Inc., Santa Fe, New Mexico, USA † brettd@sigmalabsinc.com Materials Evaluation 80 (4): 64–73 https://doi.org/10.32548/2022.me-04271 ©2022 American Society for Nondestructive Testing ME | TECHPAPER 64 M A T E R I A L S E V A L U A T I O N • A P R I L 2 0 2 2

(for example, nominal versus anomalous) is shift-invariant (Bronstein et al. 2021). It has been experimentally demon- strated that CNNs are less fragile than their nonconvolutional counterparts while processing image-based in situ AM data (Snow et al. 2021). If there are deviations in nominal/anoma- lous process signals over the course of the build plate, then this assumption is broken, and the accuracy of CNNs for detecting defects across the entirety of the build plate may be signifi- cantly reduced. A similar argument nonspecific to any deep learning architecture can be constructed: that process signals should be uniform across machines, or models trained on data from one machine will not be applicable to other machines. In order to ensure data uniformity across machines, all else being held constant, sensors need a universal standardization (cali- bration) scheme. In addition to concerns relating to system repeatability and reproducibility, a standardization scheme may enable reference to physical temperature by relating to a thermally calibrated radiant emission source. Laser powder bed fusion (LPBF) is a thermally driven process, and as such, the fabri- cation quality is intrinsically tied to the physical temperatures in and around the melt pool. The range of temperatures that exist in and around an LPBF melt pool are extremely complex, known to span several thousand degrees Celsius (Hooper 2018 Deisenroth et al. 2021), and include a multitude of thermally radiating phenomena such as hot particle spatter (Khairallah et al. 2016), plumes (Dunbar and Nassar 2018), reflections, and the like. While singular values for melt pool temperature inevitably superimpose a multitude of these different thermal radiating phenomena, a thermally standardized system may still enable computation of temperature-based data features such as radiance temperature, cooling rates, or thermal gradi- ents. In turn, these physics-based thermal metrics may offer greater sensitivity to processing quality (for example, devia- tions in these metrics may correspond to AM process devia- tions, thereby supporting more robust statistical or machine learning–based models for practical process monitoring and control applications). One standardization scheme is proposed by the authors for a dual photodiode system, which utilizes two on-axis single-point photodetectors, each one behind a band-pass filter with transmission chrematistics of λ2 and λ1, where λ2 and λ1 are two visible light wavelengths that are similar in wavelength to each other ( 50 nm difference between the two, λ1 λ2) (Lane et al. 2020). These wavelengths were chosen due to their distance from any prominent emission lines of metals commonly used in the AM process. Light emitted from the melt pool with these wavelengths is therefore not due to electron transition but due to thermal radiation. The signals from these sensors are used to calculate a temperature value using bichromatic Planck thermometry (ratio, dual-band, or two-color pyrometry) (Saunders 2007 Coates 1981), which forms the basis of a process monitoring data metric called thermal emissions Planck (TEP). The photodetectors’ analog outputs are digitally sampled at 200 kHz and stored as 32-bit floating point values. An additional photodetector, whose thermal calibration was discussed using a separate method by Lane et al. (2020), is sampled to derive another metric called thermal energy density (TED). The setup is illustrated in Figure 1. Because the path length of melt pool–emitted light into the coaxial photodetectors varies with position and variations may exist between different machines, standardization is import- ant to normalize melt pool signals to a physical phenomenon: temperature. In this paper, a tungsten calibration lamp, which is known to sufficiently replicate a blackbody emitter, is used to calibrate the photodiode system using the principles of bichro- matic Planck thermometry. Bichromatic Planck Thermometry Calculations Planck’s law indicates that the wavelength-dependent radiance of a blackbody source as a function of temperature can be given by Equation 1: (1) I(!,T) = 2hc2" !5 1 e hc ! kBT 1 where h is Planck’s constant, c is the speed of light, ε is the emissivity of the source, and kB is Boltzmann’s constant, resulting in units of I(λ,T) as W × sr−1 × m−3. The ratio of emitted intensity at two different wavelengths at the same temperature can be calculated by Equation 2. Note that the emissivity value cancels itself out, so long as the emissivity of the melt pool at each wavelength is the same. Beam splitter 1 Beam splitter 2 Beam splitter 3 F-theta lens X X Y Y Melt pool Heating laser TED sensor TEP sensor, λ 1 TEP sensor, λ 2 Heating laser Scanning galvo mirror system Radiant emission from melt pool Figure 1. Diagram of the dual photodiode system with thermal energy density (TED) and thermal emissions Planck (TEP) sensors. A P R I L 2 0 2 2 • M A T E R I A L S E V A L U A T I O N 65