Real-Time Image-Based Feedback Control of Laser Powder Bed Fusion

Laser powder bed fusion (LPBF)² is a metal additive manufacturing process. The typical LPBF system is shown in Fig. 1: a high-power laser scans the surface of the metal powder and selectively melts it in a layer-by-layer fashion. LPBF can produce fully functional metal parts of complex geometry, such as organic shapes or lattice structures. However, LPBF-produced parts are prone to defects, such as porosity or distortion. As a result, significant manual effort is required to dial in process parameters for a particular part [1].

The consensus in the existing literature is that LPBF can benefit from process monitoring and control [2]. LPBF monitoring has seen significant progress recently [3]. Particular attention has been devoted to the high-speed observation of the melt pool, as the melt pool is of prime importance for process quality outcomes [1]. In contrast, only a handful of controllers for LPBF have been experimentally demonstrated [2], major reasons being (1) the closed architecture of the commercial LPBF machines and (2) the stringent requirements on sensor resolution and system bandwidth [4]. The majority of published studies employ feedforward schemes, such as control of laser power based on local geometry [5], control of laser power based on a data-driven predictive model [6], control of a cross-sectional area of a bead on a plate [7], layer-to-layer control on a line-by-line basis [8], and layer-to-layer iterative learning control [9].

For feedback control of LPBF, the seminal work of Kruth et al. [10] hinted at the use of a camera in the feedback loop but did not provide details on the closed-loop control design. The subsequent work by Craeghs et al. [11] used only a photodiode signal to control the laser power. Another photodiode-based controller was recently reported by a different group of authors [12].

To the authors’ best knowledge, substantial work on image-based melt pool monitoring has been reported in the literature, but a systematic study of real-time image-based feedback control of the LPBF process has not been reported. Thus, the key contributions of this work are as follows:

1. design and implementation of a real-time control system integrating the melt pool imaging with real-time in-line laser power control on a custom LPBF testbed;

2. experimental modeling of the melt pool image response to changes in the input laser power under a variety of conditions; and

3. experimental validation of the melt pool reference tracking with a real-time feedback controller, on a single-track and multi-layer part-scale.

The results presented in this letter are based on the open-architecture LPBF testbed, described in Ref. [9]. The testbed follows the typical hardware architecture of the industrial LPBF machines (as in Fig. 1) and is capable of producing full-scale metal parts as large as 50 × 50 × 100 mm. The testbed is instrumented with a near-infrared (NIR) camera that observes the melt pool through laser focusing optics. The setup is described in detail in Ref. [13] and is similar to Refs. [6,8,11,14]. In such a “coaxial” setup, the camera’s field of view is always centered on the melt pool, which allows for a high spatio-temporal resolution. The camera’s integration time is 400 μs, the sampling rate is 2 kHz, and the instantaneous field of view is 22 μm per pixel. A typical melt pool image is shown in Fig. 2. The camera image acquisition is triggered by an external hardware signal that is synchronized with the laser operation.

While the sampling rate of the camera is lower than that of a photodiode (up to 100 kHz), NIR imaging can observe melt pool geometry which affects the part density and microstructure [15]. Meanwhile, a photodiode averages the thermal field it observes, and thus obscures the melt pool geometrical changes. Figure 3 shows two coaxial images acquired during a build. The images are noticeably different but the total emission in the field of view is the same for both images. A photodiode would not be able to distinguish these two melt pool states. In literature, Craeghs et al. followed up on their work [11] with a suggestion that, for control purposes, a camera-based control could be superior to a photodiode-based one, as the latter only provided “small improvement” [16]. Therefore, there is a need to study camera-based feedback.

The process monitoring and control logic is implemented via the in-house developed c++ code called from within America Makes software [17]. Unlike the commercial LPBF machines, the laser power on the testbed can be updated within individual scan lines. This is achieved by asynchronous calls to the scanner control library, overriding the power level determined from the build file. Laser power can be modified in both a feedback (based on images) and a feedforward (based on a pre-defined profile) manner. The control system presented in this letter uses melt pool image feedback to control the laser power, as shown in Fig. 4.

Figure 5 illustrates the synchronized interaction between the laser scanner control board, the melt pool camera, and the process control logic, implemented at the loop closure rate of 2 kHz. It takes less than 400 μs to retrieve the frame, process it, and issue the power update command.

To the authors’ best knowledge, this is the first report of an experimental image-based TF identification used for feedback control in LPBF. Based on the previous work by the authors [18], the dominant behavior of C₁₀₀ can be described by a first-order model P(s) = K/(τs + 1). A similar conclusion was made in Ref. [11]. Figure 6 shows the typical response of C₁₀₀ signal to the changes in the input laser power and further supports the claim of the first-order behavior.

Previously, the authors have reported that the melt pool response varied with scanning speed and substrate [18]. It also evolved with layer number, settling after an initial transient of 10–20 layers. For this work, the authors identified TF for C₁₀₀ at 800 mm/s scanning speed, as it is the speed setting used to produce the parts on the testbed. Two types of single scan line tests were performed to control for the layer transient:

1. Bare metal tracks, where no powder is present and the laser scans a bare metal solid substrate.

2. Built-up part tracks, where a layer of powder⁴ is deposited on top of the LPBF-built substrate of sufficient (50 layers, e.g., 1.5 mm) thickness, so that the layer-related transient in TF parameters has converged.

The response of the “hot spot” C₁₀₀ signature to the linearly ramping power is shown in Fig. 7. Signals from bare metal and built-up part tests differ in magnitude and slope. Moreover, the response to ramping power shows a concave shape, indicating non-linear plant behavior.

An effective feedback controller must address the variability of the plant with respect to scanning speed and substrate variation. In this work, such behavior is addressed via (1) a localized linear model around the nominal process parameters (on the testbed, 200 W for laser power and 800 mm/s for scanning speed are usually used) and (2) designing around the largest possible plant gain, i.e., worst-case gain margin. In future work, different process model structures can be investigated, e.g., a parameterized set of local models, or a fully non-linear input-output model, which is outside the scope of this letter.

Figures 8(a) and 8(b) illustrate the models identified around the nominal process parameters. For the built-up part tracks, the gain was found to be around 0.47 pixel/W. Even though the built-up part test is more representative of the actual LPBF process, the model identified in bare metal tests was used for the control design because the plant gain was found to be larger.

The estimated parameters of C₁₀₀ TF in bare metal tests were as follows: $K=0.551±0.011pixel/W$⁠, τ = 447 ± 15 μs, with normalized root-mean-square error (NRMSE) of 0.43. The estimated time constant is in reasonable agreement with existing literature [4]. The estimate is based on the bare metal test with five lines scanned with different power profiles. The test was performed three times on separate bare metal plates, i.e., 15 lines total. Uncertainty is indicated by a standard deviation of the estimate.

In this work, the image-based feedback control of the melt pool is implemented, such that a desired melt pool reference is followed, as shown in Fig. 4. The remainder of the letter presents an experimental demonstration of a reference tracking on the single-track and then on the multi-layer part scale.

A time-varying reference tracking can be achieved with the proportional-integral controller, defined in serial form as C(s) = K_p(1 + 1/sT_i), where T_i = 0.0005 in this work. Bare metal scans were used for the experimental controller verification. As illustrated in the previous section, scans on the bare metal result in the largest plant gain, while time constants in bare metal and built-up cases appear to be similar. Thus, a controller that has sufficient stability margins in the bare metal setting will have sufficient stability margins for the part-scale operation.

The closed-loop response is affected by delays (e.g., camera integration time, image transmission, actuation delay), which limit the achievable bandwidth. Conservatively, this delay T_d is assumed to be equal to one control-loop time-step, i.e., T_d = 500 μs, and the controller is designed to account for that.

The frequency response of the loop transfer function $L(s)=e−sTdP(s)C(s)$ can be evaluated analytically based on the nominal model from Sec. 3: the plant gain is $0.55pixel/W$⁠, and the time constant is 447 μs. From Fig. 9, the gain margin for the controller with $Kp=1.2W/pixel$ is at least 7 dB. However, oscillations at 400 Hz were observed in the output signal (Fig. 10(a)). Un-modeled dynamics or other effects could be responsible for this behavior and are currently being investigated. The oscillatory behavior is also evident in the track appearance, as illustrated by Fig. 11. The output of the closed-loop system with a tuned gain $Kp=0.8W/pixel$ (gain margin of 11 dB, time delay margin of 1.4 ms at 145 Hz) is shown in Fig. 10(b). The output is free of oscillations, and the time-varying output reference is tracked well.

The feedback controller designed in Sec. 4.1 was then applied to the LPBF process on a part-level scale. To ensure no transient layer-to-layer behavior, 50 layers of substrate were built first. Then, four identical square geometries of 15 × 15 mm cross-section were built in parallel: two with a constant laser power (open-loop), and another two under the feedback control. The nominal scanning parameters were as follows: laser power 200 W, scanning speed 800 mm/s, hatch spacing 90 μm. The layers were hatched at 0 deg, 45 deg, and 90 deg. Layers 1, 4, 7 · · · were scanned vertically, layers 2, 5, 8 · · · were scanned diagonally, and layers 3, 6, 9 · · · were scanned horizontally. A total of 10 layers were built, and more than 240,000 unique coaxial images were acquired as the result.

A detailed analysis of this data is not the subject of this letter. However, a brief discussion is presented below.

Figure 12 shows the filtered output signature C_f as a function of the spatial coordinates x, y on the build plate. Four squares from a single layer are shown. The bottom row squares were built with the feedback controller regulating the output, while the top row squares were built in the open-loop mode. The output time series for the OL1 and FB1 squares are shown in Fig. 13, while Fig. 14(a) shows the feedback input signal corresponding to Fig. 13(b). From Fig. 13(b), it is clear that the feedback control successfully regulated the plant’s output, while the open-loop output signal deviated from a constant level.

In addition, signals acquired from the two open-loop squares clearly differ even within the same layer. For example, the OL2 square has more locations with higher values of C_f than the OL1 square. A plethora of reasons could cause such differences in open-loop parts, e.g., variations in the powder re-coating quality, spatter from the parts landing on the build surface, or thermal interactions between the parts. Meanwhile, the feedback-controlled signals appear to match more closely. The feedback control significantly reduced the error norm for all 10 layers, as shown in Fig. 14(b). Based on these findings, real-time feedback control could result in better part-to-part consistency.

For the OL squares, an increase of the coaxial signature is observable in proximity to the corners. This effect is possibly caused by the diagonal scanning pattern at 45 deg hatch angle, where the melt pool returns to the previously heated area at an ever-increasing rate. Again, no signature increase of such kind is observable in the feedback-controlled parts.

A detailed analysis of the data from the feedback-controlled builds, as well as an investigation of the influence of the feedback control on part quality metrics, is necessary as follow-up work. In the context of this letter, the applicability of the designed system for the real-time image-based control in LPBF was successfully demonstrated.

The real-time LPBF control system presented in this letter enables easy control algorithm development and implementation, e.g., image-based process monitoring and feedback control of the melt pool signature. The system allows for real-time laser power modulation in both a feedforward and a feedback manner.

The system’s capabilities for empirical modeling were experimentally demonstrated. Prescribed input signals were used to identify empirical models for different reductions of the multidimensional image data. Models’ dependencies on process parameters and other variables, e.g., substrate material, were stated. At 2 kHz sampling, the system enables real-time implementation of feedback control schemes. A time-varying reference tracking was successfully experimentally demonstrated on a single-track scale. Furthermore, the melt pool image signature was experimentally regulated to the desired level on a part scale, in real-time, using feedback control.

Potential directions for future work include process control for the LPBF builds with complicated geometries, targeting improved process outcomes and better part quality. Process disturbances, e.g., heat buildup or scanning-induced thermal disturbances on a part scale, are also being investigated.

This work is supported by NSF Data-Driven Cyberphysical Systems Award #1645648 and by the State of New York ESD/NYSTAR program.

There are no conflicts of interest.