Large-scale metal additive manufacturing: a holistic review of the state of the art and challenges

In gas metal arc welding, an arc is struck between a substrate and a consumable wire electrode that is fed through the welding torch, where it is melted and deposited onto the substrate. The molten material is protected from moisture and oxidation through the use of shielding gases, which are typically a combination of inert (Ar) and active ( ). The shielding gas varies things like the stability of the arc, metal transfer, and penetration of the weld and is tailored to the material being deposited. The wire is continuously fed as the welding torch is translated in the geometry of the weld or AM part. The consumable electrode is either a solid wire or a cored wire, with a powdered interior in various ferrous and non-ferrous compositions. The current is directly proportional to the deposition rate but inversely proportional to the electrode extension, which is the distance between the end of the wire guide and the tip of the electrode, shown in Figure 4(a). The arc voltage is a means of electrically quantifying the physical length of the arc and can be affected by many factors, including: electrode composition and size, shielding gas composition, electrode extension and the length of the welding cable [170]. The deposition rate for GMAW in terms of AM is material dependent, but is in the range of 15–160 g min [126, 171, 172].

OPEN IN VIEWER

Table 2.

A listing of various powder fed deposition technologies and associated parameter based on the material being deposited. The values listed provide the maximum and minimum for each parameter and the authors who's parameters fall within those ranges.

Process	Material	Travel speed (mm s )	Heat input (W)	Spot size (mm)	Layer height (mm)	Material feed rate (g min )
	Steels	2.5 [50] – 20 [51]	360 [52] – 2600 [51]	1.2 [53] – 2 [51]	0.25 [52] – 0.5 [53]	2 [53] – 20.4 [50]
		Within range: [50, 52–56]	Within range: [53–56]			Within range: [50, 51, 54–56]
laser DED	Ti–6Al–4V	2 [57] – 17 [58]	330 [59] – 7000 [60]	0.3 [59] – 8.6 [61]	0.3 [57] – 3 [60]	1 [62] – 59 [63]
		Within range: [59–67]	Within range: [57, 58, 61–67]	Within range: [57, 60, 62–64, 66, 67]	Within range: [58, 61, 65]	Within range: [57–62, 64–67]
	Aluminium	6 [68] – 16 [69]	120 [70] – 3600 [68]	0.6 [71] – 3.5 [68]	0.5 [68]	0.66 [70] – 23.2 [72]
		Within range: [70–74]	Within range: [69, 71–74]	Within range: [72]		Within range: [71, 73, 74]
	Nickel (Inconel 625)	6.7 [75] – 25 [76]	1500 [77] – 3000 [75]	0.4 [77] – 3 [75]		6 [75] – 33.3 [76]
		Within range: [77]	Within range: [76]	Within range: [76]		Within range: [77]
	Nickel (Inconel 718)	2 [78] – 26.6 [79]	250 [78] – 4000 [79]	0.8 [80] – 5 [79]	0.1 [78] – 0.5 [80]	1.2 [67] – 36.6 [79]
		Within range: [67, 76, 80–84]	Within range: [67, 76, 80–84]	Within range: [67, 76, 81, 82]		Within range: [76, 81–84]
	Co–Cr	5.5 [85] – 20 [86]	200 [87] – 410 [85]	0.25 [88] – 0.7 [87]	0.25 [87] – 0.5 [85]	0.57 [88] – 5 [86]
		Within range: [87, 88]	Within range: [86, 88]	0.25 [88] – 0.7 [87]		Within range: [87]
	W	2 [89] – 5 [90]	200 [89] – 2000 [90]	0.6 [89] – 3 [90]	0.8 – 0.9 [89]	7 [89] – 8 [90]
PTA	Steels	1.3 – 1.7 [91]				25 – 35 [91]
	NiCrBSi	10 [92]	1100 [92]	4.7 [92]	0.75 [92]	20 [92]

OPEN IN VIEWER

Table 3.

A listing of various wire fed deposition technologies and associated parameter based on the material being deposited. The values listed provide the maximum and minimum for each parameter and the authors who's parameters fall within those ranges.

Process	Material	Travel speed (mm s )	Heat input (W)	Spot size (mm)	Layer height (mm)	Wire feed speed (mm s )
laser DED	Ti–6Al–4V	1.4 [93] – 10 [94]	1000 [95] – 3500 [94]	2.5 [93] – 5 [94]	1 [94] – 1.28 [96]	30 [96] – 40 [94]
		Within range: [95, 96]	Within range: [93, 96]	Within range: [96]
	Nickel (Inconel 718)		5000 [97]	1 [97]
GMAW	Steels	2.5 [98] – 30 [99]	3500 [100] – 8400 [101]		0.5 – 2 [99]	28 [102] – 166 [101]
		Within range: [102–109]				Within range: [98, 99, 102, 103, 105–109]
	Ti–6Al–4V	1.5 [110] – 9.4 [111]	1430 [110] – 12500 [112]	6 [113] – 10 [110]	14 – 16 [110]	7.2 [110] – 142 [111]
		Within range: [112–114]	Within range: [111]			Within range: [112]
	Aluminium	6.13 [115] – 22 [116]	3360 – 7360 [116]			100 [115] – 250 [116]
		Within range: [117]				Within range: [117]
	Nickel (Inconel 718)	6 [118] – 10 [119]		12.8 [119]	1.7 [119] – 2.8 [118]	10 [119] – 116.6 [118]
		Within range: 6.5 [120]				Within range: 33.3 [120]
	Nickel (Inconel 625)	6.3 [121] – 10 [122]	2160 [122]			108 [122]
	Magnesium	3.3 – 16.6 [123]	400 – 1400 [123]	5 [124]	3 [124]	30 [124] – 200 [125]
		Within range: [124, 125]	Within range: 541 – 857 [124]
	Copper alloys	6.6 [126] – 8.3 [127]	4620 [127] – 7424 [126]			117 [127]
	Co–Cr	2.1 [128]	1454 [128]		3.5 [128]	75 [128]
GTAW	Steels	2.92 [129] – 7 [130]	1920 [131]			16.67 [129] – 58 [130]
		Within range: [131]	1920 [131]			Within range: [131]
	Ti–6Al–4V	0.27 [132] – 6.7 [133]	1320 [134] – 2200 [94]	5 [133] – 9.1 [94]	1 [94]	10 [133] – 128 [133]
		Within range: [94, 130, 134–136]		Within range: [134, 135]		Within range: [94, 110, 130, 132, 134–136]
	Aluminium	3.3 [137] – 100 [138]				17 [137] – 160 [138]
	Nickel (Inconel 718)	5 [139]		10 [139] -16 [140]		25 [139]
	Magnesium	3.3 [141] – 5 [142, 143]			1.25 – 2.5 [141]	19.2 [142, 143] – 33.3 [141]
	Copper Alloys	1.6 [144]				21.6 [144]
	Co–Cr				1.1 [145]	16.6 [145]
	W	2 [146]				35 [146]
PTA	Steels	0.6 [147] – 2 [148]	350 [147] – 3510 [149]			9 [148] – 28 [147]
	Ti–6Al–4V	4 [150, 151]	2700 – 5400 [151]		1.5 [150, 151]	58 [150, 151]
	Nickel (Inconel 625)	21.6 [152]			1.2 [152]	3 [152]
EB	Ti–6Al–4V	2.4 [153] – 18 [153]	690 [154] – 8500 [155]	1.2 [156]	1 [157]	14 [158] – 141 [153]
		Within range: [154–156, 158–165]	Within range: [153, 156, 158–165]			Within range: [156, 161–165]
	Inconel 718	5 [166]	600 – 960 [166]			5.4 [166]

Three traditional transfer modes are commonly used with the GMAW process, which are: spray, globular and short circuiting [173]. Cold metal transfer (CMT) is a modified subsidiary of short circuiting, where the mechanical movement of the wire electrode is synchronized with the electrical control parameters [111]. Instead of increasing the current during the short circuit phase, the current is dropped, extinguishing the arc and limiting the amount of thermal energy transferred to the deposit [174]. The electrode is then retracted, pinching the molten material, depositing it into the melt pool. The current is then increased to reignite the arc and the process repeats [175]. The decrease in thermal energy transfer reduces the heat accumulation in multi-layer deposits, which can be characterized by the finer grain structures when compared to continuous welding techniques [111, 176]. This can be seen in Figure 5 [177], where the lower heat input and heat accumulation is characterized by the finer grain structure. Furthermore, the pulsing of the arc has been shown to sever dendrite arms, increasing the heterogeneous nucleation sites, further refining the microstructure [115, 132]. It also drastically reduces the dilution of previously deposited material, reducing the amount of material being melted with each pass and possibly reducing the number of thermal cycles [175, 178]. Thus these reasons make CMT the most viable option for wire and arc additive manufacturing (WAAM). It should be noted that although there is a reduction in heat input and thermal cycles compared to continuous welding, WAAM deposits still suffer from heat accumulation, cracking, porosity, delamination and anisotropic microstructures [179]. The first study of using GMAW for AM was conducted by Dickens et al., who tried to expand the realm of 3D welding from large pressure vessels, to more complex geometries [180].OPEN IN VIEWER

Gas Tungsten Arc Welding is similar to the GMAW process, but the arc is struck between a non-consumable tungsten electrode and the workpiece. A filler metal can be fed manually or mechanically into the arc, where it melts and is deposited onto the substrate. Multiple filler metals can be fed simultaneously to increase the deposition rate and allow for the customization of the material being deposited. Inert shielding gasses (typically Ar or He) protect the melt from oxidation while also affecting weld bead geometry. The polarity of the system can be altered from DC to AC if the material being deposited is prone to forming passive films [181]. The microstructure and mechanical properties of AM deposits are highly dependent on the material feeding orientation [182, 183]. Some of the materials that have been deposited include: TiAl [184], Fe–FeAl functionally graded material [185], FeAl [186], Ti64 [94, 133, 187, 188], Al [189] and Ni alloys [190].

GMAW and GTAW offer a cost-effective means of AM, with techniques that are already common industrial practice. The ease of integration with robotic control and gantry systems, coupled with the high deposition rates, makes these technologies enticing for large-scale additive manufacturing [191]. However, some complications reside when using a welding heat source for AM. Distortion and residual stresses are common side effects of the concentrated heat flux generated from an arc [192]. Inconsistent bead geometries can lead to poor surface finish and dimensional accuracy [193]. Research has predominately been on GMAW, which is speculated to be due to the added complexity of integrating a wire feeding system with the robotic system. Ensuring the feeding angle is constant during deposition would increase the difficulty of path planning and building strategies. The continuous heat input experienced during GTAW could cause increased heat accumulation, resulting in manufacturing defects such as the slumping of different features. Furthermore, GMAW's ability to easily strike and extinguish an arc increase the thermal control during the build by extinguishing the arc after each pass to allow for the part to cool. The tungsten electrode in GTAW also requires frequent sharpening to maintain arc characteristics, decreasing the production rate of large-scale parts.

Plasma transferred arc utilizes a non-consumable tungsten electrode, similar to that seen in GTAW; however, there are some stark differences between the processes as can be viewed in Figure 4(b). Generally, there are two inert gas inlets: the plasma and shielding gas. The gases used in this process (such as Ar) are chosen due to their low ionization potential, making it easier to strike an arc between the electrode and the substrate. The flow of the plasma gas allows the arc to be self-sustaining, while the shielding gas protects the melt from the surrounding environment [194]. The plasma is constricted by a nozzle, changing the arc shape from the traditional bell shape to columnar, increasing the energy density [195]. The feeding material can either be wire, or powdered materials, allowing for a large degree of compositions and functionally graded parts. The deposition rate is the highest of the welding techniques are 33–166 g min [194]. Some of the materials that are being explored with PTA for additive manufacturing are: Ni alloys [196–199], Ni-WC [92], Ti [200], functionally graded Fe–Ni [201, 202] and stainless steel alloys [203, 204].

2.3. Laser-based direct energy deposition

Laser-based DED techniques share the basic principles with the aforementioned plasma-based methods, where the main difference lies in the energy source. For laser systems, a series of lenses are used to focus a laser beam to melt the desired material [205]. The laser source can vary depending on the particular application. lasers are better suited for low precision, simpler geometries, where an Nd-YAG laser is better suited for finer, complex geometries [206]. The feed material for laser DED can be powder, wire, or a combination of the two, depending on the application. A schematic of a typical laser system is shown in Figure 4(c). The deposition rate can be up to 25 g min with varying deposition efficiencies depending on the components geometry [207]. Both heat sources share a Gaussian energy distribution, with the highest temperatures in the centre of the melt. However, the heat flux provided by a laser source is upwards of 1 kW mm with a 2-mm-diameter spot size [208–210], while a plasma provides upwards of over 16-mm-diameter spot size [194, 211]. Another critical distinction is the safety precautions that workers must abide by during laser DED. To strike an arc, the workpiece must be electrically grounded and can only be sustained within a certain stand-off distance. Commercial lasers do not have any of these pre-requisites, meaning they can theoretically be directed at any surface. Additionally, a laser can be reflected by certain metallic surfaces that can damage facilities or personnel. Thus proper control measures must be implemented to ensure the safety of anyone working with this equipment.

2.4. Materials

In all AM techniques, the feedstock metals can be in the form of wire or micron-size powder. Powder metals are typically much more expensive than their wire counterparts, but offers material compositions that are not able to be drawn into a wire. An example of this are higher reinforcement loaded MMC's and intermetallics, where the inherent brittle nature of these materials make it unsuitable for wire applications [212]. However, the deposition efficiency of wire fed systems is beyond what is possible with powder [213]. Moreover, the storage of metal powders requires significantly more safety precaution than that of metal wires and the higher surface area to volume ratio makes them more susceptible to oxidation [214]. The quality of the feedstock is of utmost importance, as porosity in the feedstock stock powders has been shown to drastically increase the porosity of the printed part [91]. Poor surface quality and diameter variances of wire feedstock can trap moisture and hydrocarbon residue during the deposition process, resulting in porosity in the final deposit [215–218]. This section of the report will outline the common materials and the as-built microstructures found in the above-mentioned AM techniques, as shown in Table 2 and Table 3. The variation in mechanical properties of AM deposits will be compared to conventional manufacturing where applicable, and the microstructural justification for differences will be discussed. The order of materials is as follows: first steels will be discussed, followed by titanium, aluminium, nickel, magnesium, copper, cobalt–chrome and tungsten alloys. It should be noted that there has been work done on energetic materials, typically in the form of metal–polymer composites. However, the printing modalities for these materials are currently limited to those suited for polymer materials and were deemed out of the scope of this paper. The topics discussed in Section 3 and Section 4 can be applied to the deposition of energetic materials, specifically those that utilize a deposition nozzle like direct writing, fused deposition modelling and photopolymerization [219].

Some of the first researchers to recognize the need for an advanced process planning framework capable of decomposition and multi-directional slicing of complex 3D models with overhangs were Sing and Dutta [17]. The objective of their proposed method was to improve the surface accuracy and reduce the support volume through multi-directional deposition. The decomposition sequence is as follows:

1.

choose a build direction; by default along the component's Z direction to avoid collision of the deposition head with the table,

2.

identify and decompose overhanging features (often referred to as ‘unbuildable structures’ in the literature) in build direction,

3.

determine the build direction for each sub-volume and

4.

sequence and slice each sub-volume along its computed build direction.

At the core of the approach is a recursive volume decomposition scheme meaning that overhanging features within sub-volumes are also identified. The performance of the proposed process planning framework was shown on two example 3D models, but no components were fabricated. Dwivedi et al. proposed a framework for automated process planning for LDED [18]. The process planning framework is based on first-order logic and a knowledge base consisting of rule and fact attributes represented by a semantic tree structure. The authors of the study successfully verified their framework on a radial component consisting of 5 helical blades. Ruan et al. proposed a method using the centroid axis of a component to compute the deposition direction to produce collision-free slicing directions for multi-directional deposition [291]. The basic tasks are defined as 1.

centroid axis computation and formation and

2.

collision-free multi-axis slicing based on the centroid axis.

The detection of change in build direction – and therefore slicing direction – is based on the degree of shift from the centroid axis. The slicing algorithm can produce layers of non-uniform thickness, thus requiring the deposition system to be capable of producing beads of varying geometry. The algorithm was verified on a 3D model of a hinge with overhangs on a multi-axis LDED fabrication platform. Ren et al. identified limitations with the previous centroid-axis-based decomposition algorithms for certain corner cases of axis-symmetric overhanging structures where no shift in the centroid axis occurs. Thus an algorithm combining the centroid-axis-based and boundary-based decomposition methods – where concave edges and loops mark the interface between core volume and overhanging feature (see Figure 10) – of the type as previously proposed by Singh and Dutta [17] was introduced [290]. Furthermore, the authors proposed a method for representing layers of non-uniform thickness by further decomposing the non-uniform layer into uniform sub-layers of a smaller cross-section than the parent layer. The algorithm was verified by fabricating a turbine wheel with a conical shaft and winged blades on an LDED platform.OPEN IN VIEWER

Figure 10.

Schematic representations of (a) a concave edge and (b) a concave loop as defined in [290]. (Image source: [290])

In order to further improve non-planar interfaces between a core volume and overhanging feature, Singh and Dutta further extended their previous work on multi-directional deposition [17], by introducing so-called offset slices, which are essentially non-planar layers [295]. The concept of offset slices is illustrated in Figure 11. if the base surface is non-planar, which is frequently the case for radial components with overhanging features, the build quality of the overhanging features can be significantly improved when each layer follows the same contour as the core volume and subsequently the previous layer.OPEN IN VIEWER

Figure 11.

The concept of offset slices as introduced by Singh and Dutta [295]. (a) shows a contoured base surface and (b) is the corresponding offset slices. The offset slices follow the contour of the non-planar base surface where each offset slice is equidistant to the previous one.

In order to simplify process planning and fabrication of special cases of components with overhanging features containing holes (see Figure 9a), Ding et al. proposed a framework that fills all holes and protrusions within the 3D model prior to decomposition [288]. The volume decomposition itself is boundary based, whereas, with previous algorithms, concave loops and edges are detected. After decomposition, each sub-volume is sliced into planar layers according to the identified build direction. The framework was not verified experimentally. Furthermore, due to the hole-filling operation prior to decomposition, additional post-processing is required to drill the holes.

Ding et al. introduced a process planning framework for radial components such as propellers or impellers [16], shown in Figure 12. The decomposition algorithm is based on silhouette edges, as first introduced by Singh and Dutta [17], and Dwivedi et al. [292]. The algorithm is similar to previously proposed boundary-based algorithms as it looks for concave edges and loops on the core volume. Slicing is divided into two steps (see Figure 8): 1.

planar slicing of the core volume, typically a cylindrical volume for radial components and

2.

mapping of the overhanging feature's curved geometry from a cylindrical to a cartesian coordinate system to allow for a planar representation of each curved layer, similar to the principles proposed by Singh and Dutta [295].

OPEN IN VIEWER

Figure 12.

Flowchart of an example process plan similar to the one devised by Ding et al. for propeller fabrication [16].

The process planning framework was verified on a 8-DOF robot LDED platform (see Figure 7d) by fabricating the propeller model shown in Figure 8.

It should be noted that all of the frameworks for process planning reviewed up to this point can only process components where the overhanging features are sharp concave edges or concave loops (see Figure 10), meaning that they are distinguishable from the core volume. The works reviewed in the following, however, propose process planning algorithms and frameworks designed for volumes with non-sharp edges that are more difficult to decompose (see Figure 9b and Figure 9c). Wu et al. introduced an advanced volume decomposition algorithm capable of processing volumes that are not composed of a distinguishable core and overhanging volumes (see Figure 13a) [283]. The decomposition algorithm consists of 3 major steps as illustrated in Figure 13: 1.

Coarse decomposition: A skeleton is generated based on a mean – curvature flow algorithm (see Figure 13b) followed by the computation of a distance metric–the shape diameter function (SDF) – between volume boundary and skeleton (see Figure 13c) and partitioning the mesh using the distance metric based on [296]. The partitioning algorithm identifies significant differences in the SDF and creates a boundary plane where the change occurs. When considering the bunny model, a significant change in SDF can be found at the bunny's neck, ears, and tail.

2.

Sequence planning: A graph is constructed that defines the preliminary build sequence – nodes are the sub-volumes – and the print orientation for each sub-volume is determined (see Figure 13d). The preliminary build sequence is A B C D E.

3.

Constrained fine tuning: The decomposition is refined and re-configured to satisfy manufacturing constraints (see Figure 13e and Figure 13f). For example, the bunny tail as labelled B in Figure 13d cannot be manufactured with the platform shown in Figure 7(c) due to inaccessibility. It, therefore, needs to be merged with A. In addition, A* needs to be separated into H and K since the belly of the rabbit is an overhanging feature.

OPEN IN VIEWER

Figure 13.

The volume decomposition algorithm proposed by Wu et al. [283] with (a) the input 3D model, (b) the extracted skeleton, (c) the shape diameter metric (distance of every point to skeleton), (d) the result of initial decomposition and sequence planning, (e) after merging (B into A), and (f) the final result after fine decomposition to ensure manufacturability. (Image source: [283])

The decomposition algorithm was verified experimentally on a robotic AM platform equivalent to the one shown in Figure 7(c).

One limitation of Wu et al. work is that it relies on planar layers, which imposes constraints on the manufacturability of more complex components (see Figure 9c). Dai et al. proposed a novel method utilizing curved layer decomposition relying on dimensionality reduction [284]. The algorithm is separated into the following steps as illustrated in Figure 14:OPEN IN VIEWER

Figure 14.

The volume decomposition algorithm proposed by Dai et al. [284] with (a) the input 3D model, (b) after voxel discretization and voxel sequencing where the color scheme represents the fabrication sequence by layer, (c) generated curved layers based on (b), and (d) a detailed view on a computed tool path. (Image source: [284])

1.

Discretization of the input model into a voxel grid – a discretization into small cubes – where the voxel dimensions are determined by the deposition system's resolution (Figure 14b). This is done to reduce the computational load on the following steps since the volume decomposition of the input model is posed as a global search problem.

2.

Sequencing of the voxels to obtain a sequence of voxel accumulation representing the flow of fabrication. By iterating over all voxels, satisfying manufacturing constraints can be significantly simplified. The colour scheme shown in Figure 14b represents the voxel sequencing by layer.

3.

Computation of each curved layer while avoiding voxel aliasing (see Figure 14c).

4.

Computation of a tool path for each layer using the method introduced by Zhao et al. and based on Fermat spirals [294] (see Figure 14d).

This algorithm was also verified experimentally on a robotic AM platform equivalent to the one shown in Figure 7(c). The limitations of the algorithm identified by the authors include the reliability of thin-feature deposition, fabrication errors due to the used hardware, and voids in the filling patterns of the tool path planning algorithm.

Despite the limitations of the frameworks and algorithms proposed by Wu et al. and Dai et al., their works contain important contributions to process planning of complex models with significant adoption potential to metal AM.

4.2. Tool path planning

Once the component has been decomposed and sliced into cross-sectional layers, the optimal path to accurately deposit the material within the boundaries of the cross-section is computed. This process is known as tool path planning. An optimized deposition path planning strategy results in dense parts with minimized residual stress, free of any porosity, better control of anisotropic microstructures, mitigation and minimization of heat accumulation, geometrical accuracy and a smooth surface finish [24]. In order to develop an optimal deposition path planning strategy, features that are unique to various kinematic systems and deposition technologies (consistency of deposition, motion delay, dynamics, lag) need to be considered. Notably, the varying delays and inaccuracies in deposition system motion (especially for larger systems with increased mass) and material deposition (material feeding, melting) that are difficult to predict can cause unwanted variations on the rate of deposition and therefore complicate path planning significantly [47]. Inter-layer dwell time, start-stop minimization, smooth directional changes, as well as minimization of weld path cross-overs, are some of the commonly adopted strategies to mitigate these complications [47, 49, 293]. Towards the development of an optimized path planning strategy, Ding et al. identified various requirements for WAAM such as geometrical accuracy, minimization of start-stop points, minimization of rapid directional changes caused by sharp corners in every tool-path pass, and simplicity allowing for fast implementation [47].

Ding et al. reviewed various path planning methods with respect to their suitability for WAAM, using the above-mentioned evaluation criteria. Among the reviewed path planning algorithms are: Raster [297], Zigzag [298, 305], Contour [306–308], Spiral [301, 309], Fractal Space Filling Curve [302, 310], Continuous [303, 304, 310] and Hybrid (Combination of contour and zig-zag) [19, 299]. However, Raster (see 15 a), Zig-zag (see 15 b), Contour (see 15 c), Fractal (see 15e) and Spiral (see 15f) should be entirely avoided for metal AM due to the many issues listed by Ding et al. [47]. Raster and Zig-zag suffer from poor outline accuracy due to discretization errors on non-parallel edges. Contour generates many disconnected closed curves, therefore violating the requirement to minimize start-stop points. Fractal Space Filling Curve involves many path direction change motions, violating the requirement to minimize rapid directional changes. Finally, the Spiral method is only suitable for unique geometrical models that are convex [47]. Hence, these methods will not be reviewed in detail in this section.OPEN IN VIEWER

The Hybrid method (see Figure 15h) is a combination of the Contour and Zig-zag methods in that first, the contour of the layer boundary is traversed followed by filling the interior of the layer with the Zig-zag and Contour method (see 15 d). As this method combines the advantages of the Zig-zag and Contour methods, it is particularly promising for WAAM as it meets both the geometrical accuracy and surface quality. According to Ding et al., the Hybrid method is still insufficient due to the increased amount of tool-path passes and tool-path elements [47].

Ding et al. therefore proposed a novel tool path planning method intended to address the limitations of the previously proposed methods [47] and to conform with the aforementioned requirements: geometrical accuracy, minimization of start-stop points, minimization of rapid directional changes, and simplicity of implementation. The method is henceforth referred to as Convex Polygon Generation (CPG, see Figure 15 i). In order to generate a set of simpler convex or monotone sub-polygons, and to simplify the implementation of path generation for each sub-polygon, a polygon decomposition algorithm first decomposes each 2D slice via a divide-and-conquer strategy. Then the Hybrid path planning method is used for tool path generation due to the aforementioned advantages of this planning method. After tool paths are generated for each sub-polygon, the sub-paths from each sub-polygon are connected into a closed curve that spans the entire layer, thus minimizing start–stop points [47]. This algorithm extends the Hybrid path planning method to polygons with an arbitrary complexity through convex polygon decomposition. As this method, however, also utilizes the Zig-zag method for space-filling, voids can still occur [47, 49].

To address the issue of voids while retaining geometrical accuracy, Ding et al. proposed a method based on Medial Axis Transformation (MAT), or also referred to as skeletonization, as depicted in Figure 15(j) [48]. MAT was first proposed by Blum to describe shapes [311] by generating tool paths in a contour-like fashion from the centre outwards along a skeleton to the boundary of the geometry. First, the skeleton or the branch lines are generated, followed by the generation of loops representing the tool paths at a given step-over distance, which is the distance between passes representing the resolution of the deposition system [48]. With this method, the occurrence of voids is minimized. However, there are some disadvantages, such as start- and stop points and discontinuities at the geometry boundaries and deposition beyond the geometry boundaries [49]. While these deficits can be mitigated by post-process milling, they essentially limit the MAT path planning method to hybrid manufacturing.

Further iterating on their previous work with the objective of addressing the deficits raised with MAT, Ding et al. proposed adaptive MAT [49]. The difference being that the tool-path elements are designed so that the contour of the geometry boundary is followed and discontinuous path segments are minimized (see Figure 15 k). Benefits of adaptive MAT include the capability of generating continuous tool-path elements and following the geometry contour, void-free layers, good geometrical accuracy, and thus minimal post-milling, and suitability for thin-wall structures. For adaptive MAT to produce void-free deposition, the bead geometry must be able to be varied in-situ. To facilitate bead geometry adjustment, Ding et al. developed a Neural-network-based model that takes the desired bead geometry as an input and outputs welding parameters that significantly influence the bead geometry. Moreover, the adaptive MAT algorithm is experimentally validated using the proposed deposition model [300].

In summary, some of the variants of Contour-based algorithms such as Hybrid, CPG and adaptive MAT are preferred over raster or pure Zig-zag algorithms since they are more suitable for thin wall structures and allow for improved geometric accuracy, void-free deposition, and minimization of start-stop discontinuities in tool paths. Among the more suitable tool path planning methods, adaptive MAT is preferable from the aspects of void-freeness and accuracy if in-situ bead geometry adjustments are possible or feasible for a given deposition system.

A further tool path planning method specifically designed for the particular case of thin-walled structures with varying thickness was proposed by Ma et al. [37]. The adjustment of the wall width is achieved through a weaving trajectory where the weaving amplitude is the same as the width of the thin wall. After computing the skeleton of the polygon, the centreline is then obtained (see Figure 15 l), which constitutes an approximation of the polygon's median axis. During deposition, the torch weaves about the centreline in a triangular way, as illustrated in Figure 15l. The authors of the study successfully fabricated multiple thin-walled components with gradually varying wall thickness through this weaving technique.

As can be seen from this section, process planning is an integral part of robotic metal AM and involves many algorithms and software components. The cascade of complex software needs to interface and exchange information efficiently to provide robust performance while simultaneously providing flexibility, modularity, and reusability to integrate new algorithms and software in a research environment. For robotic research platforms, the used software frameworks facilitating novel research need to be as open as possible. This enables maximum flexibility and customization for each software component across research groups within the toolchain and facilitates the integration of custom hardware (HW).

A popular open-source software framework and middleware providing such a software ecosystem for advanced robotics research is the Robot Operating System (ROS). ROS is leveraged for wide varieties of robotics research and provides structured messaging between software components, robot-specific tools and libraries, various visualization and convenience tools, HW abstraction, low-level device control, and tools and libraries for obtaining, building, writing, and executing code [312]. ROS, therefore, simplifies and facilitates robotics research and software development significantly. The ROS software package MoveIt!, for example, provides interfaces to sophisticated path planners for free-space motion and inverse kinematics solvers for industrial robot arms such as the one shown in Figure 1.

In recent years, multiple open-source software frameworks have been developed within the ROS ecosystem for the planning of complex cartesian trajectories with an emphasis on industrial robotics applications such as welding, routing, milling, deburring, and grinding. In 2015, Edwards et al. introduced a path planning software package called Descartes for semi-constrained cartesian trajectory planning [313]. The software takes a 6-DOF cartesian trajectory that can be under-defined and is generated for any industrial application. Under-defined means, for example, that there is no rotational constraint on the rotation about the vertical axis of a welding torch. This enlarges the inverse kinematics solution space such that there are more options for the joint trajectory planner to avoid collisions.

Armstrong introduced a further cartesian path planning stack (collection of packages) called Tesseract for complex industrial motion planning applications with flexibility and modularity in mind [314]. The stack offers features such as fully and semi-constraint cartesian motion planning and free space planning. A significant advantage of this package, particularly towards multi-directional deposition, is its capability to plan collision-free trajectories between two moving coordinate frames, therefore enabling planning of coordinated motion between a positioner and manipulator (see Figure 1).

While there is currently an open-source robotic AM software framework available (ROS AM) [315], providing limited 2.5-DOF slicing capabilities, tool path visualization, and AM-specific message definitions, significant limitations exist. Besides being limited to 2.5-DOF AM, there is no generalized, hardware-agnostic and computer-integrated interfacing with the hardware available since post-processors generate instructions written in a hardware-specific language that only allows for open-loop execution.

The scope of this section is limited to Ti–6Al–4V (Ti64) due to the abundance of studies conducted on this material system. There are other Ti alloys that are being studied, such as TC21 [64, 235, 237, 355–357], near β Ti alloys [358–361] and near α Ti alloys [362], but they will not specifically be mentioned.

The heat treatment of Ti64 typically includes solution annealing and aging at a range of temperatures depending on the desired mechanical properties. Typically, the solution temperature is below the β transus temperature [363]. Lower annealing temperatures result in mostly α, with some β at the grain boundaries. The higher the annealing temperature, the higher the fraction of β that will form upon cooling. However, there is a decrease in solubility of V as the temperature increases, causing the β phase to turn to α' with quenching. If any β is retained after solution treatment at higher temperatures, a martensitic transformation to α' will be induced when plastically deformed [364]. Higher cooling rates are more desirable for Ti64 to maximize the amount of supersaturated β or α', which can be decomposed to α precipitates during aging [363, 365].

Laser-based AM techniques result in a mix of columnar and equiaxed grain morphologies, depending on the thermal history of the part. Equiaxed grains tend to form closer to the edges due to the higher thermal gradient achieved at these locations [57, 58, 64, 236]. The microstructure consists of primary β with α lamellae, which form in colonies, Widmanstätten or basketweave morphology. These colonies are more prevalent along prior β grain boundaries and close to the β transus lines from the interlayer passes. This microstructure has been seen for both powder and wire fed processes [57, 60, 62, 64, 93–96, 133]. Electron beam and plasma techniques have also shown to have similar microstructures, with martensitic α' and α laths in a Widmanstätten or basketweave morphology, and a small amount of acicular α [110, 113, 114, 150, 151, 158, 161]. Defects such as pores are also prevalent in the as-built parts that cannot be removed with standard heat treatment methods [60, 366–368]. Lower annealing temperatures tend to lead to coarsening of the α laths to more plate-like morphology, with interplate transformed β [93, 95]. The α plates transform into ‘crab-like’ morphology closer to the β transus temperature [64]. Furthermore, recrystallization of β grains begins at higher solution temperatures, while the primary α laths increase in aspect ratio and decrease in volume percent. The formation of β with solution treatment has been shown to increase the corrosion resistance of AM Ti64 parts. The coarsening of the α laths decreases strength while increasing the elongation [62, 369]. Increasing the annealing time decreases the aspect ratio of the α phase while also inducing a higher amount of precipitation of secondary α in the retained β phase [57, 158]. This causes an initial spike in strength, but this decreases as the secondary α coarsens. Increasing aging times decreases the volume fraction and aspect ratios of primary α laths while increasing the volume fraction of fine secondary α. Increasing aging time slightly coarsens the secondary α but decreases the width of the primary α, causing slight increases in the strength and ductility. Aging times over 8 h will result in the globularization of the α laths. These precipitates tend to coarsen with higher subsequent aging temperatures [62]. Heat treatment has shown to reduce hardness due to grain coarsening and dislocation annihilation [64]. Under dynamic loading, heat treatment may reduce strain rate sensitivity while increasing the risks of adiabatic shear localization [57].

6.2. Ni alloys

This section will discuss the heat treatment protocols of both Inconel 718 and Inconel 625. A summary of the standard heat treatment and corresponding microstructure will be presented for each material, followed by a tabular summary of the effects of heat treatment on mechanical performance.

6.3. Steels

This section will discuss the post-processing of 316L and 17-4 stainless steel. This will include the microstructural changes from the as-built condition with heat treatment and the corresponding changes to the mechanical properties.

6.4. Al alloys

The post-fabrication heat treatment discussed in this section will be limited to the hypo-eutectic alloy AlSi10Mg, as it is the most studied of the Al alloys. An outline of the standard heat treatment will be reviewed, followed by the as-built and post heat treatment microstructure, and the corresponding effect on mechanical properties.

The typical heat treatment for Al–Si alloys is a T6 treatment, which is a solution heat treatment at 535 C, quench and artificial age hardening protocol at 158 C for 10 h, outlined in the ASM Metal Handbook vol. 2 [412]. The solution treatment is done close to the eutectic temperature of the alloy to ensure the dissolution of Mg-containing phases, homogenize the alloying elements, and create spherical eutectic Si particles. The quench is done to preserve the vacancy and solute concentration, while aging is done to form a uniform distribution of particles, increasing the strength [413, 414]. The microstructure of as-built AlSi10Mg can consist of a columnar or a rod-like dendritic structure, depending on the heat input of the energy source [71]. However, the cellular structures can vary in coarseness depending on the spatial thermal history of the sample [72]. LMD of AlSi10Mg has also shown to form cellular and divergent dendrites [68, 249]. The dendrites are α-Al, while the interdendritic regions are eutectic Si [68, 69, 71]. There are also cases of Mg Si precipitates in the interdendritic regions [69, 71, 72, 415]. Laser-based techniques have shown to result in deposits that are not fully dense [68, 69, 74] and heat treatment has not been shown to alleviate this issue [71]. Heat treatment of AlSi10Mg has shown to decrease the solubility of Si in the primary α dendrites, suggesting that Si is rejected from the primary dendrites to form Mg Si [68]. This is due to the supersaturation of α-Al resulting from the rapid solidification and undercooling of Mg Si. Increasing the solution temperature or time tends to coarsen the Si particles while also decreasing the number of particles due to particle coalescence and Oswald ripening [68, 416]. Lv et al. studied heat treatment of LMD AlSi10Mg and found that the tensile properties increase with a T6 heat treatment[68]. The formation of a fine cellular α-Al dendrites supersaturated with Si, and the localization of Mg at the grain boundaries of the as-built samples, increase the hardness of the primary dendrites. However, the T6 heat treatment allows for the diffusion of Si, which causes a slight reduction in hardness [68]. Most publications on DED AM of AlSi10Mg use L-DED; thus, more investigation using other DED methods with AlSi10Mg needs to be done. Furthermore, additional work is required to determine the optimal heat treatment protocol for AlSi10Mg using all types of DED technologies.

6.5. Co–Cr alloys

This section will discuss the heat treatment of AM deposited Co–Cr alloys. The focus will be on Stellite 21 and Stellite 6 Co–Cr–Mo alloys, as they are the most studied. An outline of the standard heat treatment will be reviewed, followed by the as-built and post heat treatment microstructure, and the corresponding effect on mechanical properties.

For cast Co–Cr–Mo alloys complying with ASTM F75, no standard heat treatment is included [417]. The as-cast condition of this composition typically consists of FCC γ Co, a σ intermetallic, and interdendritic carbides [418]. The goal of heat treatment is to homogenize the microstructure, remove cast defects and improve mechanical properties through precipitation dissolution [419–421]. The most common treatment consists of a solution treatment at temperatures of approximately 1200 C for a range of times from 1 to 4 h [420, 422, 423]. However, for pin-on-disk and hardness tests, increased performance has been shown when aging is included in the heat treatment [424, 425]. Laser deposited Stellite 21 typically consists of a range between columnar and equiaxed dendrites, with a σ intermetallic, and in the interdendritic regions [88]. WAAM of Stellite 6 results in a dendritic structure with Co-rich FCC γ primary dendrites, and eutectic γ-Co with carbides [145]. Porosity has also been found in laser deposited Co–Cr–Mo deposits [426]. For LENS of Co–Cr–Mo, increasing the solution treatment time leads to decreases in size and amount of carbides due to improved kinetics for carbide dissolution at higher temperatures. This is opposite to increasing the aging time, which increases the precipitation concentration, due to decreases in solid solubility of carbide forming elements at higher temperatures [86]. Depending on the particular application of the part, the resulting microstructure from the heat treatment will yield different performance results. Longer solutionizing times with no aging may increase corrosion resistance due to high Cr contents in the matrix [425]. Variations in temperature and hold times will result in different carbide sizes, morphologies, and distributions, which will result in a range of properties [86, 427].

6.6. Mg alloys

The use of Mg alloys for AM has not been explored in as much depth as the other alloys presented in this work. Thus no work has been published on the effects of heat treatment on the microstructure and corresponding properties of Mg deposits.

6.7. Copper alloys

The majority of the work on Cu alloys for DED technologies has been on WAAM of nickel aluminium bronze. Shen et al.deposited nickel aluminium bronze using WAAM and found that the as-built microstructure mainly consisted of Widmanstätten α phase and martensitic β phase [126, 127, 428]. Dharmendra et al. found no retained β, but instead found precipitates in the interdendritic regions [429, 430]. Homogenization at 900 C and quenching transformed the microstructure to equiaxed and columnar α with some retained β' and κ phases. This tends to decrease the strength and hardness but increases the elongation to failure, which is attributed to the absorption of some of the previously formed κ phases [430]. With post quench tempering, it was found that the equiaxed grains disintegrated to columnar grains, while the already formed columnar grains coarsened. Increasing the tempering temperature resulted in the elimination of the retained β and κ lamellae, and particles begin to form [126, 127]. This causes the mechanical properties to increase closer to the as-built condition. However, the distribution of κ phases becomes more uniform, decreasing the spatial variation in properties [430]. However, further increasing the tempering temperature results in significant coarsening of the κ phase, causing a decrease in performance [127].

6.8. Tungsten alloys

As mentioned previously, some work has been done on DED of pure W [431], W-Ni alloys [90], W-Fe alloys [432] and tungsten heavy alloys [89]. However, no work has been conducted on how post-processing affects the microstructure and the corresponding properties.

7.1. Process planning

Current path planning methods are generally limited to 2.5 DOF, with few systems available for 3–5 DOF path planning. 2.5 DOF systems are inherently inefficient due to support structures, requiring post-processing as well as design limitations. For large-scale parts, this entails additional manufacturing costs (such as labour and delivery time). 5 DOF path planning overcomes these challenges to a large extent but has limited industrial integration. A number of algorithms have been reviewed in this paper. While the algorithms are fundamentally suitable for metal AM, work is still required on non-planar tool path planning for metal AM where the generated tool path must satisfy the requirements identified in Section 4.2. Adaptive slicing offers an advantage in terms of reducing both the layer height and variation in material properties. This necessitates a fundamental understanding of bead deposition geometry, microstructure and solidification modelling. Combining this knowledge with adaptive slicing will allow efficient manufacturing of high-quality parts, but this requires a significant multidisciplinary effort in material science and mechanical and manufacturing engineering. Path planning, which is a function of part geometry, directly affects heat transfer and conduction through the part being made. This results in varying amounts of additional heat in the part at any given time and build location, resulting in varying solidification rates, thereby affecting the geometry of the build and the resultant microstructure. Therefore, it is necessary to include heat transfer modelling at an earlier stage concurrent to the path planning. Current models suffer from long simulation times, inherent assumptions to reduce computational time, and a limited set of manufacturing systems and material system availability which need to be improved through further research.

Incorporation of multi-degree of freedom path planning along with considerations of the aspects mentioned above will enable in-situ modification of material metallurgy and its mechanical and geometric properties during deposition. This will truly unleash the potential freedom of design and complexity that AM processes have to offer.

Prior to fabrication, it is also necessary in many cases to calibrate the workpiece with the fabrication platform. This is especially important when the build requires coordinated motion between the workpiece and deposition system – as is always the case during multi-directional deposition. Workpiece calibration can be automated by using a 3D or line scanner mounted on the deposition head.

7.2. Deposition technologies

An area where considerable research potential can be found in the powder delivery during multi-axis DED (e.g. using an 8 DOF robotic LDED platform). Currently, the LDED deposition head must always remain vertical and thus align with the gravity vector to provide ideal powder delivery. Developing methods to loosen these constraints on the deposition head orientation is necessary to utilize the full potential of an 8 DOF robotic LDED platform. Several challenges need to be overcome to enable this, including, but not limited to, modelling of powder flow at different angles to the build surface and the effect of shielding gas dispersion in the build area at non-vertical angles.

While in contrast to LDED, deposition at varying orientations is intrinsically possible with fewer limitations using GMAW-based deposition technologies. However, they are at a stage of lower maturity regarding monitoring the melt pool temperature and geometry and energy input. Sensing the melt pool in a GMAW-based deposition system is also challenging due to the rapidly and drastically changing lighting conditions due to the presence of the arc.

Work is also ongoing on the minimization of the energy input during GMAW-based deposition, where CMT technology plays a significant role. Owing to the highly controlled CMT process where it is possible to fine-tune the deposition process, significant potential for the optimization and adaptation to particular material considerations is possible. For example, in recent years, Fronius International GmbH has been developing custom synergic lines to further reduce the heat input during WAAM using CMT technology [L. Hudson and M. Zablocki, personal email and oral communications, March 2020]. Further potential for advanced research on optimizing the deposition process exists and should be considered. This necessitates in-situ and high-speed sensing of the welding current and voltage, providing important insight into the energy input into the build during fabrication. It can also provide valuable insight into the process, and the measurements themselves can be used as feedback for temperature control systems. Moreover, tremendous potential for robust sensor-fusion-based technologies exists.

7.3. In-situ monitoring, modelling and control

The control algorithms reviewed in Section 5.1 for bead geometry control are relatively basic and have only been developed for and tested with single-track walls. Significant research is required to advance process monitoring and control towards the objective of robust, adaptive and intelligent control methods that provide a sufficient degree of autonomy and robustness to unanticipated conditions during fabrication. Moreover, bead profile sensing and feature extraction have only been done for simple beads. The sensing and feature extraction capabilities need to be expanded and combined with modelling to provide accurate predictions of single beads and overlapping regions of multi-track deposition.

Substantial research potential is also apparent for advancing the area of layer geometry sensing and tool path re-planning during fabrication. The fact that during fabrication, a component is built layer by layer provides a unique insight into the current state of the build through the methods reviewed in Section 5.2. Impending catastrophic build failures can be detected, and the tool path for the following layer can be re-planned to mitigate and correct potential build failures.

Most work on temperature monitoring, and control has been done for LDED, as is apparent from Section 5.3. Particularly melt pool temperature sensing needs substantial work for arc-based deposition technologies. Heat accumulation is coupled with the deposition system travel speed and the material feed rate, which influence the bead and layer geometry. This means that if the bead geometry is adjusted (which is necessary), the heat input changes, which can modify the material composition.

Similarly, as for the layer geometry monitoring, IR cameras can also monitor the overall surface temperature of the component during the build to adjust dwell times and cooling rate of the substrate plate. This is especially important for maintaining consistent metallurgical properties.

7.4. Materials

Many challenges still need to be addressed in regards to materials for AM. One of the more apparent areas of exploration is expanding the number of materials available in AM. This is clearly highlighted by the chart presented in Figure 22 [433]. Although many of the materials used in conventional manufacturing are ill-suited for AM, there are still important contributions that can be made through failed experimentation. Increasing the amount of data on what materials may or may not be compatible with AM, allows for significant deductions to be made on the essential material properties a material must have to be used in AM.OPEN IN VIEWER

Another future avenue of interest is using AM to achieve manufacturing feats that are outside the realm of possibility with conventional manufacturing. Although metals toughness far exceeds any other type of material, this comes with a poor strength to weight ratio. However, with AM, the internal structure can be altered to a lattice type, drastically increasing the strength to weight ratio. Additionally, polymer-based AM techniques could be used to fabricate these structures, which can be converted to a mould, and then cast, known as hybrid investment casting. This allows for the use of well-understood material systems in the way they were originally designed.

One of the challenges that is starting to be addressed is the current material selection for AM [434]. The current materials landscape for AM is dominated by materials that were successful with conventional manufacturing techniques. These materials were not designed for the complex thermal cycling inherent to AM, which result in material defects and anisotropic microstructures. Thus the development of new materials created explicitly for AM could allow for more control over phase transformations, elemental segregation and the resulting microstructure. This is especially crucial for large components, where heat treatment procedures incur a financial cost that make them unfeasible compared to conventional manufacturing. Furthermore, microstructural control will allow for predictions in how the part will perform when in service, which is imperative for on-site fabrication. Some promising alternative methodologies are being explored to prevent the epitaxial growth of large columnar grains. The addition of inoculants to aid in the nucleation of equiaxed grains would eliminate anisotropic mechanical properties prevalent in a lot of AM deposits [435]. The addition of boron to Ti64 has been shown to form TiB, which allows for the nucleation of α-grains, resulting in an isotropic grain structure [436]. A similar phenomenon has been reported with the addition of Ti to 5356 Al [253]. Furthermore, the addition of carbon to Ti64's hypoeutectic composition decreases the solidification temperature, causing grain growth restriction through constitutional supercooling. Although a different mechanism, a similar isotropic grain structure occurred [437]. These studies highlight that the development of materials better suited for AM is going to involve understanding the fundamental material paradigms involved in grain growth and solidification, and how these can be used to manipulate the thermodynamics of the system, to mitigate some of the microstructural challenges that researchers are currently faced with. Large amounts of data can be compiled by completing the aforementioned experiments on increasing the materials being trialed for AM, trying completely new material compositions, and in-situ grain control, which can then be used as input for artificial neural networks, to synthesize new materials specifically for AM. This would also incorporate all the data from the published process planning and monitoring and control strategies, allowing the network to develop the appropriate deposition strategy for the new material. The seed to the network would be a material of known composition, as-built microstructure, and mechanical properties. The network would have the ability to simulate the deposition of the material and then predict its as-built microstructure and mechanical properties. The model would employ reinforcement learning strategies to iterate over various compositions of the material, based on the data acquired from the research, to optimize the microstructure and mechanical properties based on the part's specifications. This could completely reinvent how material selection is done and produce materials specifically tailored to the additive manufacturing of that specific part.

The materials available for large-scale robotic AM are currently limited to current alloys in either the welding or coating processes. Material development for various processes used in AM is in its infancy and will yield significant opportunities as the processes mature.

The complex thermal cycling of AM leads to microstructures that are not found in conventional casting and forging operations. Using design guidelines of traditional heat treatment protocols can result in poorer mechanical properties in some materials. This is attributed to the varying degrees of segregation or the novel grain structure that occurs during deposition [438]. Furthermore, many post-processing operations rely on HIPing to reduce internal porosity. This is problematic for large-scale AM due to the inherent cost of this procedure and the size of the processing chamber needed to contain the large part. Thus developing techniques to reduce porosity in situ will be an essential future contribution to AM. Furthermore, the poor geometric tolerance obtained from parts manufactured using particular metal DED systems will need to be improved to reduce the manufacturing costs. This problem currently necessitates hybrid manufacturing systems or some combination of additive and subtractive technologies. This requires developing frameworks that unify positioning, referencing, and planning software to negate the need to detach the part from the build plate and any post-processing. The framework would also need to include localized heat treatment and a means to control the whole part's thermal cycle to ensure the promised mechanical performance. Thus it is speculated that the next generation of large-scale AM systems will appear more similar to traditional manufacturing approaches than powder bed fusion systems. There will be some modularity, where the part will be fabricated and machined in one module and then transferred automatically to a separate heat treatment module, similar to what is seen in traditional manufacturing. It is clear that an integrated automation system will increase productivity for this type of manufacturing. However, this manufacturing system would offer the geometrical freedom and the multi-meter scalability that both traditional and PBF are unable to provide.

The challenge remains to identify the raw materials, process conditions, and process control to maximize product quality using the AM processes and minimize subsequent post-processing requirements. The novel solutions will only be met through multidisciplinary and cross-functional teams closely collaborating. For example, this paper could not have been written without the close collaboration between mechanical, process control, mechatronics, electrical and materials engineers. The future young engineers trained in AM will require a holistic knowledge base and the ability to work cooperatively with other disciplines in engineering, sciences, design and visual arts. This paper has intentionally not addressed the redesign of components from both an engineering or artistic design approach. However, the possibilities using AM technologies will reveal new opportunities that are currently not imaginable.