Using total specific cost indices to compare the cost performance of additive manufacturing for the medical devices domain

Abstract

The label additive manufacturing, also known as three-dimensional printing, serves as an umbrella term for a number of technologies designed to deposit product geometries directly from build materials and digital design information. However, as a relatively recent addition to the spectrum of manufacturing processes, the relationship between process type, system characteristics and cost performance is still broadly unclear for several technology types. To address this gap, the current research develops comprehensive and robust additive manufacturing cost models for two less-studied polymeric additive manufacturing technology variants, material jetting and mask projection stereolithography. Despite sharing the fundamental principle of photopolymerization, the operating processes of both systems are markedly different. This is reflected in the constructed cost models, which incorporate process maps to capture ancillary process elements, ensure efficient capacity utilisation through optimised build volume packing and approximate the expected cost impact of build failure. On this basis, this article estimates a set of specific cost indices reflecting the overall total cost performance of the investigated systems in an example application from the medical devices domain. Specific cost results range from £2.01 to £1.19/cm³ deposited on the Objet Connex 260 system and from £1.59 to £1.00/cm³ of material deposited on the Perfactory system. These results are discussed in the context of similar cost indices extracted from the empirical engineering literature. This article shows that next to increases in build speed, improvements in overall process automation and process stability are needed to enhance the commercial proposition of the investigated technology variants.

Keywords

Additive manufacturing cost estimation cost indices build failure cost model build volume packing material jetting vat photopolymerization

Introduction

Additive manufacturing (AM) technology, also known as 3D printing, is considered by manufacturing professionals and technology observers as an opportunity to transition to digital manufacturing, increase responsiveness, adopt innovative supply chains and realise novel product designs. However, evaluating the technology for commercial deployment in manufacturing settings requires a detailed and realistic view of the full process economics associated with the technology, which is still lacking.

Forming an umbrella term for different technology variants,¹ AM builds up products by incrementally adding material without the use of dedicated physical tooling. As discussed by Tuck et al.,² this mode of operation affords two generic advantages over conventional manufacturing. First, AM reduces the importance of manufacturability and cost constraints imposed by tooling, formalised in conventional methodologies such as Design for Manufacturability and Assembly.³ Second, AM permits the cost-efficient manufacture of products at low volumes or single units. This frees technology adopters from some pressure to create large volumes of identical products, at least on the manufacturing process level.⁴ Adopters are thus able to manufacture customised or highly differentiated products.^5,6

Despite a number of generic process limitations,⁷ AM technology diffusion is currently progressing in a broad spectrum of manufacturing applications.⁸ In the medical sector in particular, the advantages of AM are routinely employed to manufacture complex and patient-specific devices with bespoke geometry. Examples for successful medical applications are hearing aids,^9,10 surgical implants and aids,^11,12 dental products¹³ and drug delivery devices.¹⁴

The value proposition of a manufacturing technology in a commercial setting is defined by the relationship of the manufacturing performance, or its utility in creating products, and the overall cost incurred through the technology. Where manufacturing technologies are developed for new applications, as is the case for AM, cost–benefit frameworks can be employed.¹⁵ Supporting such investigations, the costs of using AM technologies can be assessed through activity-based costing models, which allocate cost elements to particular activities.¹⁶

The seminal AM cost model was proposed by Alexander et al.¹⁷ for the two technology variants filament extrusion and stereolithography. This model is driven by build time estimates and considers the cost of surrounding pre- and post-processing steps. Building on this approach, a number of extensions have been made as follows:

Hopkinson and Dickens¹⁸ investigated the cost of three major AM technology variants (stereolithography, filament extrusion and laser sintering) for the production of identical components in large numbers. By doing so, their cost model treated AM at full capacity utilisation rather than building individual parts in isolation.

Ruffo et al.⁷ studied the relationship between unit cost and the degree of capacity utilisation in laser sintering, showing that unfilled machine capacity has an adverse effect on unit cost. This result was extended by Ruffo and Hague,¹⁹ demonstrating that mixing dissimilar components in builds can lead to higher utilisation and hence lower cost.

Atzeni and Salmi²⁰ and Atzeni et al.²¹ incorporated product redesign into inter-process cost comparisons between AM and tooled conventional processes for both polymeric and metallic products. The authors thereby addressed the interdependence of process and design choice, which forms an aspect that may limit the validity of process cost comparisons relying purely on breakeven costs.

Identifying the need to specify well-configured build volumes to claim technical efficiency in analyses of AM process cost, Baumers et al.²² described a capacity utilisation problem in AM, which was resolved by integrating a computational build volume packing approach with AM cost modelling.

Rickenbacher et al.²³ proposed a cost model apportioning costs incurred per layer to individual parts, thereby increasing the accuracy of cost estimation for processes requiring the deposition of geometry on build plates or substrates.

Baumers and Holweg²⁴ experimentally assessed risk of build failure in a study of polymeric laser sintering, showing that ill-structured aspects²⁵ can severely affect expected unit cost. This research complemented other previous work investigating the cost of quality control systems in AM.^26,27

On the basis of this body of literature, this article constructs a comprehensive cost model of the detailed analysis type²⁸ for two AM technology variants that carry great promise for the medical sector, material jetting and mask projection stereolithography. Both technology variants have yet to be the subject of extensive investigations of manufacturing cost. To reach a total cost perspective, this article joins up a range of relevant aspects with the aim of generating a comprehensive and robust cost model for the investigated systems.

The investigated AM technology variants share the core principle of exposing a photoreactive monomer layer-by-layer to build up product geometries. In terms of the operating process, however, the studied systems are highly dissimilar. While the material jetting system selectively deposits a viscous monomer resin using a piezoelectric jetting head and exposes the material through an attached ultraviolet (UV) lamp, the mask projection stereolithography system selectively exposes an entire layer of resin held in a vat through a micro mirror device, known as a digital light processing (DLP) device. A short overview of the two identified AM technology variants is provided in Table 1; a more detailed introduction to both technologies is presented by Gibson et al.³¹

Table 1.

Characteristics of the investigated AM systems.

Technology variant	Material jetting	Mask projection stereolithography
System type	Stratasys Objet Connex 260	Envisiontec Perfactory 3 Mini SXGA
Operating principle	Droplet deposition of photopolymers using a jetting head	Resin vat/photopolymer, area exposure via UV light and a DLP chip
Build direction	Bottom-up	Upside-down
Supports	Yes	Yes
Multi-material capability	Yes	No
Layer thickness range	16 μm minimum, 30 μm typical	15–150 μm
Used layer thickness setting	30 μm	50 μm
X/Y build area (nominal)	255/252 mm	Max. 84/63 mm, depending on lens
X/Y build area (adapted)	As above	190/120 mm
Max. Z-height	200 mm	230 mm
X/Y resolution	600 dpi	2800/2100 pixels
Material handling	Deposited on demand	Gravity fed
Build material(s)	VeroClear RGD810, TangoBlack FLX973	R11
Manufacturer reference	Stratasys²⁹	Envisiontec³⁰

UV: ultraviolet; DLP: digital light processing.

In manufacturing applications, the ability of a process to achieve acceptable surface quality is important. Table 1 shows that the layer thickness setting on both systems, as a factor influencing surface quality, is similar, at 30 μm on the Objet Connex system and 50 μm on the Perfactory system. Surface roughness data available in the literature for material jetting indicate a mean R_a value of approximately 10.2 μm,³² whereas available surface roughness data for the Perfactory system indicate a better surface finish, with an R_a value of 0.6 μm in the vertical orientation.³³ For the purposes of this research, however, we ignore the effect of surface finish on manufacturing cost.

While the analysed Objet Connex material jetting system has a suitable configuration for an investigation of current AM technology application in the medical devices domain, it was determined that the available build space on the investigated Envisiontec Perfactory 3 Mini SXGA was insufficient. Hence, this research substitutes the nominal X/Y build area characteristic (as stated in Table 1) with the useful build volume area of the closely related Perfactory 4 Standard XL system. This adaptation appears permissible as the Perfactory system is an area exposure system and therefore does not exhibit a relationship between per-layer processing time and layer content, as it is capable of processing entire layers in a single step.³¹

This article is structured as follows: section ‘Methodology’ presents in detail the methodology used to construct the total cost estimators and the calculation of specific cost indices. Section ‘Results’ presents the results of the proposed cost models. Next, section ‘Discussion’ critically evaluates the obtained results in the context of a body of corresponding indices derived from the literature. Conclusions are drawn and recommendations for further research are made in section ‘Conclusion’.

Methodology

Process mapping

As discussed by Alexander et al.,¹⁷ AM processes do not take place in isolation but in the context of surrounding process steps. Baumers and Holweg²⁴ have demonstrated that process maps can be used to capture relevant process steps, from the handling of digital design files in the *.stl format to the final post-processing of the resulting parts. Showing that such representations are specific to individual AM technology variants, Figure 1 depicts process maps for the Objet Connex and the Perfactory processes.

Figure 1.

Process maps for material jetting (a) and mask projection stereolithography (b).

For both investigated AM processes, the mean time durations have been recorded for each process element in a series of build experiments. As the possibility of build failure may impact the successful outcome of a build, AM process maps can additionally be used to define at which point in the work flow the consequences of failure events arise. In Figure 1, this point is demarcated by placing a failure node, determining which process elements must be repeated if failure occurs.

Table 2 summarises this information and introduces a notation that distinguishes between steps that are affected by the risk of build failure R_i (i.e. by preceding it in the work flow) and those unaffected by this risk S_j. Furthermore, the table identifies whether the costs associated with each step arise for the build as a whole or must be allocated to each part contained within the build.

Table 2.

Process map for the assessed AM systems.

Notation	Material jetting (Objet Connex)			Mask projection stereolithography (Perfactory)
Notation	Step	Cost element	Duration (s)	Step	Cost element	Duration (s)
R ₁	File preparation and build setup	Labour cost, per build	480	File preparation, build setup, data loading	Labour cost, per build	120
R ₂	Nozzle cleaning procedure	Labour cost, per build	240	Calibration and compensation mask	Labour cost, per build	1620
R ₃	Build supervision	Labour cost, per build	615	Build supervision	Labour cost, per build	615
R ₄	Manual platform separation	Labour cost, per part	18	Manual platform separation	Labour cost, per part	45
Effect of possible build failure at this point
S ₁	Manual support removal (rough)	Labour cost, per part	240	Manual support removal	Labour cost, per part	120
S ₂	Water jet support removal	Labour cost, per part	360	Ultrasonic bath in isopropanol	Labour cost, per build	900
S ₃	Compressed air drying	Labour cost, per part	60	Compressed air drying	Labour cost, per part	30
S ₄	Removal of residue and initial post-processing	Labour cost, per part	780	UV exposure	Labour cost, per build	400

The general structure of the process maps shown in Figure 1 has been specified to attribute the risk of build failure to steps within the AM process. The build failure node has been placed following part removal to reflect the fact that the deposited parts will have to be removed from the build platform, irrespective of failure. Thus, the costs incurred by manual platform separation (R₄) are incurred in any case. Moreover, this specification, which has been chosen for simplicity, implies that part destruction or rejection outside of the build process is not reflected in the total cost model.

Risk of build failure

The idea guiding the integration of build failure within this model is that it does not occur with certainty. Instead, it is assumed that failure takes place with a constant probability p_Constant during each layer deposition, where p_Constant ∈ [0, 1]. By employing a probability tree to reflect the deposition of n layers, the overall probability of successfully completing the build can be established, as shown in Figure 2.

Figure 2.

Probability tree for successful build outcome based on p_Constant.

Given the constant probability of build failure per layer, the probability of successful build completion, which is also the complement of the sum of all probabilities of build failure per layer, can thus be expressed as following probability function P(n)

P (n) = 1 - \sum_{i = 1}^{n} {(1 - p_{Constant})}^{n - 1} \cdot p_{Constant} = {(1 - p_{Constant})}^{n}

(1)

To obtain the data required for the estimation of p_Constant, eight separate build experiments were performed on the Objet Connex system. The number of successfully deposited layers was 6920 with one build failure event, which was due to machine stoppage caused by control system error (the machine failed to resume build activity after pausing the build operation). This error was unrelated to the build configuration and machine setup. On the Perfactory system, 11 builds were observed, resulting in a number of cumulatively deposited layers of 4963 and two build failures, one resulting from detachment from the build platform and one resulting from part deformation. By dividing the number of observed build failure events by the total number of depositable layers, p_Constant was approximated for the Objet Connex system at 0.0145% and for the Perfactory system at 0.0403%. Figure 3 graphically plots both build failure models P(n), given the estimation of p_Constant for both systems, and also indicates the maximum build height of a single build.

Figure 3.

Comparison of the simple build failure models.

It is acknowledged that the number of build experiments used to calculate p_Constant is small, with 8 and 11 observations. Moreover, empirical data on build failure modes and probabilities are rare in the engineering literature. Baumers and Holweg²⁴ report on a similar study for laser sintering, based on 19 experiments, resulting in a mean number of depositable layers of 4041, which can be expressed as a p_Constant value of 0.0247%. Sitting between the probabilities obtained in this research, this serves as a preliminary indication of robustness.

Constructing the total cost model

Providing the parameters required for the cost models presented in this article, Table 3 summarises available data drawn from the literature and other sources. Adapting the general structure proposed by Alexander et al.,¹⁷ this information flows into build time estimation, process energy consumption estimation, and the specification of indirect overheads, technician labour costs, and machine-specific direct and indirect costs. The purchase cost of the AM machine enters through the total machine indirect cost rate. This cost rate is obtained by assuming straight line depreciation with a period of 8 years,⁷ which is considerably longer than for computers and electronic equipment. Furthermore, an annual operating time of 5000 h is assumed, as well as an estimated annual cost for maintenance and consumables.

Table 3.

Shared and machine-specific cost elements.

Model element	Parameter	Objet Connex 260		Perfactory 3 Mini SXGA		Data source
Model element	Parameter	Value	Unit	Value	Unit	Data source
Build time estimator	Machine start-up and shut down (T_Job)	254	s	60	s	Measurement using timer
	Fixed time increment per layer, includes fixed increment for Y movement and simultaneous platform indexing, α₁	0.8936	s	N/A		Baumers et al.³⁴
	X pass print head movement time (X dispersion related), β₁	0.0018	s
	Fixed time increment per X pass, β₂	3.0005	s
	Y dispersion related print head Y movement time, β₃	2.0010	s
	Fixed time increment per layer, α₂	N/A		30.58	s	Measurement using timer
	Layer thickness, lt	30	μm	50	μm	System setting
Energy consumption estimator	Fixed energy consumption per build, E_Job	0.10	MJ	0.42	MJ	Measurement using Yokogawa CW240 digital power metre³⁴
Energy consumption estimator	Energy consumption rate, Ė_Process	533.1	J/s	240.01	J/s	Measurement using Yokogawa CW240 digital power metre³⁴
Shared indirect cost elements	Rent, building area cost rate	4.53		£/h		Ruffo et al.⁷
	Hardware, software and consumables per year	1781.62		£
	Total administration overhead cost rate	0.36		£/h
	Total non-machine indirect cost rate	4.89		£/h
Labour costs	Technician full annual cost	32,420		£		University of Nottingham pay scales
	Working days net of holidays	228		days
	Total hours worked per year	1653		h
	Production labour cost rate (C_Labour)	19.61		£/h
Machine-specific indirect costs	Machine purchase	115,810.40	£	54,400.00	£	Manufacturer correspondence^29,30
	Depreciation period	8		years		Ruffo et al.⁷
	Annual operating time	5000		h		Ruffo et al.⁷
	Estimated maintenance and consumables	3044.00	£/year	3291.20	£/year	Estimate based on Baumers et al.³⁴
	Total machine indirect cost rate	3.50	£/h	2.02	£/h	–
Machine-specific direct costs	Energy cost	0.02		£/MJ		Baumers et al.²²
	First raw material price	281.33	£/kg	227.13	£/kg	Vendor quotation
	Second raw material price	280.99	£/kg	N/A		Vendor quotation
	Employed material price, per volume unit	0.33	£/cm³	0.28	£/cm³	–
	Mean material density, as deposited	1.165	g/cm³	1.215	g/cm³	Manufacturer reference
	Material waste factor	1.76	–	1.2	–	Baumers et al.³⁵ and experimental measurement
	Modified material waste factor, ω	1.936	–	1.2	–	Baumers et al.³⁵ and experimental measurement

As shown by Alexander et al.,¹⁷ the appropriate model of build time for an AM system depends on its operating principle. For the investigated Objet Connex system, this analysis employs an existing model,³⁴ which follows the intuition that the material jetting is characterised by discrete print head movements, or ‘passes’ in the available build space. The required number of passes of the print head in the X dimension can be obtained using the least integer function f(y), effectively splitting up the build volume into y* discrete slices of build space oriented parallel to the X dimension (where y^* ∈ ℕ)

f (y) = y^{*} if (y^{*} - 1) w < y \leq y^{*} w

(2)

For simplicity, this article ignores print head offsetting and interlacing and assumes a fixed width w for each print head pass in the X direction of 50 mm. Thus, the time needed for print head movement in this dimension is seen as a function of dispersion of part geometry in the Y dimension.

The duration of passes in the X dimension is again obtained by the dispersion of geometry within the layer. Movement in the Z dimension is simply the time needed to index the build platform vertically. Thus, overall build time T_Build is modelled as a sequence of discrete print head movements in the X, Y and Z dimensions, together with a fixed time increment for machine start up and cool down T_Job. The build time model used for the Objet Connex employs four parameters α₁, β₁, β₂ and β₃, as summarised in Table 3. As described by Baumers et al.,³⁴ the parameters needed for the print head movement time estimator were estimated via two ordinary least squares regressions, each yielding a slope term (β₁, β₃) and an intercept term (α₁, β₂). The R² values accompanying both ordinary least squares regressions indicate a high degree of fit (0.99 and 0.97). The used build time model for the material jetting system can be expressed as

\begin{matrix} T_{Build, Objet Connex} = \\ T_{Job} + \sum_{i = 1}^{Z} (α_{1} + \sum_{j = 1}^{Y^{*} i} (β_{1} x_{ji} + β_{2}) + β_{3} {y^{*}}_{i}) \end{matrix}

(3)

As discussed by Baumers et al.,³⁴ it should be noted that the employed approach implicitly assumes that the investigated relationships governing build time are linear; the authors also present a validation of this build time model, reporting an estimation error of 5.96% for this system at full capacity, which is deemed sufficient.

While a reasonably accurate and validated build time model was available for the Objet Connex system, no such model was available for the Perfactory system. Since its operating principle is based on full area exposure, we use a simple build time estimator treating the time needed for each layer exposure as fixed, as discussed in early literature on AM build time estimation.³⁶ As indicated by Gibson et al.,³¹ it is thus possible to treat layer deposition time α₂ as constant, such that

T_{Build, Perfactory} = T_{Job} + α_{2} l

(4)

The single parameter α₂ was obtained through a dedicated build experiment. Its robustness was demonstrated in two validation builds, resulting in errors of 6.46% and 8.50%. In the context of the AM build time estimation literature^{7,22,28,37–39} which reports errors ranging from 1.03%²² to 22.68%,³⁹ this appears acceptable.

The next step is the formulation of an energy consumption estimator E_Build. A simple specification was sought by splitting overall energy consumption into a fixed increment E_Job incurred during machine start up and cool down, and a time-dependent element, obtained by multiplying a measured energy consumption rate Ė_Process by build time. Furthermore, it is assumed that the investigated systems do not consume energy when idle. Thus, build energy consumption can be expressed as

E_{Build} = E_{Job} + ({\overset{\cdot}{E}}_{Process} \times T_{Build})

(5)

With the build time and energy consumption models in place, the remaining step is the specification of the cost estimator. As an activity-based costing model,¹⁶ the approach allocates a rate of indirect costs Ċ_Indirect, which includes non-machine and machine costs, on the basis of the build time estimate T_Build. Differing from the seminal model by Ruffo et al.,⁷ this research removes labour cost from this indirect cost rate and instead adds a fixed labour cost associated with initial machine supervision (R₃ in Table 2).

For the process model, it is of relevance that the Objet Connex system deposits two build materials whereas the Perfactory system is a single material approach (like all resin vat systems). To avoid the additional complexity of a multi-material case, a mean material price m was estimated for the Objet system (£0.33/cm³) and allocated on the basis of the total deposited part volume v, excluding supports. Energy costs are determined by a commercial electricity price ep, £0.02/MJ.

To reflect the cost of material losses due to handling and degradation, the overall raw material expense is modified by a waste factor ω. For the Objet Connex system, a waste factor was obtained from the literature.³⁵ However, as this system requires additional support structures, which the Perfactory system does not, this waste factor was modified using the mean part volume/additional support volume ratio exhibited by the test geometries. The waste factor for the Perfactory system was obtained by measuring the mean difference between the mass of the products and the overall material consumed during each build experiment. Thus, the cost model for the core build activity C_Build on both AM systems can be specified

C_{Build} = ({\overset{\cdot}{C}}_{Indirect} \times T_{Build}) + ω (v \cdot m) + (E_{Build} \times ep)

(6)

The subsequent step is to assign the labour costs associated with various process steps to arrive at the total cost model. As is evident from the process maps shown in Figure 1, some of these steps are subject to the risk of build failure, denoted in Table 2 by R_i. To form our expected cost terms for the affected elements of the AM process chain, we attach the risk of build failure as the inverse of equation (1). Other elements are unaffected by this risk and are denoted as S_i. Moreover, as described in Table 2, some costs arise per part, whereas some arise only once per build containing q units. Since these characteristics differ, two alternative total cost models C_Total were required for the investigated systems. For the Objet Connex system, where all post-build processes S_i occur on the individual part level, the total cost model is thus expressed as

\begin{matrix} C_{Total, Objet Connex} = ({[\begin{matrix} 1 \\ 1 \\ 1 \\ q \end{matrix}]}^{T} \cdot [\begin{matrix} {\overset{\cdot}{C}}_{Labour} R_{1} \\ {\overset{\cdot}{C}}_{Labour} R_{2} \\ {\overset{\cdot}{C}}_{Labour} R_{3} \\ {\overset{\cdot}{C}}_{Labour} R_{4} \end{matrix}] + C_{Build}) \frac{1}{P (n)} + q {\overset{\cdot}{C}}_{Labour} \sum_{i = 1}^{4} S_{i} \end{matrix}

(7)

For the Perfactory system, where S₂ and S₄ occur on a per build level, the following specification is used

\begin{matrix} C_{Total, Perfactory} = ({[\begin{matrix} 1 \\ 1 \\ 1 \\ q \end{matrix}]}^{T} \cdot [\begin{matrix} {\overset{\cdot}{C}}_{Labour} R_{1} \\ {\overset{\cdot}{C}}_{Labour} R_{2} \\ {\overset{\cdot}{C}}_{Labour} R_{3} \\ {\overset{\cdot}{C}}_{Labour} R_{4} \end{matrix}] + C_{Build}) \frac{1}{P (n)} + {[\begin{matrix} q \\ 1 \\ q \\ 1 \end{matrix}]}^{T} \cdot [\begin{matrix} {\overset{\cdot}{C}}_{Labour} S_{1} \\ {\overset{\cdot}{C}}_{Labour} S_{2} \\ {\overset{\cdot}{C}}_{Labour} S_{3} \\ {\overset{\cdot}{C}}_{Labour} S_{4} \end{matrix}] \end{matrix}

(8)

Having specified the total cost per build subject to the cost of build failure C_Total, the cost metric of interest in this research, C_Specific, can be obtained by dividing C_Total by the total volume of geometry deposited, v

C_{Specific} = \frac{C_{Total}}{v}

(9)

Implementing the cost model as an integrated tool

To be able to claim efficient technology utilisation, which forms a prerequisite for manufacturing cost models in any case, the next step is to integrate the specified cost model with a computational build volume packing approach, addressing the inherent irregular three-dimensional packing problem.⁴⁰ This aspect is especially pertinent when considering the use of AM in commercial manufacturing settings, such as the medical industry, where cost performance shapes the production performance frontier⁴¹ and thereby forms a determinant of the viability of technology adoption decisions.

To obtain well-specified levels of capacity utilisation, this research employs a modified version of the implementation used by Baumers et al.,³⁴ labelled the Build time, Energy and Cost Estimator (BECE). This software was implemented in the general-purpose programming language C++ using the free integrated development environment Dev-C++ (version 4.9.9.2). The BECE tool operates by manipulating voxel-based representations of build volumes and product geometries, with a voxel resolution of (5 mm)³.

With a particular emphasis on the application of AM in the medical devices domain, the BECE tool contains a number of geometries which are reflective of medical devices in size and/or shape, as shown in Figure 4. This includes horizontally oriented inhaler geometry with cap (Figure 4(a)), a vertically oriented inhaler cap (Figure 4(b)), a packing testing geometry (Figure 4(c)), a surface testing block (Figure 4(d)), an end cap for an extrusion profile (Figure 4(e)) and a delicate internal geometry for a medical device (Figure 4(f)). For an illustration of the test geometries as-manufactured on both systems, Figure 5 shows the end cap; to demonstrate multi-material capability, a logo is embedded into the part built on the Objet Connex system.

Figure 4.

Basket of test parts (a)–(f).

Figure 5.

End caps as-manufactured on (a) Objet Connex and (b) Perfactory.

The BECE tool is designed to populate the available build space in the machine using a computational build volume packing algorithm selecting and inserting the parts shown in Figure 4. Such functionalities have been discussed in the literature.^42–44 The number and type of each part to be inserted is determined by the user; the location and rotation (around the Z-axis in discrete 90° steps) is determined automatically. Following the specification of each build, the BECE tool estimates build time T_Build, process energy consumption E_Build and total cost C_Total on the basis of system attributes and a voxelised representation of the build configuration.

Results

To assess the relationship between the pattern of capacity utilisation and overall process performance in the medical devices context, this article investigates three different build configurations for each investigated system. In an initial execution A, one inhaler component (shown in Figure 4(a)) is inserted into the build volume in isolation. This is followed by execution B, in which the maximum possible number of inhalers is placed into each platform by the BECE tool, resulting in the insertion of six inhalers on the Objet Connex system and two inhalers on the Perfactory system. Execution C reflects machine utilisation at full capacity with a mix of different parts inserted and arranged in the available build space by the BECE tool. On the Object system, four inhalers (Figure 4(a)), eight vertical inhaler caps (Figure 4(b)), one packing test geometry (Figure 4(c)), one surface testing block (Figure 4(d)), two end caps (Figure 4(e)) and seven internal components (Figure 4(f)) were inserted. On the Perfactory system, the tool inserted one inhaler, four inhaler caps, zero packing test geometries, one surface testing block, four end caps and two internal components.

Figure 6 compares the absolute estimates of overall process duration (the makespan), process energy consumption and cost resulting from these executions subject to the risk of build failure. The calculated makespans range from 335 to 1826 min on the Objet Connex system and from 332 to 847 min on the Perfactory system. The analysis shows a total cost range from £90.66 to £615.47 on the Objet Connex system and from £73.49 to £176.99 on the Perfactory system. It should be noted, however, that this difference in absolute cost is not indicative of costs per part as the Objet Connex system exhibits a larger build volume, containing more parts at full capacity utilisation.

Figure 6.

Total build time, energy consumption and costs.

To make statements on relative process performance, it is necessary to break the results presented in Figure 6 down to the volume of geometry deposited. Table 4 compares the investigated systems at different levels of capacity utilisation in terms of process speed (cm³/h), specific energy consumption (MJ/kg) and specific cost (£/cm³).

Table 4.

Summary metrics of process economy.

Metric	Unit	A, Objet, one inhaler	A, Perfactory, one inhaler	B, Objet, full inhaler	B, Perfactory, full inhaler	C, Objet, full utilisation	C, Perfactory, full utilisation
Overall process speed	cm³/h	8.27	8.36	23.44	16.55	10.07	7.25
Specific energy consumption	MJ/kg	178.75	78.69	54.84	39.34	112.33	90.54
Specific cost	£/cm³	1.96	1.59	1.19	1.00	2.01	1.73

The comparison reveals that high utilisation configurations do not necessarily lead to proportionate increases in process efficiency. This aspect can be observed by comparing the single-unit inhaler configuration on both platforms (execution A) to the full build multi-part configuration (execution C). Despite a far greater degree of build utilisation, the Objet Connex system exhibits only a moderate increase in process speed, from 8.27 to 10.07 cm³/h. The Perfactory system even exhibits a decrease in process speed from 8.36 to 7.25 cm³/h. This is largely determined by machine build time and mirrored in specific energy consumption, decreasing from 178.75 to 112.33 MJ/kg on the Objet system and increasing from 78.69 to 90.54 MJ/kg on the Perfactory system. In terms of monetary inputs, the analysis shows that cost performance worsens when switching from execution A to C on both systems, from £1.96/cm³ deposited to £2.01/cm³ deposited on the Objet system and from £1.59/cm³ deposited to £1.73/cm³ deposited on the Perfactory system.

This apparently counterintuitive behaviour originates from the different Z-heights in the single-part and multi-part build configurations, where individual parts protrude into higher regions of Z-height, prolonging the build process and leading to an accumulation of the risk of build failure. To isolate the effects of increased capacity utilisation from effects of increased Z-height, the single-part execution A can be compared with a build containing multiple inhalers, execution B, resulting in a build with the same Z-height. On the Objet Connex system, the BECE tool inserted six units into the available build space. On the Perfactory system, due to a smaller available build volume, the algorithm was able to insert two components.

In this case, switching from the single-part build (execution A) to the full inhaler build (execution B) results in a substantial improvement in all investigated process performance characteristics. The deposition speed observed on the Objet Connex system increases from 8.27 to 23.44 cm³/h. Similarly, the Perfactory system shows an increase in speed from 8.36 to 16.55 cm³/h. Again mirroring this relationship, specific process energy decreases from 178.75 to 54.84 MJ/kg on the Objet system and from 78.69 to 39.34 MJ/kg on the Perfactory system. Cost performance is also dramatically improved on both systems, changing from £1.96/cm³ deposited to £1.19/cm³ deposited on the Objet system and from £1.59/cm³ deposited to £1.00/cm³ deposited on the Perfactory system. The discrepancy between the multi-part full build (execution C) and the configuration with multiple inhalers (execution B) highlights the effect of Z-height on the cost performance on both systems. The cost model proposed by Rickenbacher et al.²³ recognises this issue and resolves it by apportioning costs on the basis of per-layer content, rather than simply on the basis of volume fraction, as done in this research.

For further investigation, the overall cost metric can be decomposed into its constituent elements. This analysis is performed for the multi-part build configuration with mixed parts (execution C) as this setup is deemed to be most reflective of the level of capacity utilisation observed in practise. Figure 7 summarises the full cost metric (amounting to £615.47 on the Objet system and £176.99 on the Perfactory system) separated into the contributions of cost elements. To demonstrate the overall cost impact of build failure, these costs are shown as a separate cost element, which is obtained by subtracting a model where P(n) is set to unity from the total cost models (equations (7) and (8)). Labour costs, indirect costs and direct costs are also shown without the effect of the risk of build failure.

Figure 7.

Total process costs of the multi-part build decomposed (Objet Connex (a) and Perfactory (b)).

As can be observed in Figure 7, the relative importance of the cost elements is dissimilar on both systems. The largest relative difference, 15% versus 29%, occurs in terms of risk-related costs, which have approximately twice the impact on the Perfactory system. This is due to the larger number of build failures observed in the experimental phase of this research – indicating a somewhat lower process stability. This observation must be qualified by stating that the used model was obtained from a very small dataset and should therefore be generalised with caution. The share of labour costs observed on the Objet system is approximately 13% higher than on the Perfactory system, which is due to the Objet Connex process drawing more heavily on labour inputs for post-processing, as evident from Table 2.

Discussion

Unlike AM, the cost structure of conventional tooled manufacturing processes is often shaped by tooling expenses that are spread over production runs.^19,21 This aspect creates static process-level economies of scale arising from indivisibilities.⁴⁵ It is noteworthy that sometimes claims are made that AM does not exhibit economies of scale.^4,9 Since AM processes never employ tooling, cutting implements, moulds or dies, it would be, however, correct to state that only process-based scale economies resulting from the use of dedicated tooling are absent. Other static and dynamic economies of scale can arise if throughput more than proportionately increases where systems are physically scaled⁴⁶ or where there are learning effects. The results of this research indicate that depending on how the processes are configured, efficiency gains may be obtained from increasing throughput.

The process maps developed for the investigated technology variants show that AM processes can involve significant manual inputs and process interventions. The comparison between the Objet Connex and Perfactory systems has revealed that material jetting relies to a greater degree on manual labour inputs, which may indicate that the Perfactory system lends itself more readily to complementary automation in post-processing. It should be noted, however, that the approach taken in his research limits itself to labour expenses surrounding the core AM build process and ignores capital inputs outside of the AM system. Despite the prevalent perception of AM as a fully automated manufacturing technology requiring no manual input, this research demonstrates the importance of the human operators running the process.

Furthermore, distributed supply chains are proposed as a methodology to provide differentiated products to end users while minimising supply chain costs. However, in medical applications in particular, high marginal costs are associated with quality assurance and quality control,⁴⁷ which may negatively affect the real value proposition of AM where operated in a redistributed configuration.

The model constructed in this article suggests that the investigated AM technologies, with their roots in prototyping equipment, are prone to build failure. Due to the simple build failure model with a single failure node, this has simplified the facts by assuming that AM processes cannot fail during post-processing. In addition, it is assumed that any excessive deviation from intended part geometry or unrecoverable disturbance of the build process is classed as build failure, leading to the write-off of all parts contained in the build. In their study on the costs of laser sintering, Baumers and Holweg²⁴ provide a more detailed failure model, distinguishing between outright build failure, part rejection and material failure.

In this context, it is important to note that the dataset used into establish the probabilistic build failure models on both systems is small, with only 19 observations overall. A larger dataset could provide more details into the build failure behaviour of the investigated AM technologies, in particular to identify different build failure modes and relate them to the quality capability of the processes. Moreover, by treating the probability of build failure per layer p_Constant as fixed, this research establishes merely a rudimentary relationship between the probability of build failure and the geometry deposited in the build process. Thus, it is proposed that future work should extend the build failure model by taking one or more of the following approaches:

A Monte Carlo methodology could be implemented by generating a large artificial sample of data for testing.

A more detailed understanding of build failure could be implemented by associating the risk of failure of components and sub-processes in the build operation.

Investigations of the effect of different distributions of stochastic variables and failure models on expected cost.

As the main result of this research, the specific cost performance of both systems for a multi-part build at full capacity, £2.01/cm3 on the Objet Connex system and £1.73/cm3 on the Perfactory system, can be compared to data made available in the literature on the process economics of AM. Table 5 summarises a selection of such metrics extracted from the literature and summarises features of the underlying cost models.

Table 5.

Specific cost indices obtained from the literature.

Study	Technology variant	Material	Capacity utilisation	Support structures	Pre- and post-processing	Risk of build failure	Specific cost estimate^a
Hopkinson and Dickens¹⁸	Stereolithography	Unspecified epoxy polymer	Full build	Included	Manual pre- and post-processing	Ignored	£0.58/cm³
	Filament extrusion	Unspecified ABS	Full build	Included	Manual pre- and post-processing	Ignored	£0.49/cm³
	Laser sintering	Unspecified nylon	Full build	Not applicable	Manual pre- and post-processing	Ignored	£0.30/cm³
Ruffo et al.⁷	Laser sintering	DuraForm PA (nylon polymer)	Full build	Not required	Manual pre- and post-processing	Ignored	£0.29/cm³
Gibson et al.³¹	Stereolithography	Unspecified polymer	Full build	Included	Ignored	Ignored	£0.55/cm³
Grimm⁴⁸	Filament extrusion	ABSplus (polymer)	Single part	Included	Included	Ignored	£0.57/cm³
	Material jetting	EX200 (polymer)	Single part	Included	Included	Ignored	£1.52/cm³
	Film transfer imaging	FTI-GN (polymer)	Single part	Included	Included	Ignored	£0.77/cm³
	Material jetting	VeroWhite (polymer)	Single part	Included	Included	Ignored	£1.07/cm³
	Binder jetting	Zp150 (polymer composite)	Single part	Not required	Included	Ignored	£0.39/cm³
	Sheet lamination	PVC	Single part	Included	Included	Ignored	£1.42/cm³
Atzeni et al.²¹	Direct metal laser sintering	AlSi10Mg	Full build	Included as a 10% waste factor	Pre- and post-processing, heat treatment	Ignored	£9.48/cm³
Baumers et al.²²	Direct metal laser sintering	Stainless steel 17-4PH	Full build	Included	Only wire erosion	Ignored	£5.71/cm³
Baumers et al.³⁵	Material jetting	VeroClear	Full build	Included	Only build process	Ignored	£1.86/cm³
Baumers et al.³⁴	Selective heat sintering	SHS nylon composite	Full build	Not applicable	Only build process	Ignored	£1.49/cm³
Piili et al.⁴⁹	Direct metal laser sintering	Stainless steel PH1	Full build	Not applicable	Only build process	Ignored	£6.40/cm³
Baumers et al.⁵⁰	Electron beam melting	Ti6Al4V	Full build	Included	Only build process	Ignored	£6.18/cm³
Baumers and Holweg²⁴	Laser sintering	PA2200 (nylon)	Full build	Not applicable	Full process chain	Included	£0.84/cm³

Converted into pounds sterling according to exchange rate at the start of the publication year (where applicable). Where data used for specific cost estimation are not cited explicitly, they have been estimated from the data provided.

As can be seen from Table 5, the available cost metrics for material jetting systems range from £1.07 per cubic cm⁴⁸ to £1.86 per cubic cm.³⁵ Considering that these results ignore the expected cost of build failure, this is broadly in line with the specific cost level estimated in this article. For the assessed mask projection stereolithography system, no cost model was available in the literature. The cost models for stereolithography, which is a related AM technology, suggest a far lower specific cost range from £0.55 per cubic cm³¹ to £0.58 per cubic cm.¹⁸ This difference originates from the very high levels of capacity utilisation in the studies by Hopkinson and Dickens¹⁸ and Gibson et al.,³¹ which analyse builds that hold 190 and 1393 product units, respectively. When considering a build with uniform Z-heights, as done in execution B, the specific cost performance reported in this article for mask projection stereolithography is more similar at £1.00/cm³. Factoring in the high expected cost impact of build failure estimated at 29% in this research, the cited cost levels are broadly similar.

Conclusion

Starting with system-specific build time models, this article has developed detailed and comprehensive models of the process cost of the two investigated AM technology variants, material jetting and mask projection stereolithography. Despite using a limited dataset, the proposed models yield insight into build speed, process energy consumption and the cost performance of these systems subject to the risk of build failure. This information will help inform future research into the viability of AM for particular applications and promote commercially viable technology adoption.

The developed model helps improve the understanding of the effects of increasing throughput and capacity utilisation in the material deposition processes taking place within AM systems. As shown in this article for builds with identical Z-heights (executions A and B), increases in deposition rates lead to improvements in cost performance. However, these improvements are less than proportionate: on the Objet Connex system, an increase in the process speed of 183% led to an improvement in specific cost of 39%. On the Perfactory system, a rise in the process speed of 98% is associated with a cost improvement of 37%. The under-proportionate cost improvement has been shown to be the result of aspects outside of the cost deposition process, such as ancillary elements of the AM process, and cost relating to the risk of build failure. Especially, in the light of nascent polymeric AM processes designed for increased deposition speed, for example, the CLIP process⁵¹ or the Jet Fusion process,⁵² the requirement for a more automated work flow becomes clear.

The results of this research, obtained by coupling a computational build volume packing tool with a cost estimation technique, underline that an efficient pattern of operation is not a matter of a single manufacturing technology used in isolation. Rather, a successful wide-scale adoption of AM in manufacturing applications requires supporting complementary inputs, such as the adaptation of organisational structure, a suitable control architecture,⁵³ a supportive regulatory environment, product liability agreements and quality assurance methodologies, all of which form pertinent areas for additional research. A suitable framework for such investigations could be provided by the approach of design theory,⁵⁴ which aims to evaluate and improve practical problem solving in specific contexts, such as the adoption of AM in the medical devices domain. By utilising data generated from field testing rather than limited laboratory experiments, such research would allow more robust conclusions.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research has been funded by GSK R&D as part of an academic-industry collaborative project entitled 3D Printed Formulations: Additive Manufacture for Drug Delivery.

ORCID iD

Martin Baumers

References

ASTM 2015 ISO/ASTM52900-15. Standard terminology for additive manufacturing – general principles – terminology. West Conshohocken, PA: ASTM.

Tuck

Hague

Ruffo

et al . Rapid manufacturing facilitated customization. Int J Comp Integ Manuf 2008; 21(3): 245–258.

Boothroyd

. Product design for manufacture and assembly. Comput Aided Design 1994; 26(7): 505–520.

Weller

Kleer

Piller

. Economic implications of 3D printing: market structure models in light of additive manufacturing revisited. Int J Prod Econ 2015; 164: 43–56.

Eyers

Wong

Wang

et al . Rapid manufactured enabled mass customisation: untapped research opportunities in supply chain management. In: Proceedings of the logistics research network annual conference (ed Lyons

), Liverpool, 10–12 September 2008, pp.10–12.

Berman

. 3-D printing: the new industrial revolution. Bus Horizons 2012; 55(2): 155–162.

Ruffo

Tuck

Hague

. Cost estimation for rapid manufacturing-laser sintering production for low to medium volumes. Proc IMechE, Part B: J Engineering Manufacture Eng 2006; 220(9): 1417–1427.

Wohlers

. Wohlers report 2015. Fort Collins, CO: Wohlers Associates, 2015.

D’Aveni

. 3-D printing will change the world. Harvard Business Review, March 2013, https://hbr.org/2013/03/3-d-printing-will-change-the-world

10.

Masters

Velde

McBagonluri

. Rapid manufacturing in the hearing industry. Rapid Manuf Ind Revolut Digit Age 2006; 195–209, https://onlinelibrary.wiley.com/doi/10.1002/0470033991.ch12

11.

Bibb

Eggbeer

Evans

et al . Rapid manufacture of custom-fitting surgical guides. Rapid Prototyping J 2009; 15(5): 346–354.

12.

Bartolo

Kruth

Silva

et al . Biomedical production of implants by additive electro-chemical and physical processes. CIRP Ann: Manuf Techn 2012; 61(2): 635–655.

13.

Dawood

Marti

Sauret-Jackson

et al . D printing in dentistry. Brit Dent J 2015; 219(11): 521–529.

14.

Prasad

Smyth

. D Printing technologies for drug delivery: a review. Drug Dev Ind Pharm 2016; 42(7): 1019–1031.

15.

Jones

Webb

Summers

et al . A cost–benefit framework for assessing advanced manufacturing technology development: a case study. Proc IMechE, Part B: J Engineering Manufacture Eng 2015; 229(9): 1654–1668.

16.

Niazi

Dai

Balabani

et al . Product cost estimation: technique classification and methodology review. J Manuf Sci Eng 2006; 128(2): 563–575.

17.

Alexander

Allen

Dutta

. Part orientation and build cost determination in layered manufacturing. Comput Aided Design 1998; 30(5): 343–356.

18.

Hopkinson

Dickens

. Analysis of rapid manufacturing – using layer manufacturing processes for production. Proc IMechE, Part C: J Mechanical Engineering Science 2003; 217(1): 31–39.

19.

Ruffo

Hague

. Cost estimation for rapid manufacturing–simultaneous production of mixed components using laser sintering. Proc IMechE, Part B: J Engineering Manufacture Eng 2003; 221(11): 1585–1591.

20.

Atzeni

Salmi

. Economics of additive manufacturing for end-usable metal parts. Int J Adv Manuf Tech 2012; 62(9–12): 1147–1155.

21.

Atzeni

Iuliano

Minetola

et al . Redesign and cost estimation of rapid manufactured plastic parts. Rapid Prototyping J 2010; 16(5): 308–317.

22.

Baumers

Tuck

Wildman

et al . Transparency built-in. J Ind Ecol 2013; 17(3): 418–431.

23.

Rickenbacher

Spierings

Wegener

. An integrated cost-model for selective laser melting (SLM). Rapid Prototyping J 2013; 19(3): 208–214.

24.

Baumers

Holweg

. Cost impact of the risk of build failure in laser sintering. In: Proceedings of the solid freeform fabrication symposium, University of Texas at Austin, Austin, TX, 7–10 August 2016, http://www.programmaster.org/PM/PM.nsf/ApprovedAbstracts/497F5B35F092073685257FA8006BB5C9?OpenDocument

25.

Son

. A cost estimation model for advanced manufacturing systems. Int J Prod Res 1991; 29(3): 441–452.

26.

Schmid

Levy

. Quality management and estimation of quality costs for additive manufacturing with SLS. In: Fraunhofer Direct Digital Manufacturing Conference, ETH-Zürich, 2012.

27.

Berumen

Bechmann

Lindner

et al . Quality control of laser-and powder bed-based additive manufacturing (AM) technologies. Phy Proc 2010; 5: 617–622.

28.

Di Angelo

Di Stefano

. A neural network-based build time estimator for layer manufactured objects. Int J Adv Manuf Tech 2011; 57(1–4): 215–224.

29.

Stratasys. Corporate website, 2017, www.stratasys.com

30.

Envisiontec. Corporate website, 2015, www.envisiontec.com

31.

Gibson

Rosen

Stucker

. Additive manufacturing technologies. 2nd ed.New York: Springer, 2014.

32.

Kim

. A benchmark study on rapid prototyping processes and machines: quantitative comparisons of mechanical properties, accuracy, roughness, speed, and material cost. Proc IMechE, Part B: J Engineering Manufacture Eng 2008; 222(2): 201–215.

33.

Clare

Gao

Brkić

et al . Linear ion trap fabricated using rapid manufacturing technology. J Am Soc Mass Spectr 2010; 21(2): 317–322.

34.

Baumers

Tuck

Wildman

et al . Transparency built-in. Journal of Industrial Ecology 2013; 17(3); 418–431.

35.

Baumers

Tuck

Dickens

et al . How can material jetting systems be upgraded for more efficient multi-material additive manufacturing? In: Proceedings of the solid freeform fabrication symposium, Austin, TX, 4–6 August 2014.

36.

Cheng

Fuh

JYH

Nee

AYC

et al . Multi-objective optimization of part-building orientation in stereolithography. Rapid Prototyping J 1995; 1(4): 12–23.

37.

Campbell

Combrinck

de Beer

et al . Stereolithography build time estimation based on volumetric calculations. Rapid Prototyping J 2008; 14(5): 271–279.

38.

Munguía

Ciurana

Riba

. Neural-network-based model for build-time estimation in selective laser sintering. Proc IMechE, Part B: J Engineering Manufacture Eng 2009; 223(8): 995–1003.

39.

Wilson

. Selection for rapid manufacturing under epistemic uncertainty. Master’s Thesis, Georgia Institute of Technology, Atlanta, GA.

40.

Araújo

LJP

Özcan

Atkin

et al . Toward better build volume packing in additive manufacturing: classification of existing problems and benchmarks, 2015, https://pdfs.semanticscholar.org/a592/080bca29d7198ceb9409de4c30798483b4a8.pdf

41.

Schmenner

Swink

. On theory in operations management. J Oper Manag 1998; 17(1): 97–113.

42.

Hur

Choi

Lee

et al . Determination of fabricating orientation and packing in SLS process. J Mater Process Tech 2001; 112(2): 236–243.

43.

Ikonen

Biles

Kumar

et al . A genetic algorithm for packing three-dimensional non-convex objects having cavities and holes. In: Proceedings of the 7th international conference on genetic algorithms (ICGA), East Lansing, MI, 19–23 July 1997, pp.591–598. San Francisco, CA: Morgan Kaufmann Publishers Inc.

44.

Canellidis

Dedoussis

Mantzouratos

et al . Preprocessing methodology for optimizing stereolithography apparatus build performance. Comput Ind 2006; 57(5): 424–436.

45.

Haldi

Whitcomb

. Economies of scale in industrial plants. J Polit Econ 1967; 75: 373–385.

46.

Pine

. Mass customization: the new frontier in Business Competition. Boston, MA: Harvard Business School Press, 1993.

47.

Thurman-Newell

Harrison

Hourd

et al . Cell micro-factories: feasibility of a redistributed manufacturing model for cell based therapeutics. In: Proceedings of the EurOMA: 23th international annual EurOMA conference: ‘interactions’, Trondheim, 17–22 June 2016. Trondheim: EurOMA, http://euroma2016.euroma-online.org/www.euroma2016.org/index.html

48.

Grimm

. 2010Rapid prototyping benchmark – 3D printers. Edgewood, KY: Grimm & Associates, Inc.

49.

Piili

Happonen

Väistö

et al . Cost estimation of laser additive manufacturing of stainless steel. Phy Proc 2015; 78: 388–396.

50.

Baumers

Dickens

Tuck

et al . The cost of additive manufacturing: machine productivity, economies of scale and technology-push. Technol Forecast Soc 2016; 102: 193–201.

51.

Carbon3D. Corporate website, 2017, carbon3d.com

52.

HP Inc. Corporate website: 3D printing with Multi Jet Fusion™ technology, 2015, http://www8.hp.com/us/en/commercial-printers/floater/3Dprinting.html

53.

Eyers

. Control architectures for industrial additive manufacturing systems. Proc IMechE, Part B: J Engineering Manufacture Eng 2018; 232(10): 1767–1777.

54.

Denyer

Tranfield

Van Aken

. Developing design propositions through research synthesis. Organ Stud 2008; 29(3): 393–413.