Manufacturing Letters 31 (2022) 28–32

Contents lists available at ScienceDirect

Manufacturing Letters
journal homepage: www.elsevier.com/locate/mfglet

A scalable digital platform for the use of digital twins in additive

manufacturing
Luke Scime ⇑, Alka Singh, Vincent Paquit
Electrification and Energy Infrastructure Division, Oak Ridge National Laboratory, Oak Ridge, TN 37830, United States

a r t i c l e i n f o a b s t r a c t

Article history: While Additive Manufacturing promises to reshape the manufacturing landscape, challenges related to
Received 8 April 2021 part, and process qualification hinder its widespread adoption. The Instance-Qualified approach seeks
Accepted 24 May 2021 to qualify individual parts, even for processes with high variability, by leveraging the concept of a digital
Available online 18 June 2021
twin. This work proposes a scalable cyberphysical infrastructure to enable the construction and use of
such digital twins. This work also introduces the concept of an Augmented Intelligence Relay, which
Keywords: allows Artificial Intelligence algorithms to predict component performance for a given application even
Additive manufacturing
when it is impractical to perform a large number of physical tests.
Digital twin
Artificial intelligence
Ó 2021 Society of Manufacturing Engineers (SME). Published by Elsevier Ltd. All rights reserved.
Instance-qualified
Cyberphysical infrastructure

1. Introduction Here we propose a cyberphysical infrastructure designed to

enable the production of Instance-Qualified AM components, as it
Additively Manufactured (AM) components are constructed is being developed at ORNL’s Manufacturing Demonstration Facil-
directly from feedstock, typically in a layer-wise fashion. ity (MDF). The proposed methodology pays special attention to
Nominally, AM is ideal for fabricating complex geometries for scalability and the application of Artificial Intelligence (AI). Pro-
high-value industries [1]. However, significant variability in AM cessing knowledge collected over the last century of subtractive
processes leads to inconsistent properties (particularly with manufacturing now enables high-quality production. The proposed
respect to microstructure and defects) [2–4], hindering broader methodology strives to achieve similar results for AM in a fraction
adoption [1]. Low production volumes further complicate the of the time.
application of traditional manufacturing quality assurance
methodologies. In lieu of alternatives, industries currently leverag-
ing AM rely on non-destructive testing for part qualification; which 2. Methodology
is often not cost-effective.
For these reasons, many researchers have suggested that a new 2.1. Scalability
qualification paradigm is needed. Referred to as ‘‘Instance-
Qualified,” this approach seeks to qualify components using a We refer to the cyberphysical infrastructure required to imple-
digital twin [5] based on instance-specific, as opposed to aggregate, ment a digital twin as the digital platform. The digital platform is a
data. First proposed by Grieves and Vickers [6], a digital twin is a network of manufacturing equipment, data storage systems, and
computer model of a real object, assembly, or system which is computation capabilities. In our implementation, only metadata
updated based on data collected from its physical twin. This are stored within a relational database while the data are stored
approach is particularly promising as the layer-wise nature of within indexed file systems. By enforcing a minimally-viable data
AM allows data to be collected during production. While several structure, we can integrate new manufacturing processes more
papers have explicitly addressed the application of digital twins quickly into the digital platform. Data can be queried and accessed
to AM, they have primarily focused narrowly on thermo- using a web portal or an application programming interface
mechanical modeling [7,8]. embedded in custom software. While the web portal supports
some data visualization, most visualization and analytics are per-
formed using software deployed on networked computers.
⇑ Corresponding author. To construct a digital twin for an AM component, the entire
E-mail address: [email protected] (L. Scime). manufacturing process, including design, simulation, printing,

https://doi.org/10.1016/j.mfglet.2021.05.007
2213-8463/Ó 2021 Society of Manufacturing Engineers (SME). Published by Elsevier Ltd. All rights reserved.
L. Scime, A. Singh and V. Paquit Manufacturing Letters 31 (2022) 28–32

post-processing, and characterization, must be captured. There- approach is that each part must be uniquely identifiable to facili-
fore, we define the digital workflow as the interaction between tate the linkage between the digital and physical twins. Strictly,
the physical manufacturing process, the technicians and engineers, each digital thread stored in the digital platform’s database is the
and the digital platform. To maintain scalability, each manufactur- sequential list of operations which were performed during manu-
ing process is decomposed into a sequence of operations. These facturing, while the collected data remain associated with the cor-
operations may create a part, modify a part, characterize a part, responding operation. While this less-structured approach is not
act on a feedstock, or modify a printer. During each operation, data ideal for data analytics, it is necessary because (1) a given operation
are collected and transmitted to the digital platform either manu- may act on multiple parts which may not be separable in the data,
ally or automatically via operation-specific software tools. Fig. 1 (2) any part may be broken into child parts or integrated into an
shows an example digital workflow for the manufacture of silicon assembly and the duplication of data should be minimized, and
carbide fuel elements for a nuclear reactor, starting with binder (3) the relevant datasets can be diverse requiring flexibility in their
jet printing and ending with component characterization [9,10]. transmission, storage, and analysis. The digital thread for each part
The totality of the design intent and data collected during the must also include historical information about the feedstock and
fabrication of a given component is considered a digital thread machines used in its creation. Fig. 2 shows a digital thread within
and it instantiates the corresponding digital twin. A corollary of this the digital platform.

Fig. 1. The digital workflow for producing a silicon carbide fuel element for a nuclear reactor via binder jet AM.

29
L. Scime, A. Singh and V. Paquit Manufacturing Letters 31 (2022) 28–32

Fig. 2. Digital threads for parts produced via laser powder bed fusion AM. Parts 2–5 are considered children of part 1 and inherit its digital thread. Each part’s digital thread
links to the feedstock and manufacturing equipment at a specific point in time.

Fig. 3. An example AIR which uses in-situ imaging to predict localized tensile properties for a laser powder bed fusion AM process.

30
L. Scime, A. Singh and V. Paquit Manufacturing Letters 31 (2022) 28–32

2.2. Artificial intelligence to ongoing research. In this work, we argue that decomposition
of manufacturing processes into discrete operations is key to
Creating an Instance-Qualified component requires modeling deploying digital workflows. Furthermore, we propose the use of
the physical twin’s performance. While such models may be an AIR to achieve part property prediction based on in-situ data.
physics-based, the large quantity and high dimensionality of data Notably, the proposed methodology only extends to instantiation
collected during AM suggests that targeted use of AI is appropriate. of a digital twin, the full value of a digital twin will only be realized
While the use of machine and deep learning [11] for AM part qual- once data from the entire product lifecycle can be integrated into
ification is actively researched, most work has focused primarily on the digital model. Finally, while this work focused on AM, the pro-
defect detection [12–17]. Less progress has been made toward posed digital platform is broadly applicable across manufacturing
scalable property predication, particularly as it relates to geometry processes.
agnosticism and applicability across materials and printers.
Ideally, an AI would ingest digital threads and learn to predict Declaration of Competing Interest
part performance based on the collected in-situ processing data
and characterization results. Unfortunately, contemporary AI The authors declare that they have no known competing finan-
requires vast quantities of labeled training data [11,18]. For exam- cial interests or personal relationships that could have appeared
ple, using deep learning to directly predict part performance based to influence the work reported in this paper.
on in-situ imaging might require 109 physical tests, which is
impractical. This problem is compounded when part geometry, Acknowledgements
application environment, and post-processing operations are con-
sidered. Therefore, we propose an Augmented Intelligence Relay The authors thank the MDF team for supporting this work
(AIR) in which a sequence of AIs solve different aspects of a which is sponsored by the US Department of Energy’s Advanced
problem. Manufacturing Office and the Transformational Challenge Reactor
In an AIR, human experts observe intermediate results at the program. This manuscript has been authored by UT-Battelle, LLC,
interface between each AI and perform feature-engineering to under contract DE-AC05-00OR22725 with the US Department of
structure the data. While the work remains preliminary, Fig. 3 Energy (DOE). The US government retains and the publisher, by
shows one implementation of an AIR. In this scenario, in-situ accepting the article for publication, acknowledges that the US
images from a laser powder bed fusion printer are used to predict government retains a nonexclusive, paid-up, irrevocable, world-
local tensile properties by leveraging data from three different wide license to publish or reproduce the published form of this
types of builds (1) SS-J3 tensile samples, (2) representative tube manuscript, or allow others to do so, for US government purposes.
geometries, and (3) application-specific geometries. First, data col- DOE will provide public access to these results of federally spon-
lected from all three sets of builds – representing 108 labeled pixels sored research in accordance with the DOE Public Access Plan
– are used to train the anomaly detection deep learning algorithm (https://energy.gov/downloads/doe-public-access-plan).
implemented by Peregrine [12]. At the first interface, a human engi-
neers voxelized feature vectors which summarize the anomaly
References
detections, encode local part geometry information, and incorpo-
rate post-processing operation data. These feature vectors are input [1] Wohlers T, Campbell I, Diegel I, Huff R, Kowen J. Wohlers report 2020: 3D
to a machine learning algorithm which predicts voxelized tensile printing and additive manufacturing global state of the industry. CO: Fort
properties and is trained by 104 tensile testing operations. Finally, Collins; 2020.
[2] Fisher B, Mireles J, Ridwan S, Wicker R, Beuth J. Consequences of part
validation can be performed by combining the voxelized properties temperature variability in electron beam melting of Ti-6Al-4V. J Mater 2017.
with a physics-based model of the tube geometries and performing https://doi.org/10.1007/s11837-017-2597-y.
102 burst tests. Throughout the AIR, the AIs become less data- [3] Cunningham R, Narra SP, Montgomery C, Beuth J, Rollett AD. Synchrotron-
based X-ray microtomography characterization of the effect of processing
intensive as the ground truths become more difficult to acquire. variables on porosity formation in laser power-bed additive manufacturing of
Ti-6Al-4V. J Mater 2016;69. https://doi.org/10.1007/s11837-016-2234-1.
[4] Foster BK, Reutzel EW, Nassar AR, Hall BT, Brown SW, Dickman CJ. Optical,
3. Results and discussion layerwise monitoring of powder bed fusion. Solid Free. Fabr. 2015.
[5] Kritzinger W, Karner M, Traar G, Henjes J, Sihn W. Digital twin in
manufacturing: a categorical literature review and classification. IFAC-
Development of the MDF digital platform initially focused on
PapersOnLine 2018;51:1016–22. https://doi.org/10.1016/j.ifacol.2018.08.474.
seven powder bed printers and is currently scaling to another 80 [6] Grieves M. Digital Twin: Manufacturing Excellence through Virtual Factory
+ printers. In CY2020, 6.7 TB of data from 410 builds were Replication This paper introduces the concept of a A Whitepaper by Dr .
uploaded to the digital platform and analyzed. During this process, Michael Grieves. 2015.
[7] DebRoy T, Zhang W, Turner J, Babu SS. Building digital twins of 3D printing
several axioms became clear – flexibility over structure, automa- machines. Scr Mater 2017;135:119–24. https://doi.org/10.1016/j.
tion over manual data entry, and scalability necessitates stream- scriptamat.2016.12.005.
lined software tools. Another realization is that the success of a [8] Mies D, Marsden W, Warde S. Overview of additive manufacturing informatics:
‘‘a digital thread”. Integr Mater Manuf Innov 2016;5:114–42. https://doi.org/
digital platform is dependent upon the data discipline of the overall 10.1186/s40192-016-0050-7.
organization. We found that providing immediate ‘‘digital” value to [9] Terrani KA, Jolly BC, Trammell MP, Vasudevamurthy G, Schappel D, Ade B, et al.
the person uploading the data (e.g. allowing builds to be remotely Architecture and properties of TCR fuel form. J Nucl Mater 2021;547:. https://
doi.org/10.1016/j.jnucmat.2021.152781152781.
monitored) improved acceptance of new data collection proce- [10] Terrani K, Jolly B, Trammell M. 3D printing of high-purity silicon carbide. J Am
dures. Finally, addressing process scalability has facilitated rapid Ceram Soc 2020;103:1575–81. https://doi.org/10.1111/jace.16888.
transfer of MDF-developed technologies to the broader manufac- [11] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436–44. https://
doi.org/10.1038/nature14539.
turing community.
[12] Scime L, Siddel D, Baird S, Paquit V. Layer-wise anomaly detection and
classification for powder bed additive manufacturing processes: a machine-
agnostic algorithm for real-time pixel-wise semantic segmentation. Addit
4. Conclusions Manuf 2020;36:. https://doi.org/10.1016/j.addma.2020.101453101453.
[13] Zur Jacobsmuhlen J, Kleszczynski S, Witt G, Merhof D. Detection of elevated
While the schemas presented in this work will evolve, the cur- regions in surface images from laser beam melting processes. Conf. IEEE Ind.
Electron. Soc. 2016:1270–5. https://doi.org/10.1109/IECON.2015.7392275.
rent methodology has allowed data visualization and analytics [14] Aminzadeh M, Kurfess TR. Online quality inspection using Bayesian
efforts to rapidly scale across the MDF and add substantial value classification in powder-bed additive manufacturing from high-resolution

31
L. Scime, A. Singh and V. Paquit Manufacturing Letters 31 (2022) 28–32

visual camera images. J Intell Manuf 2018:1–19. https://doi.org/10.1007/ neural networks. Addit Manuf 2017. https://doi.org/10.1016/j.
s10845-018-1412-0. addma.2017.11.012.
[15] Khanzadeh M, Bian L, Shamsaei N, Thompson SM. Porosity detection of laser [17] Meng L, McWilliams B, Jarosinski W, Park H-Y, Jung Y-G, Lee J, et al. Machine
based additive manufacturing using melt pool morphology clustering. Solid learning in additive manufacturing: a review. JOM 2020;72:2363–77. https://
Free. Fabr. 2016:1487–94. doi.org/10.1007/s11837-020-04155-y.
[16] Shevchik SA, Kenel C, Leinenbach C, Wasmer K. Acoustic emission for in situ [18] Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, et al. ImageNet large
quality monitoring in additive manufacturing using spectral convolutional scale visual recognition challenge. Int J Comput Vis 2015;115:211–52. https://
doi.org/10.1007/s11263-015-0816-y.

32