Application of Machine Learning in Additive

Manufacturing
Comprehensive Review of Machine Learning Applications Across the Additive
Manufacturing Lifecycle

Nehaa Vijayakumar
400607375
28 March 2025

although the various table text styles are provided. The formatter
I. INTRODUCTION will need to create these components, incorporating the
Additive manufacturing (AM), also referred to as 3D applicable criteria that follow.
printing, is a transformative technology that builds objects layer-
by-layer from digital models. Unlike traditional subtractive II. CATEGORIES OF MACHINE LEARING IN ADDITIVE
manufacturing processes, AM enables the creation of highly MANUFACTURING
complex geometries, minimizes material waste, and supports
on-demand production. However, despite its numerous The integration of machine learning (ML) into additive
advantages, AM faces critical challenges that limit its manufacturing (AM) processes has opened new avenues for
widespread adoption in high-stakes industries such as aerospace, process optimization, defect detection, quality control, and
medical, and automotive. These challenges include inconsistent predictive maintenance. The classification of ML applications in
part quality, high defect rates, complex process optimization AM can be effectively mapped across different stages of the
requirements, and limited real-time control. To address these
production pipeline. This section elaborates on six main
challenges, the integration of Machine Learning (ML) into the
categories, with detailed explanations and expanded discussions
AM workflow has become increasingly important.
of their impact and potential.
Machine learning, a subset of artificial intelligence, allows
systems to learn from data and make predictions or decisions A. Design phase Applications
without being explicitly programmed. In the context of AM, ML In the design phase, ML contributes to topology
can be applied at various stages of the manufacturing pipeline, optimization, generative design, and support structure
from design and material selection to in-situ monitoring and prediction. Generative design tools use ML algorithms to
post-process inspection. ML models can analyse large volumes explore countless design alternatives that meet specified
of sensor data to detect anomalies, optimize process parameters constraints, such as stress thresholds and manufacturability. By
for better outcomes, predict part properties, and even enable analysing historical CAD models and simulation data, ML can
closed-loop control systems that adapt in real-time. These propose more efficient designs that optimize both performance
capabilities not only enhance part quality and reduce defects but and material usage. Support prediction tools reduce material
also improve the overall efficiency and cost-effectiveness of consumption and ensure part stability during printing
AM operations. (Chowdhury & Anand, 2016).
This paper explores the multifaceted applications of ML in
AM, categorizing them into distinct areas based on the B. Pre-processing Applications
manufacturing workflow. It provides detailed examples for each During the pre-processing stage, ML is used for tasks like
category and discusses the relevance of supervised and material classification, slice optimization, and thermal
unsupervised learning paradigms. Moreover, it highlights the deformation compensation. Data-driven models can forecast
contributions of specific Convolutional Neural Network (CNN) distortions caused by temperature gradients and suggest
architectures in classification and segmentation tasks, presents a geometric adjustments. These models reduce build failures and
case study involving reinforcement learning, and offers ensure dimensional accuracy. Neural networks, for instance, can
suggestions for future research. By understanding and predict the heat-induced warping of components and
leveraging the power of ML, stakeholders in AM can overcome recommend pre-compensated geometries (Chowdhury &
existing limitations and unlock new levels of performance, Anand, 2016).
reliability, and scalability (Wang et al., 2018). of electronic
products, and (3) conformity of style throughout a conference C. In-process Monitoring
proceedings. Margins, column widths, line spacing, and type
ML plays a pivotal role in real-time monitoring of AM
styles are built-in; examples of the type styles are provided
processes. Sensor data, such as images from thermal cameras,
throughout this document and are identified in italic type, within
melt pool dimensions, and acoustic emissions are analysed to
parentheses, following the example. Some components, such as
detect anomalies. Convolutional Neural Networks (CNNs) can
multi-levelled equations, graphics, and tables are not prescribed,
detect defects like pores, cracks, or balling based on melt pool III. BRIEF DISCRIPTION OF EACH CATEGORY
imagery. When integrated into a closed-loop control system, ML Each of the six categories of ML application in AM can be
allows for real-time corrections to mitigate defects as they expanded with detailed explanations and practical examples to
emerge (Baumgartl et al., 2020). demonstrate the value added at every stage. Below is an
elaboration on each category.
D. Process Parameter Optimization
Optimizing process parameters—such as laser power, scan A. Design phase Applications
speed, and hatch spacing—is critical for ensuring quality. ML In the design phase, ML algorithms can recommend part
regression models learn from past builds and simulate outcomes geometries that are structurally optimized while still being
based on input conditions. Gaussian Process Regression (GPR) manufacturable. Deep learning algorithms can be trained on
and polynomial regression have been used to create process historical CAD data to recommend support structures that
maps, which predict porosity and surface finish across parameter minimize material usage while maintaining structural integrity.
ranges (Tapia et al., 2018). Generative design platforms use ML to automatically generate
multiple design alternatives based on specific input constraints,
E. Post-processing Inspection such as load-bearing capacity or thermal dissipation. These
High-resolution imaging, combined with ML, allows for algorithms can drastically reduce the time engineers spend on
automated defect detection during post-processing. CNNs can trial-and-error iterations, increasing innovation and part
classify defects by type and severity, streamlining quality performance.
assurance. This reduces reliance on manual inspection and
Example: Design feature recommendation systems analyse
improves repeatability. Vision-based inspection systems can
detect surface anomalies, ensuring that only conforming parts existing designs and suggest features that improve performance
proceed to the next production phase (Zhang et al., 2019). or reduce manufacturing complexity. ML-based support
structure prediction tools ensure stability during the printing
process, as discussed by Chowdhury and Anand (2016).
F. Cost and Quality Estimation
ML aids in predicting build time, energy usage, and material B. Pre-processing Applications
consumption. These insights are essential for resource planning
and cost optimization. Models trained on historical production Pre-processing involves preparing the digital model for
data help in estimating the cost per part and flag potential quality printing, selecting materials, and compensating for thermal
issues before printing begins (Seltzer et al., 2015). stresses. ML can be applied to predict thermal distortion and
suggest geometric adjustments. This predictive approach
significantly reduces build failures and increases dimensional
accuracy. Material classification using vision-based ML models
also ensures that only suitable powders or filament types are
chosen for printing.

Example: Chowdhury and Anand used an ANN to predict

and compensate for geometric deformation caused by thermal
stresses in the AM process.
C. In-process Monitoring
Real-time monitoring using ML is crucial for detecting and
preventing defects during the build. ML models can analyse
sensor signals (e.g., thermographic, acoustic, or camera-based)
to detect anomalies and take corrective actions. This results in
less material waste and increased part reliability. Sensors
combined with ML allow the creation of closed-loop systems
Figure 1. Topology Optimization and Machine that adjust parameters during printing to eliminate detected
Learning Design for 3D-Printed Materials flaws on the fly.
Example: Baumgartl et al. implemented CNN-based
thermographic monitoring to detect porosity and other defects
during laser powder bed fusion, leading to early detection and
Understanding these categories enables manufacturers to
strategically implement ML across the AM workflow. The improved print quality.
ability to predict outcomes, monitor processes, and optimize D. Process Parameter Optimization
parameters contributes to enhanced reliability, reduced waste,
ML models are widely used to optimize process parameters
and improved scalability. As AM technologies evolve, the role
such as laser power, scanning speed, and hatch spacing. These
of ML in facilitating intelligent manufacturing will only grow
stronger. parameters critically affect part quality and mechanical
properties. By learning from historical builds, ML can
recommend optimal combinations to achieve specific goals such that combine both approaches are being developed for robust,
as maximum density or minimum surface roughness. adaptive control in AM.
Example: Tapia et al. used Gaussian Process Regression to As AM continues to generate large volumes of multi-modal
create surrogate models that map input parameters to outputs data (thermal, visual, acoustic), the importance of both
like porosity, enabling optimization without exhaustive supervised and unsupervised learning will grow. Leveraging
experimentation. their strengths helps manufacturers increase automation, reduce
E. Post-Processing Inspection errors, and build intelligence directly into the AM pipeline.
Automated inspection using ML reduces the need for
manual labour and increases consistency in quality checks. V. CNN IN AM RESEARCH
Image-based inspection systems use trained CNNs to evaluate Convolutional Neural Networks (CNNs) are a class of deep
part surfaces, identifying defects that may not be visible to the learning models particularly suited for analysing spatial
human eye. These models can be integrated into production lines hierarchies in data, especially images. Their use in additive
for real-time validation. manufacturing (AM) has revolutionized how sensor data,
Example: Zhang et al. used CNNs to identify surface including thermal and optical images, are interpreted. CNNs
roughness and dimensional inaccuracies in AM parts. play a vital role in improving process control, part quality, and
defect detection.
These examples show how ML is deeply integrated into every A notable application of CNNs in AM is real-time melt pool
stage of AM. Through practical implementation, these analysis. For instance, Fang et al. (2021) used a U-Net-based
technologies can drastically reduce failures, improve quality, CNN to perform semantic segmentation on in-situ melt pool
and make AM more scalable for industrial use. images in laser powder bed fusion (LPBF). The architecture
enabled precise delineation of thermal profiles, which in turn
helped detect anomalies like keyholing or lack of fusion. Such
IV. SUPERVISED AND UNSUPERVISED LEARING IN AM insights are critical for real-time control systems aiming to
Machine learning techniques are broadly classified into reduce defects during printing.
supervised and unsupervised learning, each offering unique CNNs have also been used in both online and offline
advantages in the context of additive manufacturing (AM). monitoring systems. Online CNN systems process thermal or
These paradigms differ in how the model learns from data and visible images during printing, flagging anomalies instantly.
the types of problems they solve. Offline systems, on the other hand, analyse high-resolution
Supervised learning involves training models on labelled images of finished parts to detect surface defects, porosity, or
datasets, where each input is paired with a known output. It is dimensional errors. These applications reduce the need for
especially useful for predictive tasks, such as classifying defects, manual inspection and ensure higher repeatability.
predicting mechanical properties, or estimating build outcomes Advanced CNN architectures, such as ResNet and
based on process parameters. These models require large, EfficientNet, are now being trained on larger datasets using
curated datasets but deliver high accuracy when well-trained. transfer learning techniques. This has increased their robustness
They are commonly used in AM for tasks like defect across different AM platforms and materials. In addition, CNNs
classification, melt pool anomaly detection, and dimensional are being integrated with other ML models for multi-modal
accuracy predictions. analysis, combining thermal, acoustic, and optical data.
Example: Gaja and Liou (2018) applied supervised learning The success of CNNs in AM stems from their ability to
techniques such as logistic regression and artificial neural generalize across complex defect patterns and their adaptability
networks (ANNs) for defect classification in laser metal to new tasks with limited training data. As the quality and
deposition, using features like energy and rise time as input. quantity of AM datasets improve, CNNs will continue to drive
automation, real-time quality control, and intelligent process
Unsupervised learning, in contrast, deals with unlabelled management in AM systems.
data. It is used to uncover hidden structures, group similar data
points, or detect anomalies without predefined classes.
Clustering algorithms, dimensionality reduction methods like VI. CLASSIFICATION CNN MODEL
PCA, and autoencoders are widely used in this domain. In AM,
unsupervised learning is used for tasks like powder While MobileNet is widely known for its efficiency on
classification, pattern discovery in process data, and defect mobile platforms, several other CNN models are particularly
detection where labels are scarce or unavailable. well-suited for defect classification tasks in additive
manufacturing (AM). One such model is LeNet, originally
Example: Clustering algorithms have been applied to developed for digit recognition, which has found its niche in
identify variations in powder particle morphology, allowing industrial applications due to its simplicity and speed.
classification of feedstock quality and detecting outliers that LeNet consists of a sequence of convolutional layers
may cause defects. followed by subsampling (pooling) layers and fully connected
The two paradigms are often complementary. For example, layers. It can handle simple classification problems effectively,
unsupervised methods can be used to pre-process or segment especially in scenarios where high accuracy is needed without
data before feeding it into a supervised model. Hybrid models extensive computational resources. In AM, LeNet can classify
thermal or visual images into categories such as “porosity,” industries demanding high precision and traceability, such as
“crack,” or “no defect.”
For instance, LeNet can be deployed in real-time systems that
analyse melt pool imagery or post-process scans. By training the
model on labelled datasets of known defect types, it learns to
associate specific visual patterns with specific classes. This
capability reduces reliance on manual inspection and enables
automated defect tracking during or after a print job.
LeNet is also favoured in academic research for
benchmarking and prototyping. Researchers can quickly
implement it using frameworks like TensorFlow or PyTorch and
test new ideas, such as novel preprocessing techniques or dataset
augmentation strategies. The architecture serves as a baseline for
evaluating more complex models like ResNet, Inception, or
DenseNet.
Moreover, LeNet's simplicity allows it to be deployed on edge
devices or embedded systems used in in-situ monitoring. It
provides fast, reliable classification results that are essential for Figure 2. The structure of the Mask R-CNN architecture
real-time feedback loops.
Overall, LeNet remains a valuable and versatile CNN model aerospace and biomedical sectors.
for AM classification tasks, offering a balance between
performance, interpretability, and computational efficiency.
VIII. SEGMENTATION CNN MODEL
Reinforcement Learning (RL) is a machine learning paradigm
VII. SEGMENTATION CNN MODEL where an agent learns to make decisions through interaction with
Beyond U-Net, one of the most effective models for an environment, aiming to maximize a reward signal. Its
segmentation tasks in additive manufacturing (AM) is Mask R- application in additive manufacturing (AM) is growing,
CNN. This architecture extends Faster R-CNN by incorporating particularly in areas involving real-time control, adaptive
an additional branch that generates pixel-level segmentation process tuning, and predictive maintenance.
masks for detected objects. Unlike U-Net, which focuses on In a typical RL setup for AM, the environment represents the
semantic segmentation, Mask R-CNN excels in instance printing process, and the agent is a controller that adjusts
segmentation, enabling it to identify and separate multiple parameters like laser power, scan speed, or layer thickness. The
overlapping defects in a single image. reward function is designed to optimize quality metrics such as
In AM, Mask R-CNN is especially useful during post- dimensional accuracy, surface finish, or defect minimization.
processing inspection where multiple defect types—such as Over time, the RL agent learns policies that lead to better
porosity clusters or surface cracks—need to be detected and outcomes through exploration and feedback.
segmented individually. The model provides a bounding box, a One example is the work of Lee and Prabhu (2016), who
class label, and a mask for each detected object. This detailed proposed a simulation-driven framework for thermal control in
output enables precise defect tracking, quality grading, and LPBF processes. While their approach wasn’t fully RL-based, it
defect density analysis. introduced core RL principles like feedback-based learning and
The architecture relies on a backbone network, typically adaptive control. Future implementations could build on this by
ResNet, to extract features. These are then used by the Region integrating true RL algorithms such as Q-learning or Deep
Proposal Network (RPN) to suggest candidate object regions. Deterministic Policy Gradients (DDPG) to optimize complex
The final stage of the model performs classification, bounding control policies.
box regression, and mask generation. This multi-task learning In practice, RL is most promising when combined with
approach enhances accuracy without significantly increasing digital twins—virtual replicas of the AM process that allow safe
computational cost when GPUs are available. and rapid experimentation. Agents can be trained in simulation
Mask R-CNN’s performance in AM has been demonstrated and later deployed in real systems, minimizing real-world risks.
in research focusing on defect detection in powder bed fusion RL can also be integrated with other ML techniques to form
processes, where the ability to distinguish individual defect hybrid control systems that respond to sensor feedback in real-
instances is critical. Furthermore, the model's output can be fed time.
into data analytics pipelines to predict process health or guide As sensor resolution and computational power improve, RL
rework strategies. will likely become an integral part of closed-loop control in AM.
Despite being computationally heavier than U-Net, Mask R- Its ability to learn from experience and adapt to changing
CNN remains a strong choice where detailed inspection and conditions makes it ideal for optimizing complex, dynamic
classification of multiple defects are required. Its integration into manufacturing environments.
AM quality control workflows is growing, particularly in
 IX. SUGGESTIONS FOR FUTURE WORK I extend my sincere appreciation to my research supervisor,
Mr. Hossein Rezaeifar, for his invaluable mentorship,
Despite recent progress, several areas remain ripe for further
constructive feedback, and continuous support throughout the
research and innovation in applying ML to AM. One major
course of this study. His expertise in cyber-physical systems and
opportunity lies in developing more generalized models that can advanced manufacturing has been instrumental in shaping the
adapt across different machines, materials, and process types. direction and depth of this work, ensuring its academic rigor and
Most existing models are tailored for specific setups, limiting relevance.
scalability and adoption.
I also acknowledge McMaster University for providing
access to essential research facilities, academic databases, and
A. Digital Twin Integration technical resources that greatly facilitated the comprehensive
Future research should focus on coupling ML with digital review of literature in the domains of machine learning and
twins of the AM process. Digital twins enable virtual additive manufacturing. The support from faculty members and
experimentation, rapid process optimization, and predictive administrative staff has played a significant role in overcoming
maintenance. When integrated with real-time sensor data, these challenges encountered during this research.
systems can significantly reduce trial-and-error and enhance Additionally, I am grateful to my colleagues and peers for
process reliability. their insightful discussions, collaborative spirit, and valuable
feedback. Their contributions have enriched the analysis and
interpretation of machine learning applications within the
B. Reinforcement Learning for Adaptive Control
additive manufacturing lifecycle, ultimately strengthening the
RL offers great potential for real-time adjustment of process overall quality and clarity of this article.
parameters. RL agents could continuously adapt scan paths,
laser intensities, or powder flow rates based on sensor feedback
to minimize defects dynamically. REFERENCES
[1] Baumgartl, H., Tomas, J., Buettner, R., & Merkel, M. (2020). A deep
C. Transfer Learning and Domain Adaptation learning-based model for defect detection in laser-powder bed fusion
using in-situ thermographic monitoring. Progress in Additive
These techniques allow models trained on one dataset or Manufacturing, 5(4), 309–318. https://doi.org/10.1007/s40964-020-
machine to be reused in a different context with minimal 00149-2
retraining. This is crucial for scaling ML solutions across [2] Chowdhury, S., & Anand, S. (2016). Artificial neural network based
multiple AM platforms and reducing data requirements. geometric compensation for thermal deformation in additive
manufacturing processes. Proceedings of the ASME 2016 International
Manufacturing Science and Engineering Conference, V003T08A006.
D. Multi-modal Sensor Fusion https://doi.org/10.1115/MSEC2016-8750
Combining data from thermal, visual, acoustic, and other [3] Fang, Q., Tan, Z., Li, H., Shen, S., Liu, S., Song, C., Zhou, X., Yang, Y.,
sensors can lead to richer insights. ML models that analyse this & Wen, S. (2021). In-situ capture of melt pool signature in selective laser
melting using U-Net-based convolutional neural network. Journal of
fused data can identify complex failure modes that would be Manufacturing Processes, 68, 347–355.
missed with a single modality. https://doi.org/10.1016/j.jmapro.2021.06.032
[4] Lee, J., & Prabhu, V. V. (2016). Simulation modeling for control of
E. Standardized Datasets and Benchmarks additive manufacturing processes. Journal of Manufacturing Science and
Engineering, 138(12). https://doi.org/10.1115/1.4033724
To support comparative research, there is a need for open-
[5] Seltzer, D. M., Wang, X., Nassar, A. R., Schiano, J. L., & Reutzel, E. W.
source datasets with standardized evaluation metrics. This (2015). System identification and feedback control for directed-energy,
would facilitate collaboration and accelerate innovation in the metal-based additive manufacturing. International Solid Freeform
field. Fabrication Symposium.
[6] Tapia, G., Khairallah, S., Matthews, M., King, W. E., & Elwany, A.
By addressing these challenges, future work can help fully (2018). Gaussian process-based surrogate modeling framework for
realize the potential of ML in AM. The goal is to create process planning in laser powder-bed fusion additive manufacturing of
intelligent, autonomous systems capable of delivering high- 316L stainless steel. The International Journal of Advanced
quality parts with minimal human intervention and maximal Manufacturing Technology, 94, 3591–3603.
https://doi.org/10.1007/s00170-017-0952-2
efficiency.
[7] Zhang, B., Jaiswal, P., Rai, R., Guerrier, P., & Baggs, G. (2019).
Convolutional neural network-based inspection of metal additive
X. ACKNOWLEDGEMENT manufacturing parts. Rapid Prototyping Journal, 25(3), 512–519.
https://doi.org/10.1108/RPJ-04-2018-0097
This review article was made possible through the support
and guidance of several individuals and institutions, to whom I
am deeply grateful.