Additive Manufacturing (AM)

1,2 Guo Dong Goh¹, Xi Huang ¹and Wai Yee Yeong

1 School of Mechanical and Aerospace Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798, Singapore 2 Singapore Centre for 3D Printing, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798, Singapore E-mail: wyyeong@ntu.edu.sg

Status

Additive manufacturing, commonly known as 3D printing, has matured from a prototyping tool into a viable production technology across industries ranging from aerospace and biomedical to electronics and construction. By building objects layer-by-layer, AM enables the fabrication of complex geometries and customized, functionally graded parts with minimal material waste. Ensuring reliability and consistency of printed parts, however, remains a critical concern as defects or process variations can compromise mechanical properties and impede AM’s adoption for end-use, safety-critical components (1). In recent years, the convergence of AM with machine learning (ML) has been increasingly viewed as a key enabler for smart manufacturing, addressing these challenges by extracting insights from data and automating decision-making.

Machine learning algorithms excel at recognizing complex patterns in large datasets, and in AM they are being leveraged to unravel the intricate relationships between process parameters, material behavior, and part quality (2). Early successes of ML in AM have been demonstrated across the workflow: in design (e.g., ML-driven topology optimization and generative design for lightweight structures), in materials development (predicting formulations or microstructures to achieve desired properties), in process optimization (tuning print parameters for quality and efficiency), and in in situ monitoring for defect detection (Figure 1).

Current and future challenges

Despite the enthusiasm, several key challenges must be addressed to fully realize ML’s potential in AM. Data acquisition and quality is a foundational hurdle: ML models require large, high-quality datasets, yet AM experiments are time-consuming and sensors can be costly, making data scarce or siloed. Printing conditions vary widely between machines and materials, and there is a lack of standardized data formats and sharing mechanisms across the industry. As a result, models trained on one dataset may struggle to generalize. For instance, an ML model for defect detection might need thousands of labeled images covering different defect types, build geometries, and lighting conditions – data that is often unavailable or expensive to obtain. In metal AM, researchers noted the difficulty of obtaining ground-truth defect data for training computer vision models; high-speed optical cameras capture only surface phenomena, missing subsurface defects, and ex- situ X-ray CT scans are hard to align with the images (3). This highlights a broader sensor and labeling challenge: how to efficiently acquire rich, synchronized data (vision, thermal, acoustic, etc.) and accurate labels (defect locations, material properties) during the 3D printing process.

Figure 1. The figure provides a summary of how ML is incorporated into AM workflows. On the left, ML

approaches are categorized into supervised, unsupervised, semi-supervised, and reinforcement learning, with a note on the rising use of transformer architectures. The right side outlines different AM technologies. At the center, it highlights the advantages ML brings to AM, and along the bottom, it lists practical implementations spanning industries—from aerospace and defense to electronics and food—demonstrating the broad impact of combining ML with advanced manufacturing (2). Reproduced under Creative Commons CC BY license.

Another major challenge is the generalizability of ML models in AM. A model trained for one printer or material often underperforms when applied to a different setup due to variations in machine hardware, calibration, or process dynamics. Adapting an ML-based process optimizer or quality predictor to a new AM

machine typically requires laborious data collection and retraining for that specific context. This hampers scalability in production environments where fleets of printers or new machine models are introduced. Techniques like transfer learning and domain adaptation are being explored (4), but ensuring robust cross-machine performance remains non-trivial. Likewise, scalability and real-time implementation pose challenges: embedding ML into the real-time control loop of a printer demands fast inference speeds and reliable hardware/software integration. Many deep learning models are computationally intensive, which could slow down fabrication if not optimized. For example, a complex neural network might detect defects accurately but could become a bottleneck if it cannot run at the printer’s frame rate for live monitoring, especially in cases whether high frame rate is required such as melt pool monitoring in powder bed fusion technique (5). Achieving millisecond-level response times may require model compression, edge computing devices, or dedicated accelerators, all of which increase system complexity. There are also practical deployment challenges. The stochastic nature of some AM processes (e.g., powder bed fusion spatter, filament feed variability) means ML models must handle noisy, high-dimensional data and rare events. Ensuring that models not only detect anomalies but also make reliable corrective decisions without human intervention is a frontier that involves risk: a mistimed or incorrect correction could itself cause a failure. Moreover, the black-box nature of many ML algorithms can reduce user trust in critical manufacturing settings. Engineers and certifying agencies may be wary of decisions made by opaque models, highlighting the need for explainable AI and rigorous validation standards (6). Qualification and certification of ML-augmented AM processes is largely uncharted territory – there is a lack of standards on how to approve parts made with ML-driven parameter adjustments or defect correction. Finally, organizational and skill barriers exist; implementing these advanced systems requires interdisciplinary expertise (materials, ML, software) that manufacturing teams are still building.

Advances in science and technology to meet challenges

Research efforts are actively advancing the state of the art to address the above challenges, yielding promising results on several fronts. One significant area of progress is in real-time defect detection and correction during printing. For extrusion-based 3D printing, computer vision models have been developed to automatically detect print anomalies such as filament under-extrusion or over-extrusion and intervene mid-build (7). Brion & Pattinson address the need for a truly generalizable error-correction system (1). They built a multi-head neural network trained on 1.2 million automatically labeled images spanning 192 parts, multiple geometries, materials, printers, and toolpaths. By labeling deviations from optimal printing parameters during acquisition, they created a diverse dataset that lets the network detect and correct errors in real time across different extrusion methods. Their control loop not only corrects defects but also provides visualizations of its decision process, enhancing transparency and applicability across varied AM setups.

Another domain of notable progress is multi-objective process optimization using ML, which tackles the challenge of balancing competing quality metrics without exhaustive trial-and-error. Traditional process tuning in AM often involves iterative experiments to achieve a trade-off (e.g., maximizing strength while minimizing porosity). ML-driven surrogate models and optimization algorithms can accelerate this search. Researchers used ML-driven surrogate models to optimize intense pulsed light sintering for aerosol-jet printed nanoink films, balancing film electrical resistance and surface roughness—factors that traditionally trade off (8). Training on a small experimental dataset, their multi-objective algorithm identified process settings yielding both low sheet resistance and low roughness, revealing an optimal window that manual tuning would likely miss. This approach demonstrates how ML can navigate complex AM trade-offs and improve material performance without new hardware. Similarly, Bayesian optimization and reinforcement learning schemes are being explored to tune dozens of AM process parameters simultaneously, accelerating

process qualification. For instance, transfer-learning-based frameworks have been able to predict optimal laser processing parameters for new machines using knowledge from prior machines, reducing the effort needed when adopting a new printer model (9). These advances point to a future where “self-optimizing” printers automatically adjust to achieve target outcomes.

In the realm of materials and properties, ML techniques are enabling breakthroughs in achieving application-specific material performance via AM. For instance, researchers trained a neural network on 216 PolyJet-printed samples mixing hard and soft photopolymers to predict Shore hardness and elastic modulus with <1% error—outperforming response surface models (10). By inverting this model, they could specify a desired tissue stiffness and directly obtain the needed material ratios and layer structure. This enables patient-specific anatomical models or prosthetics with tunable tactile properties that trial and error cannot achieve. More broadly, ML is accelerating materials development for AM by identifying complex process– structure–property linkages: for instance, in bioelectronics and bioprinting, where living cells or soft polymers are printed, data-driven models have helped in discovering printable bio-ink formulations and in calibrating process parameters to ensure viability and performance of printed tissues. In electronics printing, ML has been used to predict how printing parameters affect conductivity and to adjust them to produce functional circuits with minimal defects (11). These case studies underscore that by learning from experimental data, ML algorithms can navigate the enormous design space of multi-material and functional printing to meet specific targets.

Researchers are tackling generalizability by combining physics-informed neural networks and digital twin simulations with empirical ML to ground models in physical reality. In metal powder bed fusion, for instance, pore-detection accuracy rose to 87% by augmenting limited experimental data with high-fidelity melt-pool simulations (12). Emerging architectures like transformers are also under exploration for their ability to model sequential, high-dimensional AM data and catch subtle defects (13). Meanwhile, initiatives such as the NIST Additive Manufacturing Material Database are building open benchmarks—compiling build logs, in situ sensor readings, and quality metrics—to spur development of more generalizable AM ML models (14). Together, advances in sensing, data augmentation, algorithm efficiency, and hybrid modeling are transforming AM from a manual, experience-driven practice to a data-driven, adaptive process, building confidence that ML integration will overcome current limitations and unlock higher automation and performance.

In summary, advances in sensing, data augmentation, algorithm efficiency, and hybrid modeling are jointly pushing the boundaries: what was once a manual, experience-driven practice is evolving into a data-driven, adaptive process. With each demonstrated success – from real-time correction systems to predictive material tuning – confidence grows that the integration of ML will resolve many of AM’s current limitations and unlock higher levels of automation and performance.

Concluding remarks

Machine learning will transform additive manufacturing into a smart, data-driven paradigm. By enabling smarter design, self-optimizing parameters, and autonomous quality control, ML makes production more reliable and efficient. High-quality process data is as essential as hardware for scaling AM to industry. Although challenges remain in data sharing, model transferability, and real-time deployment, ongoing advances—bridging simulation and experiment, standardizing data formats, and developing validation protocols—are paving the way. Future AM systems will continuously learn from each build, reducing errors, improving yield, and expanding design possibilities. Integrating ML with AM thus provides the precision and flexibility needed for agile factories capable of producing complex, customized products with minimal human intervention.

Acknowledgements

The authors acknowledge the support of National Research Foundation for NRF Investigatorship Award No.: NRF-NRFI07-2021-0007.

The research is supported by the National Research Foundation, Prime Minister’s Office, Singapore under its Medium-Sized Centre funding scheme.

References

[1] Brion D A J and Pattinson S W 2022 Generalisable 3D printing error detection and correction via multi-head neural networks *Nat. Commun.*134654 (doi: 10.1038/s41467-022-31985-y) [2] Ng W L, Goh G L, Goh G D, Ten J S J and Yeong W Y 2024 Progress and opportunities for machine learning in materials and processes of additive manufacturing *Adv. Mater.*362310006 (doi: 10.1002/adma.202310006) [3] Bonato N, Zanini F and Carmignato S 2023 Deformations modelling of metal additively manufactured parts and improved comparison of in-process monitoring and post-process X-ray computed tomography *Addit. Manuf.*75103736 (doi:

10.1016/j.addma.2023.103736) [4] Zhu J, Su Z, Wang Q, Lan Z, Chan F S, Han Z *et al.*2024 Surface quality prediction and quantitative evaluation of process parameter effects for 3D printing with transfer learning-enhanced gradient-boosting decision trees *Expert Syst. Appl.*237 121478 (doi: 10.1016/j.eswa.2023.121478) [5] Ma H, Mao Z, Feng W, Yang Y, Hao C, Zhou J *et al.*2022 Online in-situ monitoring of melt pool characteristic based on a single high-speed camera in laser powder bed fusion process *Appl. Therm. Eng.*211118515 (doi:

10.1016/j.applthermaleng.2022.118515) [6] Ukwaththa J, Herath S and Meddage D P P 2024 A review of machine learning (ML) and explainable artificial intelligence (XAI) methods in additive manufacturing (3D printing) *Mater. Today Commun.*41110294 (doi: 10.1016/j.mtcomm.2024.110294) [7] Goh G D, Hamzah N M B and Yeong W Y 2023 Anomaly detection in fused filament fabrication using machine learning 3D Print. *Addit. Manuf.*10428–437(doi: 10.1089/3dp.2021.023) [8] Goh G L, Zhang H, Goh G D, Yeong W Y and Chong T H 2022 Multi-objective optimization of intense pulsed light sintering process for aerosol jet printed thin film *Mater. Sci. Addit. Manuf.*110–10 (doi: 10.36922/msam.26) [9] Pashmforoush F and Seyedzavvar M 2023 A transfer learning-based machine learning approach to predict mechanical properties of different material types fabricated by selective laser melting process Proc. Inst. Mech. Eng. E 238 09544089231215683 (doi: 10.1177/09544089231215683) [10] Goh G D, Sing S L, Lim Y F, Thong J L J, Peh Z K, Mogali S R *et al.*2021 Machine learning for 3D printed multi-materials tissue-mimicking anatomical models *Mater. Des.*211110125 (doi: 10.1016/j.matdes.2021.110125) [11] Abbasi Shirsavar M, Taghavimehr M, Ouedraogo L J, Javaheripi M, Hashemi N N, Koushanfar F *et al.*2022 Machine learning-assisted E-jet printing for manufacturing of organic flexible electronics *Biosens. Bioelectron.*212114418 (doi:

10.1016/j.bios.2022.114418) [12] Ren Z, Gao L, Clark S J, Fezzaa K, Shevchenko P, Choi A *et al.*2023 Machine learning–aided real-time detection of keyhole pore generation in laser powder bed fusion Science 37989–94 (doi: 10.1126/science.add4667) [13] Wang X Q, Jin Z, Zheng B and Gu G X 2025 Transformer-based approach for printing quality recognition in fused filament fabrication *npj Adv. Manuf.*215 (doi: 10.1038/s44334-025-00025-0) [14] Lane B 2020 Process Monitoring Dataset from the Additive Manufacturing Metrology Testbed (AMMT): Overhang Part X4 Natl. Inst. Stand. Technol. Data 2020