Leveraging Machine Learning for Porosity Prediction in AM Using FDM for Pretrained Models and Process Development
Abstract
1. Introduction
2. Materials and Methods
2.1. Experimental 3D Printing Work
2.1.1. Printing Using FDM
- The 5 mm × 5 mm × 5 mm cubes: Referred to as ‘small’ in the rest of this study, it is the standard specimen size for defect assessment in metal additive manufacturing.
- The 10 mm × 10 mm × 10 mm cube: Referred to as ‘large’ in the rest of this study, it is printed under identical process parameters to evaluate dimensional scalability effects.
- Material: PLA was the only material explored through this investigation. It is widely used, recyclable, and does not require special storage conditions.
- Bed Temperature: The recommended value for PLA is 80 °C [40].
- Infill density: Fixed at 100%, since it is mainly for scaffolds and metamaterials with intentional voids.
- Printing orientation: Z-axis.
2.1.2. Test Coupons
2.2. A Convolutional Neural Network for Image Classification
2.2.1. Image Selection
2.2.2. Data Preparation
- Defective: Black, overly dark, or incomplete layers, where porosity analysis fails.
- Exploitable: Center of slice visible and well-formed, complete with closed contours, where porosity analysis succeeds.
2.2.3. Model Construction
2.3. Extraction of Porosity Metrics
2.3.1. Porosity Percentage
2.3.2. Porosity Distribution
- Elliptical voids: These elongated voids typically appear stretched along a primary axis. They arise when two parallel extruded filaments fail to fuse completely during deposition, leaving an interstitial gap aligned with the bead deposition path.
- Circular voids: These approximately symmetrical voids appear with similar dimensions along all axes. They most often occur in central cross-sections due to insufficient cohesion between adjacent material regions, interlayer bonding deficiencies, or incomplete merging of filament paths.
- Image Preprocessing:
- Converts the input RGB image to grayscale.
- Applies intensity thresholding to isolate porosity features.
- Performs contour extraction to precisely delineate void regions.
- Geometric Classification:
- Employs two distinct mathematical approaches to approximate extracted contours: Covariance matrix analysis combined with eigenvalue decomposition, and Principal Component Analysis (PCA).
- These methods collectively determine the major/minor axes and spatial orientation of each void.
- Calculates aspect ratios to classify voids as either elliptical or circular
- Data Output and Visualization:
- Generates comprehensive .csv files containing:
- ▪
- Source image identifiers.
- ▪
- Corresponding Z-height positions.
- ▪
- Complete listings of all the classified porosity features with their geometric parameters.
- Includes optional visualization modules that plot:
- ▪
- Original extracted contours.
- ▪
- Mathematically fitted ellipses and circles.
- ▪
- Comparative overlays showing approximation accuracy.
- Calculates IoU and plots results:
- ▪
- Shows how well the elliptical and circular voids capture the morphology of the void.
- ▪
- Plots indicators.
2.3.3. Repeatability Study
- Fabricated in separate print jobs (first/mid/last prints of the day).
- CT scanned individually at 30 µm resolution (1 h 15 min per scan).
- Reconstructed as 3D models using the scanner’s proprietary software.
2.4. A Multi-Layer Perceptron for Porosity Prediction
2.4.1. Data Preparation
2.4.2. Model Construction
- Data-Determined Parameters: Certain features were explicitly defined by the structure of the experimental data, such as the output layer—it consists of precisely one neuron, corresponding directly to the single output variable (porosity percentage).
- Empirically validated defaults: Other parameters were selected based on their established performance in similar regression applications:
- ▪
- ReLU (Rectified Linear Unit) as the activation function, effective in regression tasks.
- ▪
- Mean Squared Error (MSE) as the loss function, a standard for continuous value prediction tasks.
- Computationally optimized choices: The remaining configuration options were implemented to balance model performance with practical computational constraints:
- ▪
- Adam optimizer (η = 0.001) for efficient gradient-based weight updates.
- ▪
- A total of 1000 iterations specified as the initial training cycle count.
- ▪
- Random seed fixed at 42 to ensure reproducibility.
- Preprocessing:
- ▪
- Normalization/scaling of input features.
- ▪
- Categorical encoding of string variables (e.g., infill type).
- ▪
- Randomized training.
- Cross-Validation:
- ▪
- k-fold partitioning to assess generalization.
- ▪
- Iterative training on k-1 subsets with testing on held-out data.
- ▪
- Averaged metrics across all folds.
- Bias Mitigation:
- ▪
- Script-enforced random sampling.
- ▪
- Validation against parameter variability (scale/type/range).
2.5. Dimensional Scalability Investigation
3. Results
3.1. CNN Image Classifier
3.2. Porosity Analysis
3.3. Investigation on Repeatability
- Maximum porosity variation of 0.5% between specimens.
- Consistently low variance values, at 0.0001.
- Average porosity ranging from 1.85% to 2.36%.
- Specimen 11-1 exhibits incomplete layer formation.
- Specimen 11-2 lacks closed contours.
- Specimen 11-3 demonstrates ideal contour integrity.
3.4. Porosity Prediction
3.5. Scalability Investigation
4. Conclusions and Future Prospects
4.1. Key Findings
- Porosity Analysis: Development of a modular algorithmic framework capable of layer-specific porosity quantification (percentage, void location, and classification into ellipse- or circle-like geometries), offering insight into the anisotropic and heterogeneous nature of FDM components.
- Repeatability study: Establishment of a methodology to assess process variability under identical parameters, showing consistent porosity measurements within ±0.47% variation bounds.
- Porosity Prediction: Implementation of convolutional and perceptron networks, achieving over 90% classification accuracy for CNN, an R2 of 54.4% for MLP on small cubes, increasing to 77.6% on the large ones.
- Dimensional Scalability Effects: Component scale significantly influences porosity characteristics. Despite identical printing conditions, 10 mm × 10 mm × 10 mm specimens exhibited distribution patterns distinct from their 5 mm × 5 mm × 5 mm counterparts. Statistical analysis confirmed systematic differences, with an average correlation coefficient of 0.34, consistently extreme T-statistics (avg. 4.52) and p-values (avg. 0.15), thus confirming porosity distributions between small and large cubes are fundamentally different.
- Industrial Relevance: Development of practical predictive tools to reduce empirical parameter optimization and enable model transfer to higher-cost AM processes.
4.2. Limitations of the Present Research
- Dataset Diversity: The multiscale analysis was restricted to cubic geometries at micro/meso-scales, limiting direct extrapolation to full-scale components.
- Vertical porosity distribution: Exported slices were along the vertical direction. Including additional orthogonal orientations would strengthen porosity distribution assessment.
- Model Complexity: The CNN-MLP framework may not capture all porosity formation mechanisms under non-standard printing conditions or atypical defect morphologies.
- Validation scope: The repeatability assessment would benefit from a larger sample size to strengthen variation quantifications.
- Material Focus: While PLA was used to establish a cost-effective methodological pipeline, the framework is material-agnostic and should be validated on other polymers such as ABS and Nylon.
4.3. Recommendations for Future Work
- Expanding datasets to include more materials, geometries, and machine parameters, while incorporating geometric descriptors (e.g., volume, surface-area-to-volume ratio, and critical distances from edges) to improve model robustness and generalizability.
- Enhance ML capabilities to predict void morphologies and their distribution along the build height.
- Develop data-driven models to explore parameter interactions and establish phenomenological laws for cross-scale porosity prediction.
- Extend the framework to PBF processes, where datasets are more expensive to obtain, allowing exploration of the assumption that low-cost AM data can inform defect prediction in high-cost processes.
4.4. Industry Impact
- Parameter Optimization: Applying identified process windows to minimize porosity, with validation for specific material-machine combinations.
- ML System Integration: Deployment of the trained models as quality control modules within existing AM workflow softwares for continuous monitoring.
- Knowledge Transfer: Testing the transfer of porosity prediction between different FDM machines.
- Quality Assurance: The integration of ML models provides real-time predictive insights, allowing proactive parameter adjustments to enhance part reliability.
- Cost Efficiency: By replacing trial-and-error methods with data-driven approaches, production costs and time can be minimized.
- Methodological Innovation: The hybrid ML framework establishes a scalable template for quality assessment across AM technologies and materials.
- Cross-Sectoral Applicability: The findings demonstrated capabilities directly addressing stringent reliability requirements in aerospace and medical device manufacturing.
Author Contributions
Funding
Data Availability Statement
Acknowledgments
Conflicts of Interest
Abbreviations
BES | Bald Eagle Search |
PSO | Particle Swarm Optimization |
ML | Machine Learning |
AM | Additive Manufacturing |
FDM | Fused Deposition Modeling |
CNN | Convolutional Neural Network |
MLP | Multi-Layer Perceptron |
MSE | Mean Squared Error |
PCA | Principal Component Analysis |
ReLU | Rectified Linear Unit |
Appendix A
Feature | Value | Comment |
---|---|---|
Classes | 2 (Defective, Exploitable) | Binary classification task |
CNN Network’s Type | Sequential | Linear layer stack for straightforward image processing |
Convolutional Layers | 3 | Initial depth for feature extraction |
Pooling Layers | 3 | Downsampling to reduce spatial dimensions |
Fully Connected Layers | 2 | Final layers for classification |
Batch size | 10 | Balances memory usage and gradient stability |
Epoch | 15 | Initial training cycles |
Activation Function 1 | ReLU | For non-linearity |
Activation Function 2 | Sigmoid | For binary probabilities. |
Dataset size | 2282 | Number of images |
Network’s Features | Value | Comments |
---|---|---|
hidden_layer_sizes | (200, 100) | Two hidden layers with 200 and 100 neurons, respectively |
activation | ReLU | Best candidate for empirical success in regression |
output Layer | One neuron | Single output (porosity) |
optimizer | Adam | Stochastic gradient-based optimization |
learning_rate_init | 0.00597 | Initial learning rate for Adam optimizer |
alpha | 0.001835 | L2 Regularization |
batch_size | 32 | Batch size for stochastic optimization |
beta_1 | 0.99 | Exponential decay rate for first moment estimates |
beta_2 | 0.999 | Exponential decay rate for second moment estimates |
early_stopping | True | Validation-based early stopping |
n_iter_no_change | 20 | Number of iterations with no improvement to wait before stopping |
max_iter | 1000 | Maximum number of iterations |
Random seed | 42 | Reproducibility |
Total number of data | 1708 | Full dataset |
Epoch | 100 | Total number of iterations of all the training data in one cycle |
Network’s Features | Value | Comments |
---|---|---|
hidden_layer_sizes | (200, 100) | Two hidden layers with 200 and 100 neurons, respectively |
activation | ReLU | Best candidate for empirical success in regression |
output Layer | One neuron | Single output (porosity) |
optimizer | Adam | Stochastic gradient-based optimization |
learning_rate_init | 0.0066 | Initial learning rate for Adam optimizer |
alpha | 0.0069 | L2 Regularization |
batch_size | 32 | Batch size for stochastic optimization |
beta_1 | 0.95 | Exponential decay rate for first moment estimates |
beta_2 | 0.999 | Exponential decay rate for second moment estimates |
early_stopping | True | Validation-based early stopping |
n_iter_no_change | 20 | Number of iterations with no improvement to wait before stopping |
max_iter | 1000 | Maximum number of iterations |
Random seed | 42 | Reproducibility |
Total number of data | 3745 | Full dataset |
Epoch | 100 | Total number of iterations of all the training data in one cycle |
References
- Jonas, P.; Tomblin, J.; Lovingfoss, R.; Andrulonis, R.; Davies, C. Polymer Additive Manufacturing Materials, Machine, Processing and Quality Requirements Specification for ULTEM 9085 and Stratasys Inc. Fortus 900mc Machine. 2019. Available online: https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://www.wichita.edu/industry_and_defense/NIAR/Documents/NPS-89085-Rev-C-Process-Spec-March-12-2019.pdf&ved=2ahUKEwin1OPbnfWPAxUyj68BHUxKKe4QFnoECBoQAQ&usg=AOvVaw2tCjiR7PMBY-gqbaryI80J (accessed on 22 September 2025).
- Kechagias, J.; Zaoutsos, S. Effects of 3D-printing processing parameters on FFF parts’ porosity: Outlook and trends. Mater. Manuf. Process. 2024, 39, 804–814. [Google Scholar] [CrossRef]
- Kristiawan, R.B.; Imaduddin, F.; Ariawan, D.; Ubaidillah; Arifin, Z. A review on the fused deposition modeling (FDM) 3D printing: Filament processing, materials, and printing parameters. Open Eng. 2021, 11, 639–649. [Google Scholar] [CrossRef]
- Chacón, J.M.; Caminero, M.A.; García-Plaza, E.; Núñez, P.J. Additive manufacturing of PLA structures using fused deposition modelling: Effect of process parameters on mechanical properties and their optimal selection. Mater. Des. 2017, 124, 143–157. [Google Scholar] [CrossRef]
- Gebisa, A.; Lemu, H. Investigating Effects of Fused-Deposition Modeling (FDM) Processing Parameters on Flexural Properties of ULTEM 9085 using Designed Experiment. Materials 2018, 11, 500. [Google Scholar] [CrossRef]
- Durgun, I.; Ertan, R. Experimental investigation of FDM process for improvement of mechanical properties and production cost. Rapid Prototyp. J. 2014, 20, 228–235. [Google Scholar] [CrossRef]
- Rankouhi, B.; Javadpour, S.; Delfanian, F.; Letcher, T. Failure Analysis and Mechanical Characterization of 3D Printed ABS with Respect to Layer Thickness and Orientation. J. Fail. Anal. Prev. 2016, 16, 467–481. [Google Scholar] [CrossRef]
- Reindl, T.; Meisner, D.; Hierl, S. Benchmarking of Plastic-Based Additive Manufacturing Processes. RTe J. 2023, 1–20. [Google Scholar] [CrossRef]
- Kamaal, M.; Anas, M.; Rastogi, H.; Bhardwaj, N.; Rahaman, A. Effect of FDM process parameters on mechanical properties of 3D-printed carbon fibre–PLA composite. Prog. Addit. Manuf. 2021, 6, 63–69. [Google Scholar] [CrossRef]
- Ding, S.; Zou, B.; Wang, P.; Ding, H. Effects of nozzle temperature and building orientation on mechanical properties and microstructure of PEEK and PEI printed by 3D-FDM. Polym. Test. 2019, 78, 105948. [Google Scholar] [CrossRef]
- Sehhat, M.H.; Mahdianikhotbesara, A.; Yadegari, F. Impact of temperature and material variation on mechanical properties of parts fabricated with fused deposition modeling (FDM) additive manufacturing. Int. J. Adv. Manuf. Technol. 2022, 120, 4791–4801. [Google Scholar] [CrossRef]
- Zhou, T.; Song, Z.; Sundmacher, K. Big Data Creates New Opportunities for Materials Research: A Review on Methods and Applications of Machine Learning for Materials Design. Engineering 2019, 5, 1017–1026. [Google Scholar] [CrossRef]
- Paul, A.; Acar, P.; Liao, W.; Choudhary, A.; Sundararaghavan, V.; Agrawal, A. Microstructure optimization with constrained design objectives using machine learning-based feedback-aware data-generation. Comput. Mater. Sci. 2019, 160, 334–351. [Google Scholar] [CrossRef]
- Sharp, M.; Ak, R.; Hedberg, T. A survey of the advancing use and development of machine learning in smart manufacturing. J. Manuf. Syst. 2018, 48, 170–179. [Google Scholar] [CrossRef]
- Pilania, G.; Wang, C.; Jiang, X.; Rajasekaran, S.; Ramprasad, R. Accelerating materials property predictions using machine learning. Sci. Rep. 2013, 3, 2810. [Google Scholar] [CrossRef] [PubMed]
- Ward, L.; Agrawal, A.; Choudhary, A.; Wolverton, C. A general-purpose machine learning framework for predicting properties of inorganic materials. npj Comput. Mater. 2016, 2, 1–7. [Google Scholar] [CrossRef]
- Wang, J.; Ma, Y.; Zhang, L.; Gao, R.X.; Wu, D. Deep learning for smart manufacturing: Methods and applications. J. Manuf. Syst. 2018, 48, 144–156. [Google Scholar] [CrossRef]
- Wu, D.; Jennings, C.; Terpenny, J.; Gao, R.X.; Kumara, S. A Comparative Study on Machine Learning Algorithms for Smart Manufacturing: Tool Wear Prediction Using Random Forests. J. Manuf. Sci. Eng. 2017, 139, 71018. [Google Scholar] [CrossRef]
- Taguchi, G. Introduction to Quality Engineering: Designing Quality into Products and Processes; Quality Resources: Cleveland, OH, USA, 1986. [Google Scholar]
- Heidari-Rarani, M.; Ezati, N.; Sadeghi, P.; Badrossamay, M. Optimization of FDM process parameters for tensile properties of polylactic acid specimens using Taguchi design of experiment method. J. Thermoplast. Compos. Mater. 2022, 35, 2435–2452. [Google Scholar] [CrossRef]
- Anitha, R.; Arunachalam, S.; Radhakrishnan, P. Critical parameters influencing the quality of prototypes in fused deposition modelling. J. Mater. Process. Technol. 2001, 118, 385–388. [Google Scholar] [CrossRef]
- Wang, C.C.; Lin, T.; Hu, S. Optimizing the rapid prototyping process by integrating the Taguchi method with the Gray relational analysis. Rapid Prototyp. J. 2007, 13, 304–315. [Google Scholar] [CrossRef]
- Thrimurthulu, K.; Pandey, P.M.; Reddy, N.V. Optimum part deposition orientation in fused deposition modeling. Int. J. Mach. Tools Manuf. 2004, 44, 585–594. [Google Scholar] [CrossRef]
- Nancharaiah, T.; Raju, D.; Raju, V. An experimental investigation on surface quality and dimensional accuracy of FDM components. Int J Emerg Technol 2010, 1, 106–111. [Google Scholar]
- Akande, S. Dimensional Accuracy and Surface Finish Optimization of Fused Deposition Modelling Parts using Desirability Function Analysis. Int. J. Eng. Res. 2015, 4, 196–202. [Google Scholar] [CrossRef]
- Zhang, J.W.; Peng, A.H. Process-Parameter Optimization for Fused Deposition Modeling Based on Taguchi Method. Advanced Materials Research 2012, 538–541, 444–447. [Google Scholar] [CrossRef]
- Zhang, Y.; Chou, K. A parametric study of part distortions in fused deposition modelling using three-dimensional finite element analysis. Proc. Inst. Mech. Eng. Part B J. Eng. Manuf. 2008, 222, 959–968. [Google Scholar] [CrossRef]
- Sood, A.K.; Ohdar, R.K.; Mahapatra, S.S. Improving dimensional accuracy of Fused Deposition Modelling processed part using grey Taguchi method. Mater. Des. 2009, 30, 4243–4252. [Google Scholar] [CrossRef]
- Lee, B.H.; Abdullah, J.; Khan, Z. Optimization of Rapid Prototyping Parameters for Production of Flexible ABS Object. J. Mater. Process. Technol. 2005, 169, 54–61. [Google Scholar] [CrossRef]
- Torres, J.; Cotelo, J.; Karl, J.; Gordon, A.P. Mechanical Property Optimization of FDM PLA in Shear with Multiple Objectives. JOM 2015, 67, 1183–1193. [Google Scholar] [CrossRef]
- Garg, A.; Tai, K. An Ensemble Approach of Machine Learning in Evaluation of Mechanical Property of the Rapid Prototyping Fabricated Prototype. Appl. Mech. Mater. 2014, 575, 493–496. [Google Scholar] [CrossRef]
- Mohamed, O.A.; Masood, S.H.; Bhowmik, J.L. Mathematical modeling and FDM process parameters optimization using response surface methodology based on Q-optimal design. Appl. Math. Model. 2016, 40, 10052–10073. [Google Scholar] [CrossRef]
- Chandrashekarappa, M.P.G.; Chate, G.R.; Parashivamurthy, V.; Kumar, B.S.; Bandukwala, M.A.N.; Kaisar, A.; Giasin, K.; Pimenov, D.Y.; Wojciechowski, S. Analysis and Optimization of Dimensional Accuracy and Porosity of High Impact Polystyrene Material Printed by FDM Process: PSO, JAYA, Rao, and Bald Eagle Search Algorithms. Materials 2021, 14, 7479. [Google Scholar] [CrossRef]
- Equbal, A.; Equbal, M.; Sood, A.K.; Ravindrannair, P.; Equbal, M. A Review and Reflection on Part Quality Improvement of Fused Deposition Modelled Parts. IOP Conf. Ser. Mater. Sci. Eng. 2018, 455, 12072. [Google Scholar] [CrossRef]
- Al-Maharma, A.Y.; Patil, S.P.; Markert, B. Effects of porosity on the mechanical properties of additively manufactured components: A critical review. Mater. Res. Express 2020, 7, 122001. [Google Scholar] [CrossRef]
- Pak, P.M.-W.; Ogoke, F.; Polonsky, A.; Garland, A.; Bolintineanu, D.S.; Moser, D.R.; Heiden, M.J.; Farimani, A.B. ThermoPore: Predicting Part Porosity Based on Thermal Images Using Deep Learning. arXiv 2024, arXiv:2404.16882. [Google Scholar] [CrossRef]
- Mohammed, A.S.; Almutahhar, M.; Sattar, K.; AlHajeri, A.; Nazir, A.; Ali, U. Deep learning based porosity prediction for additively manufactured laser powder-bed fusion parts. J. Mater. Res. Technol. 2023, 27, 7330–7335. [Google Scholar] [CrossRef]
- Wang, L.; Zhu, S.-P.; Luo, C.; Niu, X.; He, J.-C. Defect-driven physics-informed neural network framework for fatigue life prediction of additively manufactured materials. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. 2023, 381, 20220386. [Google Scholar] [CrossRef]
- Goh, G.D.; Hamzah, N.M.B.; Yeong, W.Y. Anomaly Detection in Fused Filament Fabrication Using Machine Learning. 3D Print. Addit. Manuf. 2023, 10, 428–437. [Google Scholar] [CrossRef]
- Ultimaker. Technical Data Sheet: PLA. MatterHackers. Available online: https://www.matterhackers.com/downloads/AMIfv94gLougK1mY_0WwRUphoscP3graPFaxdGJFrDA0eh-kG-OqeFXz9tiU9QQNfIIB9zldfqy5IfV-oSHeoIEwcvui5_r3gDVHy5IIYGsHa52OJCSSlJrqnfNE0KYcdPJfo8_Ojh3chdeG0JALLK9s1UwzA3OAIclUQ8atpJDZXB1vaSOUe04.pdf (accessed on 23 September 2025).
- MatterHackers. Technical Data Sheet—Build Series PLA. MatterHackers. Available online: https://www.matterhackers.com/r/d0cX8m (accessed on 23 September 2025).
- MatterHackers. Pulse 3D Printer. Available online: https://www.matterhackers.com/store/l/pulse-3d-printer (accessed on 19 September 2025).
- Schmidhuber, J. Deep Learning in Neural Networks: An Overview. Neural Netw. 2015, 61, 85–117. [Google Scholar] [CrossRef]
- Pang, B.; Nijkamp, E.; Wu, Y.N. Deep Learning with TensorFlow: A Review. J. Educ. Behav. Stat. 2020, 45, 227–248. [Google Scholar] [CrossRef]
- Ketkar, N. Introduction to Keras in Deep Learning with Python: A Hands-on Introduction; Ketkar, N., Ed.; Apress: Berkeley, CA, USA, 2017; pp. 97–111. [Google Scholar] [CrossRef]
- Fisher, R.A. Statistical Methods for Research Workers. In Breakthroughs in Statistics; Kotz, S., Johnson, N.L., Eds.; Springer: New York, NY, USA, 1992; pp. 66–70. [Google Scholar] [CrossRef]
- Kohavi, R. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection. In Proceedings of the 14th International Joint Conference on Artificial Intelligence—Volume 2, Montreal, QC, Canada, 20–25 August 1995; ACM Digital Library: New York, NY, USA, 1995. Available online: https://dl.acm.org/doi/10.5555/1643031.1643047 (accessed on 30 October 2024).
- Vieweger, D.; Diel, S.; Schweiger, H.-G.; Tetzlaff, U. Mechanical Properties of Raw Filaments and Printed Specimens: Effects of Fiber Reinforcements and Process Parameters. Polymers 2024, 16, 1576. [Google Scholar] [CrossRef]
- Keleş, Ö.; Anderson, E.H.; Huynh, J.; Gelb, J.; Freund, J.; Karakoç, A. Stochastic fracture of additively manufactured porous composites. Sci. Rep. 2018, 8, 15437. [Google Scholar] [CrossRef] [PubMed]
- Ouajjani, K. Leveraging Machine Learning for Porosity Prediction in Additive Manufacturing: Case of Fused Deposition Modeling. Ph.D. Thesis, Wichita State University, Wichita, KS, USA, 2024. Available online: https://hdl.handle.net/10057/29169 (accessed on 19 September 2025).
- Clarkson, E. Stratasys Certified ULTEMTM 9085 Fortus 900mc Additively Manufactured Polymer Material Qualification Statistical Analysis Report; Wichita State University: Wichita, KS, USA, 2019. [Google Scholar]
- Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830. [Google Scholar]
- Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv 2015, arXiv:1409.1556. [Google Scholar] [CrossRef]
- Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; Fei-Fei, L. ImageNet: A large-scale hierarchical image database. In Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009; pp. 248–255. [Google Scholar] [CrossRef]
- MSFC-STD-3716; Standard for Additively Manufactured Spaceflight Hardware by Laser Powder Bed Fusion in Metals. Marshall Space Flight Center: Huntsville, Alabama, 2017. Available online: https://standards.nasa.gov/standard/MSFC/MSFC-STD-3716 (accessed on 17 September 2025).
- Federal Aviation Administration. AC 33.15-3—Powder Bed Fusion Additive Manufacturing Process for Aircraft Engine Parts; Federal Aviation Administration (FAA): Washington, DC, USA, 2023. Available online: https://www.faa.gov/regulations_policies/advisory_circulars/index.cfm/go/document.information/documentID/1041938 (accessed on 17 September 2025).
- ASTM F2971; Standard Practice for Reporting Data for Test Specimens Prepared by Additive Manufacturing. ASTM International: West Conshohocken, PA, USA, 2021. Available online: https://store.astm.org/f2971-13r21.html (accessed on 17 September 2025).
Parameter | Range |
---|---|
Infill Density [%] | 0–100% |
Printing Orientation | X–Y–Z–XY–YZ–XZ |
Bed Temperature [°C] | 80 °C–115 °C ± 5 °C |
Material | PLA–ABS–Nylon |
Print Temperature [°C] | 190–290 °C ± 5 °C |
Layer Height [µm] | 100–300 |
Infill Overlap [%] | 0–100% |
Printing Speed [mm/s] | 0–100 |
Infill Type | Grid–Gyroid–Concentric–Hexagon–Triangle–Lines |
Designation | Print T | Layer Height | Infill Overlap | Printing Speed | Infill Type |
---|---|---|---|---|---|
- | [°C] | [µm] | [%] | [mm/s] | - |
1-190-100-35-30-Grid | 190 | 100 | 35 | 30 | Grid |
2-190-200-65-70-Hexagon | 190 | 200 | 65 | 70 | Hexagon |
3-190-300-90-100-Lines | 190 | 300 | 90 | 100 | Lines |
4-200-100-65-70-Hexagon | 200 | 100 | 65 | 70 | Hexagon |
5-200-200-90-100-Lines | 200 | 200 | 90 | 100 | Lines |
6-200-300-35-30-Grid | 200 | 300 | 35 | 30 | Grid |
7-210-100-90-100-Lines | 210 | 100 | 90 | 100 | Lines |
8-210-200-35-30-Grid | 210 | 200 | 35 | 30 | Grid |
9-210-300-65-70-Hexagon | 210 | 300 | 65 | 70 | Hexagon |
10-210-300-30-35-Grid | 210 | 300 | 30 | 35 | Grid |
11-210-300-90-100-Lines * | 210 | 300 | 90 | 100 | Lines |
Scanning Features | Value |
---|---|
Reconstructed pixel pitch | 0.02991 mm |
Slice spacing | 29.91 µm |
Pixel size X, Y, Z | 29.91 µm/pixel |
Voxel size | 29.91 µm × 29.91 µm × 29.91 µm |
Designation | Number of Slices [Small] | Number of Slices [Large] |
---|---|---|
1-190-100-35-30-Grid | 215 | 382 |
2-190-200-65-70-Hexagon | 216 | 376 |
3-190-300-90-100-Lines | 211 | 397 |
4-200-100-65-70-Hexagon | 229 | 369 |
5-200-200-90-100-Lines | 197 | 384 |
6-200-300-35-30-Grid | 181 | 392 |
7-210-100-90-100-Lines | 211 | 372 |
8-210-200-35-30-Grid | 198 | 416 |
9-210-300-65-70-Hexagon | 213 | 371 |
10-210-300-30-35-Grid | 213 | 375 |
11-210-300-90-100-Lines | 198 | 367 |
Designation | N. of Slices | N. of Exploitable Slices | N. of Defective Slices | Percentage of Defective Slices |
---|---|---|---|---|
- | - | - | - | [%] |
1-190-100-35-30-Grid | 215 | 162 | 53 | 24.65% |
2-190-200-65-70-Hexagon | 216 | 157 | 59 | 27.31% |
3-190-300-90-100-Lines | 211 | 162 | 49 | 23.22% |
4-200-100-65-70-Hexagon | 229 | 160 | 69 | 30.13% |
5-200-200-90-100-Lines | 197 | 127 | 70 | 35.53% |
6-200-300-35-30-Grid | 181 | 162 | 19 | 10.50% |
7-210-100-90-100-Lines | 211 | 161 | 50 | 23.70% |
8-210-200-35-30-Grid | 198 | 157 | 41 | 20.71% |
9-210-300-65-70-Hexagon | 213 | 147 | 66 | 30.99% |
10-210-300-30-35-Grid | 213 | 160 | 53 | 24.88% |
11-210-300-90-100-Lines | 198 | 153 | 45 | 22.73% |
ID | - | 11-1 | 11-2 | 11-3 |
---|---|---|---|---|
Print Temp | [°C] | 210 | ||
Layer height | [µm] | 300 | ||
Infill Overlap | [%] | 90 | ||
Printing Speed | [mm/s] | 100 | ||
Infill Type | - | Lines | ||
x-length * | [mm] | 5.65 | 5.6 | 5.29 |
y-length * | [mm] | 5.62 | 5.77 | 5.62 |
z-length * | [mm] | 5.4 | 5.13 | 5.22 |
Printing Time | - | Mid-Print of the day | 1st print of the day | Last print of the day |
Input Data | Output |
---|---|
Print Temperature Layer Height Infill Overlap Printing Speed Infill Type Z-Height | Porosity |
Cube ID | Input | Output | |||||
---|---|---|---|---|---|---|---|
Print Temperature | Layer Height | Infill Overlap | Printing Speed | Infill Type | Z-Height | Porosity | |
- | [°C] | [µm] | [%] | [mm/s] | - | [µm] | [%] |
1 | 190 | 100 | 35 | 30 | Grid | 135 | 1.64 |
1 | 190 | 100 | 35 | 30 | Grid | 216 | 1.92 |
2 | 190 | 200 | 65 | 70 | Hexagon | 3564 | 0.01 |
2 | 190 | 200 | 65 | 70 | Hexagon | 3591 | 0.03 |
10 | 210 | 300 | 30 | 35 | Grid | 4536 | 4.27 |
10 | 210 | 300 | 30 | 35 | Grid | 4563 | 4.64 |
7 | 210 | 100 | 90 | 100 | Lines | 3780 | 0.01 |
7 | 210 | 100 | 90 | 100 | Lines | 3807 | 0.01 |
6 | 200 | 300 | 35 | 30 | Grid | 3321 | 2.03 |
6 | 200 | 300 | 35 | 30 | Grid | 3348 | 2.26 |
Cube ID | Input | Output | |||||
---|---|---|---|---|---|---|---|
Print Temperature | Layer Height | Infill Overlap | Printing Speed | Infill Type | Z-Height | Porosity | |
- | [°C] | [µm] | [%] | [mm/s] | - | [µm] | [%] |
1 | 190 | 100 | 35 | 30 | Grid | 870 | 6.38 |
2 | 190 | 200 | 65 | 70 | Hexagon | 3540 | 1.05 |
2 | 190 | 200 | 65 | 70 | Hexagon | 3570 | 1.17 |
10 | 210 | 300 | 30 | 35 | Grid | 6960 | 0.95 |
10 | 210 | 300 | 30 | 35 | Grid | 6990 | 1.36 |
7 | 210 | 100 | 90 | 100 | Lines | 7410 | 0.00 |
7 | 210 | 100 | 90 | 100 | Lines | 7440 | 0.00 |
6 | 200 | 300 | 35 | 30 | Grid | 2580 | 1.84 |
6 | 200 | 300 | 35 | 30 | Grid | 2610 | 2.64 |
Performance Metric | Defective Class | Exploitable Class | Total |
---|---|---|---|
Mean ± Std | Mean ± Std | Mean ± Std | |
Precision | 0.977 ± 0.026 | 0.970 ± 0.021 | - |
Recall | 0.925 ± 0.052 | 0.992 ± 0.009 | - |
F1-Score | 0.949 ± 0.026 | 0.981 ± 0.010 | - |
Accuracy | - | - | 0.972 ± 0.014 |
Designation | Predicted Exploitable | Actual Exploitable Slices | Predicted Defective | Actual Defective |
---|---|---|---|---|
1-190-100-35-30-Grid | 328 | 325 | 54 | 57 |
2-190-200-65-70-Hexagon | 325 | 319 | 51 | 57 |
3-190-300-90-100-Lines | 325 | 311 | 72 | 86 |
4-200-100-65-70-Hexagon | 328 | 322 | 41 | 47 |
5-200-200-90-100-Lines | 330 | 326 | 54 | 58 |
6-200-300-35-30-Grid | 320 | 318 | 72 | 74 |
7-210-100-90-100-Lines | 329 | 319 | 43 | 53 |
8-210-200-35-30-Grid | 327 | 325 | 89 | 91 |
9-210-300-65-70-Hexagon | 317 | 303 | 54 | 68 |
10-210-300-30-35-Grid | 326 | 322 | 49 | 53 |
11-210-300-90-100-Lines | 324 | 300 | 43 | 67 |
Cube ID | 11-1 | 11-2 | 11-3 |
---|---|---|---|
Number of generated images | 198 | 205 | 228 |
Exploitable images | 153 | 152 | 157 |
First height for porosity from exploitable cross-section [µm] | 540 | 432 | 297 |
Last height for porosity from exploitable cross-section [µm] | 4644 | 4509 | 4509 |
Min porosity [%] | 0.33% | 0.77% | 0.12% |
Max porosity [%] | 5.55% | 6.63% | 5.30% |
Avg. porosity [%] | 2.22% | 2.36% | 1.85% |
Range | 0.0522 | 0.0586 | 0.0518 |
Variance | 0.0001 | 0.0001 | 0.0001 |
Metrics | Value |
---|---|
Mean standard deviation | 0.0047 |
Maximum standard deviation | 0.0149 |
Relative uncertainty (std/mean) | 21.55% |
RMS difference between datasets 1 and 2 | 0.0069 |
RMS difference between datasets 1 and 3 | 0.0087 |
RMS difference between datasets 2 and 3 | 0.0066 |
Average pairwise difference | 0.0074 |
Specimen | N of Data Points | Small Cubes | Large Cubes | Comparison | Correlation * | |||
---|---|---|---|---|---|---|---|---|
Mean | Standard Deviation | Mean | Standard Deviation | Mean | Standard Deviation | |||
1-190-100-35-30-G | 162 | 0.48 | 0.25 | 0.21 | 0.23 | 0.56 | 0.08 | −0.01 |
2-190-200-65-70-H | 157 | 0.12 | 0.18 | 0.49 | 0.21 | −3.08 | −0.17 | −0.18 |
3-190-300-90-100-L | 161 | 0.29 | 0.25 | 0.06 | 0.13 | 0.79 | 0.48 | 0.08 |
4-200-100-65-70-H | 160 | 0.09 | 0.14 | 0.08 | 0.11 | 0.11 | 0.21 | 0.80 |
5-200-200-90-100-L | 127 | 0.15 | 0.19 | 0.02 | 0.10 | 0.87 | 0.47 | −0.12 |
6-200-300-35-30-G | 162 | 0.18 | 0.19 | 0.22 | 0.17 | −0.22 | 0.11 | 0.61 |
7-210-100-90-100-L | 161 | 0.10 | 0.26 | 0.03 | 0.14 | 0.70 | 0.46 | 0.06 |
8-210-200-35-30-G | 157 | 0.19 | 0.17 | 0.10 | 0.15 | 0.47 | 0.12 | 0.58 |
9-210-300-65-70-H | 160 | 0.26 | 0.19 | 0.11 | 0.17 | 0.58 | 0.11 | −0.06 |
10-210-300-30-35-G | 147 | 0.24 | 0.17 | 0.09 | 0.15 | 0.63 | 0.12 | 0.68 |
11-210-300-90-100-L | 153 | 0.36 | 0.22 | 0.06 | 0.16 | 0.83 | 0.27 | 0.43 |
Min Value | 127 | 0.09 | 0.14 | 0.02 | 0.10 | −3.08 | −0.17 | −0.18 |
Max Value | 162 | 0.48 | 0.26 | 0.49 | 0.23 | 0.87 | 0.48 | 0.80 |
Average Value | 155.18 | 0.22 | 0.20 | 0.13 | 0.16 | 0.20 | 0.21 | 0.34 |
Specimen | N of Datapoints | T-Statistic | p-Value | Area Between Curves |
---|---|---|---|---|
190-100-35-30-Grid | 162 | 9.758 | 0.000 | 0.36 |
190-200-65-70-Hex | 157 | −16.701 | 0.000 | 0.42 |
190-300-90-100-Lines | 161 | 9.971 | 0.000 | 0.27 |
200-100-65-70-Hex | 160 | 0.210 | 0.834 | 0.05 |
200-200-90-100-Lines | 127 | 7.230 | 0.000 | 0.16 |
200-300-35-30-Grid | 162 | −2.203 | 0.028 | 0.13 |
210-100-90-100-Lines | 161 | 3.257 | 0.001 | 0.11 |
210-200-35-30-Grid | 157 | 5.274 | 0.000 | 0.12 |
210-300-65-70-Hex | 160 | 7.576 | 0.000 | 0.23 |
210-300-30-35-Grid | 147 | 8.473 | 0.000 | 0.17 |
210-300-90-100-Lines | 153 | 13.303 | 0.000 | 0.30 |
Min Value | 127 | −16.70 | 0.00 | 0.05 |
Max Value | 162 | 13.30 | 0.83 | 0.42 |
Average Value | 155.18 | 4.52 | 0.15 | 0.18 |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Ouajjani, K.; Steck, J.E.; Olivares, G. Leveraging Machine Learning for Porosity Prediction in AM Using FDM for Pretrained Models and Process Development. Materials 2025, 18, 4499. https://doi.org/10.3390/ma18194499
Ouajjani K, Steck JE, Olivares G. Leveraging Machine Learning for Porosity Prediction in AM Using FDM for Pretrained Models and Process Development. Materials. 2025; 18(19):4499. https://doi.org/10.3390/ma18194499
Chicago/Turabian StyleOuajjani, Khadija, James E. Steck, and Gerardo Olivares. 2025. "Leveraging Machine Learning for Porosity Prediction in AM Using FDM for Pretrained Models and Process Development" Materials 18, no. 19: 4499. https://doi.org/10.3390/ma18194499
APA StyleOuajjani, K., Steck, J. E., & Olivares, G. (2025). Leveraging Machine Learning for Porosity Prediction in AM Using FDM for Pretrained Models and Process Development. Materials, 18(19), 4499. https://doi.org/10.3390/ma18194499