Machine Learning in Laser-based Manufacturing
Yung C Shin
Mechanical Engineering, Purdue University, West Lafayette, Indiana, U.S.A.
E-mail: shin@purdue.edu
Status
Machine learning (ML) has been finding increasing adoption in various areas of laser-based manufacturing, such as in predictive modelling, process monitoring, process control, defect detection, prediction of microstructure and mechanical properties, and process parameter optimization. Laser-based manufacturing processes such as laser welding, additive manufacturing and laser cutting involve complex physical mechanisms: including, but not limited to, laser energy absorption, heat transfer, melting, fluid flow, evaporation, solidification, etc. Achieving optimal operating conditions to get the desired mechanical properties and microstructure often involves an extensive amount of experiments with the variation of operating parameters or multi-physics numerical simulations that incur high computational costs and time. As industry is striving to reduce the lead time and the cost of implementing laser processing, machine learning has emerged as a promising approach to establishing data-driven or surrogate models that can significantly reduce the high cost of iteratively finding cause-effect relationships or that can replace the prohibitively computationally expensive physics-based high fidelity modelling in some cases [1,2,3]. In recent years, one can find many examples of using machine learning for process monitoring, particularly with the use of a vision sensor to detect molten pool boundaries [4], surface defects [5], incomplete welds and cuts [6], keyhole depth [7], etc. It has served as a useful tool for automatic process control due to its ability to predict the process condition in real time [8], once developed. Machine learning can also be useful for tuning process parameters or process optimization based on the generated data [9]. It has also been used for predicting the resultant microstructure and hardness after laser processing [3,10]. In addition, some successful efforts have been made to synthesize new materials via machine learning by using additive manufacturing processes. For example, attempts have been made to predict thermodynamically stable phases in high entropy alloys [11,12]. As evidenced by these examples, it is undeniable that the role and use of machine learning will only be increasing as the scientific field of machine learning further advances. In some sense, machining learning might be the only way of realizing predictive science for the optimization, process control and robust implementation of many laser processes in material processing, because the Moore’s law indicates that it will take at least another two decades until the computational capabilities, even with massive parallel processing, catch up with the computational speed needed for high fidelity modelling that can be used for real time design, optimization and control.
Current and future challenges
Despite the rapidly increasing adoption of machine learning in various applications of laser processing, much of the current machine learning requires an extensive amount of data, which can be very expensive to generate from experiments with physical systems. Furthermore, data-driven models are often applicable only to the specific setup or operation used for the development of the data-driven model, thus lacking the generalization capability to a wide range of process conditions, unlike physics-based predictive models. For example, a data-driven model developed for a particular type of laser and workpiece material may not be readily extendable to another set of laser and material combinations. This will require establishing separate data-driven models for each combination of laser and material. In order to expand its general applicability and reduce the cost of generating a lot of data, more efficient methods of establishing machine learning models would be desirable. For example, physics-informed machine learning would be a promising approach to achieving this goal by integrating well-known physical laws or governing equations that have been developed over the last several decades through extensive scientific research. This will result in a drastic reduction in the amount of data needed to establish a data-driven model and is likely to expand the generalization capability of machine learning models. Another issue lies in how to utilize the existing data, often scattered, albeit abundant. For many of the laser processes for commonly used laser-material combinations, there have been a lot of data generated over the years, but they cannot be easily utilized for constructing a data-driven model since they exist in various formats, sizes, images and resolutions.
Therefore, the community may need to work on establishing the standard for data format or data repositories so that they can be used for developing data-driven models by machine learning. Another challenge is how to combine different types of heterogeneous machine learning models for system-level monitoring, control or optimization. For each laser process, an integrated frame for process monitoring, quantification, and control might be needed. Figure 1 illustrates a possible approach to an integrated quality inspection, process monitoring and feedback control for laser additive manufacturing processes.
Figure 1. Illustration of a machine learning-based process monitoring and control system for a laser
additive manufacturing system
Advances in science and technology to meet challenges
Many learning methods have been developed over the years, which can be applied to various aspects of laser processing of materials. In finding optimal process parameters, machine learning techniques such as Bayesian optimization, random forests, and various paradigms of artificial neural networks have been utilized. Convolution neural networks, Long Short-Term Memories (LSTMs) and Kalman filters with ML enhancement were often the choices for melt pool monitoring and control. Various convolution neural networks have been popular for the application to defect detections during laser processing with vision systems, x-ray scans, ultrasound scans or scanning electron microscope (SEM) images. People have tried to develop surrogate models of complex physical problems via various neural fuzzy models and physics-informed neural networks. Continuing this success, people need to evaluate a wider range of machine learning models for each application so that the best approaches can be established. The community also needs to work on integrated machine learning models for system level optimization and control. The laser processing community can also piggyback on the rapid advances in artificial intelligence (AI) and machine learning, as more advanced theories and methods are introduced. They also need to pay attention to new types of sensors and sensing techniques that can expand the ML-based process monitoring and diagnostics. Commonly used sensors are cameras, infrared sensors, acoustic emission sensors, photodiodes, spectrometers, etc., while in-situ x-ray devices have also been successfully used for monitoring of molten pool, spattering, etc. These sensors must be easily integrated into commercial laser processing equipment,
and provide the requisite speed and resolutions as some of the laser processes, such as laser powderbed fusion and laser welding, are performed at very high speeds.
Concluding remarks
As described above, machine learning has a very promising future in various laser-based manufacturing processes for process monitoring, control, part quality monitoring and optimization. However, various challenges mentioned in this article must be overcome for a wide use of machine learning in industry, and further advancements in the requisite sensing techniques and sensors must follow. The community needs to work together to establish standards in data formats and repositories so that efforts are not fragmented.
Acknowledgements
The author wishes to thank many of his former and current students who have contributed to the generation of concepts, advancement in theories and applications of machine learning to many manufacturing processes, which have been used for generating this article.
References
[1] Tapia, G., Khairallah, S., Matthews, M., King, W.E., Elwany, A. Gaussian process-based surrogate modeling framework for process planning in laser powder-bed fusion additive manufacturing of 316L stainless steel, *Int. J. Adv. Manuf. Technol.*94 (2018) 3591–3603. https://doi.org/10.1007/s00170-017-1045-z. [2] Jones, K., Yang, Z., Yeung, H., Witherell, P., Lu, Y., Hybrid Modeling of Melt Pool Geometry in Additive Manufacturing Using Neural Networks, in: Vol. 2 41st Comput. Inf. Eng. Conf., American Society of Mechanical Engineers, 2021. https://doi.org/10.1115/DETC2021-71266. [3] Porro, M., Zhang, B., Parmar, A. and Shin, Y.C., "Data-driven Modeling of Mechanical Properties for 17-4 PH Stainless Steel Built by Additive Manufacturing", Integrating Materials and Manufacturing Innovation, 2022. [4] Luo, M. and Shin, Y.C., "Vision-based Weld Pool Boundary Extraction and Width Measurement during Keyhole Fiber Laser Welding", Optics and Lasers in Engineering, Volume 64, January 2015, Pages 59-70. [5] Zhang, Y., You, D., Gao, X.g, Zhang, N., Gao, P.P., Welding defects detection based on deep learning with multiple optical sensors during disk laser welding of thick plates, Journal of Manufacturing Systems, Volume 51, 2019, Pages 87-94. https://doi.org/10.1016/j.jmsy.2019.02.004. [6] Stadter, C., Schmoeller, M., von Rhein, L., Zaeh, M.F., Real-time prediction of quality characteristics in laser beam welding using optical coherence tomography and machine learning, *J. Laser Appl.*32, 022046 (2020). https://doi.org/10.2351/7.0000077 [7] Luo, M. and Shin, Y.C., "Estimation of Keyhole Geometry and Prediction of Welding Defects Based on a Visual System and a Radial Basis Function Neural Network", International Journal of Advanced Manufacturing Technology, Volume 81, Issue 1 (2015), Page 263-276. [8] Kershaw, J, Yu, R, Zhang, Y.M., Wang, P., Hybrid machine learning-enabled adaptive welding speed control, Journal of Manufacturing Processes, Volume 71, 2021, Pages 374-383. https://doi.org/10.1016/j.jmapro.2021.09.023. [9] Park, H.S., Nguyen, D.S., Le-Hong, T. et al. Machine learning-based optimization of process parameters in selective laser melting for biomedical applications. J Intell Manuf 33, 1843–1858 (2022). https://doi.org/10.1007/s10845-021-01773-4 [10] Barrionuevo, G.O., Walczak, M., Ramos-Grez, J., Sánchez-Sánchez, X., Microhardness and wear resistance in materials manufactured by laser powder bed fusion: Machine learning approach for property prediction, CIRP Journal of Manufacturing Science and Technology, Volume 43, 2023, Pages 106-114. https://doi.org/10.1016/j.cirpj.2023.03.002. [11] Wang, C., Zhong, W., Zhao, J., Insights on phase formation from thermodynamic calculations and machine learning of 2436 experimentally measured high entropy alloys, Journal of Alloys and Compounds, Volume 915, 2022, 165173. https://doi.org/10.1016/j.jallcom.2022.165173. [12] Zhang, Y., Wen, C., Wang, C., Antonov, S., Xue, D., Bai, Y., Su, Y., Phase prediction in high entropy alloys with a rational selection of materials descriptors and machine learning models, Acta Materialia, Volume 185, 2020, Pages 528-539. https://doi.org/10.1016/j.actamat.2019.11.067.