edited by Lisa Lock, reviewed by Robert Egan
Editors' notes
Before and after: Above a laser weld seam before "learning," below a weld seam from the process optimized by the algorithm. Credit: Swiss Federal Laboratories for Materials Science and Technology
Laser-based processes for metals are considered to be particularly versatile in industry. Lasers can be used, for example, to precision-weld components together or produce more complex parts using 3D printing—quickly, precisely and automatically. This is why laser processes are used in numerous sectors, such as the automotive and aviation industries, where maximum precision is required, or in medical technology, for example for the production of customized titanium implants.
However, despite their efficiency, laser processes are technically challenging. The complex interactions between the laser and the material make the process sensitive to the smallest of deviations—whether in the material properties or in the settings of the laser parameters. Even minor fluctuations can lead to errors in production.
"To ensure that laser-based processes can be used flexibly and achieve consistent results, we are working on better understanding, monitoring and control of these processes," says Elia Iseli, research group leader in Empa's Advanced Materials Processing laboratory in Thun.
In line with these principles, Giulio Masinelli and Chang Rajani, two researchers from his team, want to make laser-based manufacturing techniques more affordable, more efficient and more accessible—using machine learning.
Vaporize or melt?
First, the two researchers focused on additive manufacturing, i.e., the 3D printing of metals using lasers. This process, known as powder bed fusion (PBF), works slightly differently to conventional 3D printing. Thin layers of metal powder are melted by the laser in exactly the right spots so that the final component is gradually "welded" out of them.
PBF allows the creation of complex geometries that are hardly possible with other processes. Before production can begin, however, a complex series of preliminary tests is almost always required. This is because there are basically two modes for laser processing of metal, including PBF.
In conduction mode, the metal is simply melted. In keyhole mode, it is even vaporized in some instances. The slower conduction mode is ideal for thin and very precise components. Keyhole mode is slightly less precise, but much faster and suitable for thicker workpieces.
Where exactly the boundary between these two modes lies depends on a variety of parameters. The right settings are needed for the best quality of the final product—and these vary greatly depending on the material being processed. "Even a new batch of the same starting powder can require completely different settings," says Masinelli.
Better quality with fewer experiments
Normally, a series of experiments must be carried out before each batch to determine the optimum settings for parameters such as scanning speed and laser power for the respective component. This requires a lot of material and must be supervised by an expert. "That is why many companies cannot afford PBF in the first place," says Masinelli.
Masinelli and Rajani have now optimized these experiments using machine learning and data from optical sensors that are already incorporated in the laser machines. The researchers "taught" their algorithm to "see" which welding mode the laser is currently in during a test run using this optical data.
Based on this, the algorithm determines the settings for the next test. This reduces the number of preliminary experiments required by around two-thirds—while maintaining the quality of the product.
"We hope that our algorithm will enable non-experts to use PBF devices," summarizes Masinelli. All it would take for the algorithm to be used in industry is integration into the firmware of the laser welding machines by the device manufacturers. The research is published in the journal Additive Manufacturing.
Real-time optimization
PBF is not the only laser process that can be optimized using machine learning. In another project, Rajani and Masinelli focused on laser welding—but went one step further. They not only optimized the preliminary experiments, but also the welding process itself. Even with the ideal settings, laser welding can be unpredictable, for example if the laser beam hits tiny defects on the surface of the metal.
"It is currently not possible to influence the welding process in real time," says Rajani. "This is beyond the capabilities of human experts." The speed at which the data have to be evaluated and decisions to be made is a challenge even for computers. This is why Rajani and Masinelli used a special type of computer chip for this task, a so-called field-programmable gate array (FPGA).
"With FPGAs, we know exactly when they will execute a command and how long the execution will take—which is not the case with a conventional PC," explains Masinelli.
Nevertheless, the FPGA in their system is also linked to a PC, which serves as a kind of "backup brain." While the specialized chip is busy observing and controlling the laser parameters, the algorithm on the PC learns from this data.
"If we are satisfied with the performance of the algorithm in the virtual environment on the PC, we can 'transfer' it to the FPGA and make the chip more intelligent all at once," explains Masinelli.
The two Empa researchers are convinced that machine learning and artificial intelligence can contribute a great deal more in the field of laser processing of metals. That is why they are continuing to develop their algorithms and models and are expanding their area of application—in collaboration with partners from research and industry.
More information: Giulio Masinelli et al, Autonomous exploration of the PBF-LB parameter space: An uncertainty-driven algorithm for automated processing map generation, Additive Manufacturing (2025). DOI: 10.1016/j.addma.2025.104677
Provided by Swiss Federal Laboratories for Materials Science and Technology