Industrial laser processes, vital for precise manufacturing, are being revolutionised by machine learning. This innovation promises to simplify complex operations in metal processing, making them more efficient and accessible for various applications, from automotive to medical
Industrial laser processes, long praised for their precision and speed in manufacturing everything from automotive parts to medical implants, are notoriously complex. The delicate balance between the industrial laser and material often requires extensive, time-consuming, and costly preliminary testing for metal processing.
To help remedy this, researchers at the Swiss Federal Laboratories for Materials Science and Technology (Empa) are ushering in a new era of efficiency and accessibility, leveraging the power of machine learning to simplify and optimise these sophisticated techniques in metal processing.
Overcoming the laser’s intricacies in metal processing
Laser-based metal processing, including precision welding and 3D printing (additive manufacturing), offers unparalleled versatility and the ability to create intricate geometries. This makes them indispensable in industries demanding the highest precision, such as automotive, aviation, and medical technology.
Yet, the sensitivity of these industrial laser processes to minute deviations in material properties or laser parameters can lead to costly errors and inconsistencies in metal processing.
As Elia Iseli, research group leader at Empa’s Advanced Materials Processing laboratory, emphasises, the goal is to achieve a better understanding, monitoring, and control for flexible and consistent results in metal processing. Empa researchers Giulio Masinelli and Chang Rajani are spearheading this effort, aiming to make laser-based metal processing more affordable, efficient, and accessible through machine learning.
Unlocking optimal 3D printing with fewer experiments
A key challenge in laser-based 3D printing, specifically Powder Bed Fusion (PBF), lies in determining the correct operating mode for effective metal processing. PBF involves melting thin layers of metal powder with an industrial laser to gradually build a component.
The process can operate in either “conduction mode,” where the metal is simply melted for thin, precise components, or “keyhole mode,” which involves some vaporisation, offering faster processing for thicker workpieces but with slightly less precision.
The boundary between these modes, and thus the optimal settings for parameters like scanning speed and laser power, varies significantly with different materials and even new batches of the same powder in metal processing.
Traditionally, optimising these settings has required a series of preliminary experiments, consuming considerable material and demanding expert supervision – a barrier for many companies involved in metal processing.
Masinelli and Rajani have addressed this issue by ‘teaching’ an algorithm to identify the welding mode during test runs using optical sensor data already available in industrial laser machines. Based on this real-time insight, the algorithm intelligently determines the settings for subsequent tests, effectively reducing the number of required preliminary experiments by approximately two-thirds without compromising product quality in metal processing. This breakthrough, Masinelli hopes, will empower non-experts to utilise PBF devices, requiring only integration into the industrial laser welding machine firmware by manufacturers.
Real-time refinement for flawless metal welds
The benefits of machine learning extend beyond preliminary optimisation in metal processing. Rajani and Masinelli have also focused on industrial laser welding, taking their innovation a step further by optimising the welding process itself in real time. Even with ideal initial settings, unexpected variables like tiny surface defects can compromise weld quality in metal processing. The sheer speed at which data needs to be evaluated and decisions made during the welding process is beyond human capability, and even conventional PCs struggle with this rapid throughput.
To overcome this, the researchers employed a specialised computer chip known as a Field-Programmable Gate Array (FPGA). FPGAs offer predictable execution times, which are crucial for the instantaneous decision-making required in the real-time control of industrial lasers. While the FPGA meticulously observes and controls laser parameters during metal processing, a linked PC acts as a “backup brain,” continuously learning from the gathered data. This dual system allows for continuous improvement: once the algorithm’s performance is validated in a virtual environment on the PC, it can be seamlessly “transferred” to the FPGA, making the chip incrementally more intelligent for metal processing applications.
Masinelli and Rajani are confident that machine learning and artificial intelligence hold immense potential for further advancements in the field of industrial laser processing of metals. Their ongoing work, in collaboration with research and industry partners, continues to expand the application of their innovative algorithms and models, paving the way for more precise, cost-effective, and accessible laser metal processing across various industries.
![Figure 1: Sketch of the evolution of the Universe over the last 13.77 billion years. It started with the Big Bang, followed by an extremely short period of rapid exponential expansion. The furthest we can see is the cosmic microwave background, when radiation decoupled from matter, approximately 380,000 years after the Big Bang. This is followed by the ‘dark ages,’ during which this radiation redshifted from the visible regime into infrared and sub-mm wavelengths. The occurrence of the first stars, about 400 million years after the Big Bang, ended this phase, spearheading the formation of galaxies as we see them today. [Credit: NASA/WMAP Science Team, public domain]](https://www.openaccessgovernment.org/wp-content/uploads/2025/05/Fig-1_1200-100x70.jpg "How did the first stars form in space?")
