Artificial intelligence meets friction research

Friction is everywhere – whether in engines, human joints, or tiny technical devices. Nevertheless, our understanding of friction is often still based on simple, inaccurate assumptions. One reason for this is that the physical processes involved in frictional contact – i.e., where two surfaces meet and move against each other – are difficult to study experimentally. Computer simulations of such tribological systems, which deal with friction, lubrication, and wear, can help to better understand these processes.

Researchers at the Institute for Microsystems Engineering (IMTEK) at the University of Freiburg and the Karlsruhe Institute of Technology (KIT) have now developed a new simulation method that can describe friction at the molecular level much more accurately – with the help of artificial intelligence. The study was published under the title “Active learning for non-parametric multiscale modeling of boundary lubrication” in the prestigious journal Science Advances.

The new method combines physical models on different length scales with machine learning techniques. This allows friction processes on the smallest scale – i.e., at the level of individual molecules – to be transferred to large, technically relevant systems. What makes this special is that the models contained in the study improve themselves independently with the help of so-called “active learning” methods. This involves the automatic generation of new training data, which continuously improves the models' predictions without drifting into speculative errors, as can sometimes happen with AI applications that have too little data.

“We show that friction can be predicted very realistically at the molecular level by combining artificial intelligence – i.e., machine learning – with physical expertise in a targeted manner,” explains Prof. Lars Pastewka, head of the simulation group at IMTEK. Dr. Hannes Holey, first author of the study, adds: “This is a real breakthrough. It allows us to better understand and calculate even very complex friction systems, known as tribological systems, such as those found in engines, joints, or precision devices. This opens up completely new possibilities for the development of low-friction or particularly efficient technical systems.”

The new model is seen as a significant advance because it combines data-based approaches with classical material simulation. This combination of machine learning and physical simulation opens up new possibilities for better understanding material behavior under stress. This allows materials to be improved in a more targeted manner and their service life to be increased.

The publication is the result of a joint research project between IMTEK at the University of Freiburg and KIT as part of the Research Training Group 2450. The work combines expertise from the fields of tribology, material simulation, and artificial intelligence.

Link to publication in Sciences Advances

Contact:
Prof. Dr. Lars Pastewka
Albert Ludwig University of Freiburg
Institute for Microsystems Technology (IMTEK)
Simulation
Email: lars.pastewka@imtek.uni-freiburg.de

Kerstin Steiger-Merx
PR/Marketing Officer
Faculty of Engineering
Albert Ludwig University of Freiburg
Tel.: +49 (0)761/203-8056
Email: steiger-merx@tf.uni-freiburg.de