Can Neural Networks Learn Atomic Stick-Slip Friction?

Area of Science:

• Surface Science
• Tribology
• Computational Physics

Background:

• Nanofriction experiments generate force traces with atomic stick-slip oscillations.
• Traditional analysis relies on ad hoc algorithms, lacking standardization.

Purpose of the Study:

• To explore machine learning (ML) for interpreting nanofriction force traces.
• To automatically extract Prandtl-Tomlinson (PT) model parameters using ML.
• To demonstrate the transferability of ML models from simulation to experimental data.

Main Methods:

• A neural network (NN) perceptron was trained on synthetic force traces from simulations.
• Physics-based descriptors were incorporated into synthetic data to improve model transferability.
• The trained NN was applied to analyze experimental nanofriction data.

Main Results:

• The ML model successfully analyzed experimental nanofriction force traces.
• The NN extracted Prandtl-Tomlinson (PT) model parameters automatically.
• Incorporating physics-based descriptors resolved transferability issues between synthetic and experimental data.

Conclusions:

• Machine learning offers a powerful, automated approach to analyze stick-slip nanofriction.
• Physics-informed machine learning enhances model robustness and applicability to real-world data.
• This study provides a proof-of-concept for advanced ML applications in nanofriction research.