Can Neural Networks Learn Atomic Stick–Slip Friction?

Nanofriction experiments typically produce force traces exhibiting atomic stick–slip oscillations, which researchers have traditionally analyzed with ad hoc algorithms. This study successfully unravels the potential of machine learning (ML) to interpret nanofriction force traces and automatically extract Prandtl–Tomlinson (PT) model parameters. A prototypical neural network (NN) perceptron was trained on synthetic force traces generated by simulations across a wide parameter range. Despite its simplicity, this NN successfully analyzed experimental data, marking the first application of a network trained solely on computational data to experimental nanofriction. Challenges encountered in developing the NN model proved to be instructive and revealing. Poor transferability from synthetic to experimental data sets was resolved by incorporating physics-based descriptors into the synthetic training data, without experimental input. Our protocol’s simplicity underscores its proof-of-concept nature, paving the way for advanced approaches. Validation with experimental data, such as graphene-coated AFM tips on 2D materials, highlights the promise of this ML approach for stick–slip nanofriction studies.

纳米摩擦实验通常可获得呈现原子级黏滑振荡的力迹，过往研究者多采用特设算法对其进行分析。本研究成功挖掘了机器学习（ML）在解读纳米摩擦力迹、自动提取普朗特-汤姆林森（Prandtl–Tomlinson，PT）模型参数方面的潜力。研究团队在宽参数范围的仿真合成力迹数据集上训练了一款典型的神经网络感知机（NN）。尽管结构简单，该感知机仍成功完成了实验数据的分析，这是首个完全基于计算数据训练的神经网络应用于实验纳米摩擦研究的案例。开发该神经网络模型过程中遇到的挑战颇具启发性与揭示意义。通过在合成训练数据中引入基于物理的特征描述符，无需实验输入即可解决合成数据与实验数据集间迁移性不佳的问题。本研究方案的简洁性凸显了其概念验证属性，为后续先进研究方法的开发铺平了道路。通过对实验数据（如二维材料表面石墨烯涂层原子力显微镜（AFM）探针的相关测试数据）开展验证，本研究所采用的机器学习方法在黏滑型纳米摩擦研究中展现出良好的应用前景。