Data for ''Deep learning-based prediction of time-resolved adhesive forces in viscoelastic Hertzian contacts''

This repository contains the dataset and computational materials associated with the study “Deep learning-based prediction of time-resolved adhesive forces in viscoelastic Hertzian contacts.” The dataset comprises 12,450 dimensionless force trajectories generated using a boundary-element model of a viscoelastic substrate with a Lennard–Jones traction–separation law. It covers loading and unloading rates spanning four orders of magnitude, different dwell times and indentation depths, and Tabor parameters ranging from 0.2 to 3.2. The repository includes the BEM-generated trajectories, data processed using the proposed fixed-measurement-step (FMS) representation, model-training and evaluation scripts, trained-model outputs, and materials required to reproduce the principal figures and results. Eighteen sequence-to-sequence architectures are evaluated, demonstrating the importance of stateful LSTM networks for learning the history-dependent response across different loading scales and adhesion regimes. The selected LSTM model predicts complete adhesive-force histories with median errors of approximately 2.2% in pull-off force and 1.1% in hysteresis area, while providing predictions roughly three orders of magnitude faster than the reference BEM calculations. The research was supported by the European Union through the Marie Skłodowska-Curie Actions Postdoctoral Fellowship REAL-ADHERE (Grant Agreement No. 101285102) and the ERC Starting Grant SURFACE (Project No. 101039198).