MTL-PINN: an uncertainty-weighted multi-task learning physics-informed neural network for two-dimensional transient temperature field prediction in wet friction components

Abstract High power density wet friction components serve as critical transmission components in heavy-duty equipment, and thermal failure induced by prolonged sliding under extreme conditions significantly undermines operational stability. Therefore, accurate prediction of interface temperature and analysis of thermal field characteristics hold substantial engineering value for improving system reliability. To investigate the spatiotemporal evolution and distribution characteristics of the friction interface temperature field, a physics-informed neural network is constructed by integrating neural networks with the two-dimensional transient heat conduction theory of wet friction components that accounts for the time-varying characteristics of key thermodynamic parameters. Data supervision is further incorporated to enhance the model’s capability to characterize the actual temperature evolution process. Meanwhile, the uncertainty-based multi-task learning loss weighting mechanism is introduced to enable the simultaneous optimization of constraint weights and network parameters, thereby achieving dynamic balancing and coordinated optimization among multiple constraint loss terms. In addition, the systematic evaluation of the model architecture and constraint mechanisms is carried out through the multi condition degradation experiment. The model is trained using 64 000 spatiotemporal collocation points and evaluated using 40 temperature measurement points for validation. Results demonstrate that the proposed model markedly improves the prediction accuracy of the temperature field in wet friction components, with a mean absolute percentage error of 0.224%, a relative squared error of 1.884%, and a coefficient of determination of 0.981. Moreover, the model achieves an average computational speed of 37 it/s. The temperature field exhibits a radial distribution that first increases and then decreases from the inner radius to the outer radius, with the highest temperatures observed in the mid-outer regions. Furthermore, the multi condition degradation experiment confirms the validity of the model framework as removing any constraint leads to reduced prediction accuracy and degraded convergence behavior.