Ultrasonic-Net: a physics-informed neural network for ultrasonic field prediction in cleaning systems

Ultrasonic cleaning systems generate acoustic fields using high-frequency transducer vibrations. While transducer parameters can be tuned, a unified and efficient model for predicting field behavior across diverse cleaning conditions remains unavailable. In this work, we present Ultrasonic-Net, a Physics-Informed Neural Network (PINN) developed to predict ultrasonic fields, specifically targeting cleaning applications. The JAX-Fluids solver is used to generate ground truth for evaluation. The complex system, comprising the ultrasonic transducers (acoustic sources) and the target devices, is modeled using a sharp interface method. The design of Ultrasonic-Net is based on a synthesis of detailed physics and established knowledge. In the proposed architecture, Multi-scale Convolutional Neural Networks (Multi-Scale CNN) and Fast Kolmogorov-Arnold Networks (FastKAN) have been integrated into the Spline-based PINN. The Attention Fusion Layer integrates features from the multi-scale local and global branches by performing adaptive channel-wise recalibration on their concatenated representations. During training, the wave equation is employed as a physical constraint. The vibration frequency, along with the variations in transducers, and the configuration of the devices to be cleaned, is incorporated through initial and boundary condition (IBC) constraints. The decomposition method combined with vectorization is designed to facilitate the enforcement of these constraints. The combination is motivated by ultrasonic cleaning-specific physical requirements. Systematic ablation studies across seven PINN variants and multiple hyperparameter groups confirm that the proposed architecture and training configuration achieve the best accuracy–efficiency trade-off. Once trained, Ultrasonic-Net efficiently predicts the spatiotemporal evolution of acoustic waves under varying conditions, closely matching ground truth while significantly reducing computational cost. An extended experimental comparison is provided to further validate the model's performance. The model's practical utility is demonstrated through application to acoustically driven bubble dynamics. Finally, a systematic evaluation confirms its accuracy and generalization capacity across diverse conditions.