Design and implementation of a novel high-efficiency ultrasonic 3D synthetic data generation framework

Abstract This study introduces a Physics-Informed Neural Operator framework for high-efficiency, three-dimensional air-coupled ultrasonic simulation, effectively dismantling the barriers of data scarcity and computational latency that hinder next-generation Artificial Intelligence-driven sensing. By embedding the governing heterogeneous acoustic wave equations directly into the neural operator’s architecture, our framework achieves a 22-times speedup over traditional Fourier pseudo-spectral method without compromising physical fidelity. To enable large-scale, long-horizon acoustic forecasting, we propose an innovative patch-based simulation strategy coupled with a Factorized Fourier Neural Operators post-processing calibration model, which suppresses error accumulation and overcomes GPU memory constraints. This Physics-Informed Neural Operator core is seamlessly integrated into the NVIDIA Omniverse ecosystem, creating a fully automated digital-twin workflow for rapid scene configuration and dataset synthesis. Extensive experimental validation demonstrates an alignment between synthetic signals and real-world sensor measurements, achieving a correlation coefficient exceeding 0.99. This platform establishes a new benchmark for scalable acoustic simulation, significantly accelerating the research-to-deployment cycle for advanced ultrasonic haptics, object recognition, and human-computer interaction.