Abstract
Short-dipole diode sensors loaded with highly resistive lines are commonly used to measure the time-averaged square of the high-frequency electromagnetic field amplitude directly. Their precision, simplicity, broadband, high dynamic range capability, and minimal scattering make them ideal for application in the near-field of sources, particularly for demonstrating compliance with exposure limits. However, the usage of these sensors to cover multiple orders of magnitude of field amplitude requires signal-specific linearization of the sensor response. Traditionally, linearization had been performed for each signal or modulation by measurement and, more recently, by simulations based on a calibrated sensor model. These approaches have become prohibitively expensive with the launch of the fifth generation of mobile communication (5G), which added thousands of diverse and complex modulation schemes. In response to these challenges, we first developed an innovative approach to accelerate sensor model simulations with an enhancement of accuracy, which allows us to subsequently establish a data set comprising a large number of probe parameters and signal characteristic configurations. Subsequently, a physics-informed neural network (PINN) was trained with readily accessible signal characteristics to obtain on-the-fly linearization parameters with acceptable uncertainties across the relevant dynamic range. In contrast to traditional artificial intelligence (AI) models that predominantly rely on pattern recognition from precomputed data, our approach ensures that the model captures the intrinsic relationships and system dynamics inherent to the physical phenomena under study. Our AI-based approach achieves an error below 0.4 dB at peak specific absorption rate (SAR) values of up to > 200 W kg -1 . In addition, AI accelerates the determination of linearization parameters by a factor > 34,000 × and reduces storage requirements > 350,000 times, allowing linearization parameters to be computed on site.