Physics-Informed Neural Networks for Low-Power Real-Time Edge Biosensing Applications

The growing need for real-time, energy-efficient biosensing has driven interest in deploying artificial intelligence (AI) at the edge, particularly for applications in healthcare and environmental monitoring. Traditional neural networks, though powerful, often require extensive training data and computational resources, making them unsuitable for low-power biosensor platforms. This paper presents a novel approach using Physics-Informed Neural Networks (PINNs) to enable accurate and efficient biosensing on edge devices. This research focuses on carbon nanotube (CNT)-based biosensors, where the underlying physical behavior can be described by partial differential equations. By embedding these equations into the training process, the proposed PINN model reduces data dependency and enhances interpretability while maintaining high prediction accuracy. The model was implemented on a Raspberry Pi 4 platform with a power budget of under 5W. Results show that our PINN achieved a 92.8% classification accuracy for target biomolecules, with an average inference latency of 24.6 ms and a 23.8% reduction in power consumption compared to a baseline convolutional neural network. These findings demonstrate that PINNs are well-suited for edge deployment, combining physical modeling with efficient AI.