Networked nanosensor data are high-dimensional, incomplete, noisy and nonlinear, challenging both first-principles and purely data-driven models. This review presents physics-informed supervised learning that embeds conservation laws, stochastic noise and network dynamics into regression and classification. It also covers cleaning, feature engineering, fusion, Kalman filtering, Fourier transforms and weighted-variance methods across environmental, biomedical and industrial cases, aiming to improve interpretability, robustness and generalization.
Key findings
- Domain constraints reduce solution space and improve plausibility; fusion integrates heterogeneous sensors; classical filters remain important for noise; scalability, out-of-distribution generalization and real-system integration remain open.
Why this matters globally
The approach may improve climate monitoring, outbreak sensing, predictive maintenance and medical monitoring by discouraging physically impossible outputs, provided standardized data and uncertainty validation are available.
Thai researcher contribution
Pradeep Bhadola and Vishal Chaudhary of Mahidol University contribute the cross-domain synthesis of sensor physics and machine intelligence.
Limitations to consider
The broad review lacks a common benchmark framework. ‘Physics-informed’ spans heterogeneous methods, incorrect or incomplete constraints can introduce bias, and successful case studies do not quantify deployment failure rates.
Verify the original sources
Applied Physics ReviewsRead the original article↗DOI: 10.1063/5.0309891