We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Integrating nano crystal sensor with explainable deep learning for nutrients and microplastic-toxicity detection
Summary
Scientists have designed a highly sensitive light-based sensor (tested through computer simulations, not yet in a lab) that can detect soil nutrients and microplastic contamination with over 99% accuracy, using AI to interpret the readings. This matters because microplastics in soil can work their way into crops and our food supply, so a fast, precise way to spot this contamination could eventually help catch problems before they reach our plates. The tool still needs to be built and tested in real-world conditions before it can be used on farms.
This work proposes a simulation-based photonic-AI sensing framework for soil nutrient and microplastic detection. This framework integrates a 2D dual-ring cavity photonic crystal (PhC) sensor with a Deep & Cross Network model (DCN). The PhC sensor demonstrates strong optical confinement and spectral selectivity, achieving quality factors up to 18,244, with resonance wavelengths spanning 1528.5-1824.4 nm, high sensitivity as 631 nm/RIU for nutrients and 432 nm/RIU for Low density polyethylene (LDPE) microplastic detection. The PhC sensor further achieves figures of merit up to 3440 RIU⁻¹ and detection limits as low as 3 × 10⁻⁵ RIU, and stable operation under fabrication tolerances and temperature variations. The proposed DCN architecture effectively analyzes the resultant spectral responses of the different soil elements and contaminants. It precisely captures nonlinear spectral patterns without relying on refractive index as an explicit feature, avoiding classification ambiguity at similar concentration levels of soil elements. The introduced DCN model achieves high classification/identification performance of 99.87% accuracy and near-perfect precision, recall, and F1-score. For explainability, SHAP and LIME are employed to quantify spectral feature contributions and explain individual predictions. This explainable AI (XAI) analysis confirms the physical relevance of dominant spectral features used for decision-making. The performed Fabrication tolerance and temperature analyses confirm stable operation within practical limits. Although experimental validation is beyond the scope of this study, the compact sensor design and lightweight inference model support future hardware integration. These results demonstrate the potential of physics-based photonic sensing combined with explainable AI for intelligent soil monitoring.