A hybrid physics-informed neural and explainable AI approach for scalable and interpretable AQI predictions

Area of Science:

• Environmental Science
• Data Science
• Computational Science

Background:

• Air pollution poses significant risks to public health, climate, and ecosystems.
• Accurate prediction and classification of Air Quality Index (AQI) are challenging due to complex pollution dynamics.
• Conventional machine learning (ML) and deep learning (DL) models struggle to incorporate physical laws, limiting prediction accuracy.

Purpose of the Study:

• To introduce AirSense-X, a novel approach combining Physics-Informed Neural Networks (PINN) and Explainable AI (XAI) for improved AQI classification.
• To enhance AQI prediction by integrating physical laws governing air pollution into the neural network model.
• To provide a comparative analysis of the proposed approach against conventional ML and DL models.

Main Methods:

• Utilized Physics-Informed Neural Networks (PINN) for AQI regression, incorporating physical laws into the learning process.
• Employed a structure mapping technique for AQI classification based on predicted values.
• Integrated Explainable AI (XAI) techniques for model interpretability and understanding prediction drivers.

Main Results:

• Achieved high accuracy (98%), precision (97%), recall (95%), and F1 score (0.96) in AQI classification.
• Demonstrated superior performance compared to conventional and ensemble ML/DL models.
• The confusion matrix showed accurate classification of 21,306 instances with only 268 misclassifications.

Conclusions:

• The AirSense-X approach effectively enhances AQI prediction accuracy and reliability by integrating physical laws.
• PINN models offer a significant advantage over purely data-driven models in capturing complex environmental phenomena.
• The combination of PINN and XAI provides accurate, interpretable, and reliable air quality index classification.