Machine Learning Driven Analysis and Prediction of Hazardous Gas Concentrations: A Comprehensive Review of Sensing Data, Algorithms, and Emerging Applications

Hazardous gases-carbon monoxide, hydrogen sulphide, ammonia, nitrogen oxides, methane, and volatile organic compounds-pose severe risks to health, industrial safety, and environmental quality, making continuous monitoring and early prediction a critical engineering priority.Conventional sensing systems, based on metal-oxidesemiconductor (MOS), electrochemical, and infrared transducers, suffer from crosssensitivity, baseline drift, and non-linear response, limiting the reliability of threshold-based alarms.Over the past decade, machine learning (ML) and deep learning (DL) have emerged as powerful tools for extracting selective, drift-resilient, and temporally predictive information from sensor signals, enabling accurate gas identification and multi-step-ahead concentration forecasting.This review surveys thirty recent studies (2020-2026) coupling hazardous gas sensing hardware with ML pipelines across domains including IoT-based indoor and rural air quality monitoring, underground coal mine safety, electronic nose (eNose) gas mixture identification, power transformer dissolved gas analysis, industrial emission control, and cyber-physical anomaly detection.Methodologies include classical algorithms (SVM, Random Forests, k-NN), ensembles (XGBoost, Gradient Boosting, Stacking), neural networks, and sequence-to-sequence models (LSTM, GRU, BiLSTM, BiGRU, Transformer, Autoformer).The comparison proves that hybrid and bidirectional architectures deliver better performance in multi-step forecasting tasks compared to a single model, while tree-based ensembles continue to be relevant for low-latency edge computation tasks.The review highlights the difficulties connected with sensor drifting, dataset limitations, cross-sensitivity, and the challenge of transferring models from simulated environment to the real world, suggesting future research directions in federated learning, benchmarks, explainable AI, and edge-AI compression.