Integrated equivalent circuit and deep neural network approach for real-time toxic gas detection and device integrity monitoring

• Integrated ECM and deep neural networks for real-time toxic gas detection. • Graphene-based metamaterial absorber designed for wearable sensor monitoring. • ECM offers fast, accurate alternatives to full-wave electromagnetic simulations. • Evolutionary neural networks enhance gas classification accuracy to 95 %. • Framework validated by multi-method benchmarking and practical sensor scenarios. This study introduces an integrated framework that combines a physics-based Equivalent Circuit Model (ECM) with deep neural networks (DNNs) for real-time detection of toxic gases and continuous monitoring of sensor performance. The ECM models the electromagnetic behavior of a graphene-based sensor using equivalent circuit components resistors, inductors, and capacitors to represent its graphene layers, dielectric substrates, and metal backings. By incorporating graphene's surface conductivity and chemical potential, the model enables dynamic tuning of circuit elements, allowing precise control of absorption responses under varying environmental conditions. It offers a computationally efficient alternative to full-wave finite element simulations, enabling fast impedance predictions critical for real-time and wearable applications. Absorption data derived from frequency-dependent reflection and incident wave measurements capture sensor sensitivity to gases like CO, CO₂, CH₄, and O₃. These datasets are processed using convolutional and recurrent neural networks (CNNs and RNNs) optimized through an evolutionary neural network (ENN) approach. The ENN enhances model accuracy and generalization via biologically inspired operations selection, crossover, and mutation that prevent local minima and accelerate convergence. The multilayer metamaterial absorber design, incorporating flexible substrates and graphene resonator rings, enhances sensitivity and adaptability for wearable platforms. Experimental validation shows the ENN-optimized model achieves 95 % accuracy with a six-hour convergence time, outperforming gradient descent (88 %, 12 h), genetic algorithms (91 %, 9 h), and reinforcement learning optimization (93 %, 8.5 h). The proposed framework demonstrates superior computational efficiency, robustness, and real-time capability for intelligent gas-sensing systems.