Hybrid CNN-BiLSTM with Attention Mechanism for Predicting Photocatalytic Activity of CeO₂ Nanoparticles Synthesized via Green Routes

Abstract

We propose a hybrid convolutional neural network with Bidirectional Long Short‐Term Memory (CNN-BiLSTM) integrated with multi-head self-attention for predicting the photocatalytic activity of environmentally friendly synthesized CeO₂ nanoparticles using green routes like Azadirachta indica leaf extract. The architecture utilizes CNN layers to automatically extract spatial features hierarchically, BiLSTM units for sequential dependencies modeling and on top of that the attention mechanism to gather global context which has ultimately resulted in excellent prediction accuracy compared to traditional machine learning methods. Multimodal input features were experimental data from 1,200 synthesized samples characterized (structural (XRD), morphological (FESEM, TEM), optical (UV-Vis, PL) and chemical factors (XPS, EDAX)) for the individualized CeO₂ nanoparticle signatures.
The newly designed model known as Hybrid CNN-BiLSTM-Attention (HCBA) provides a prediction accuracy of 97.6%, Mean Absolute Error (MAE) scores equal to 0.023, RMSE = 0.031, and R² score = 0.979 which is further improved against the SVM, Random Forest as well as standalone CNN-LSTM variants B2 in TSI data domain setting and scope for detailed comparison with baseline Models from [23]. 10-fold cross-validation confirmed model robustness. SHAP-based importance analysis identified particle size, zeta potential and the Ce³⁺/Ce⁴⁺ oxidation ratio among the most predictive features for photocatalytic degradation efficiency. This proposed intelligent prediction framework accelerates the design of versatile nanomaterial for environmental remediation, especially for effective degradation of industrial dyes and purification of water/wastewater.