Application of Neural Networks for Advanced Ir Spectroscopy Characterization of Ceria Catalysts Surfaces

This study presents a novel convolutional neural network (CNN) architecture that represents a significant advancement in the unsupervised analysis of data from infrared (IR) spectroscopy, both in IRRAS (infrared reflection absorption spectroscopy) and in DRIFTS (diffuse reflection infrared Fourier transform spectroscopy). After measuring reference data for single‐crystal samples using IRRAS, DRIFTS allows the characterization of surfaces exposed by cerium oxide powder particles through the stretch frequency of adsorbed probe molecules. To enable real‐time monitoring of catalyst modification during exposure to reactive gases under reaction conditions, a rapid, unsupervised analysis of the DRIFTS data is required. It is demonstrated that this goal can be achieved by using a CNN with an optimized architecture. This model is proficient in determining the intensities of the adsorbed CO bands, which depend on the crystallographic orientation and oxidation state of the exposed facets. The CNN design incorporates parallel 1D convolutional layers with varied kernel sizes. These layers work in tandem to capture spectral features. To address the challenge of overfitting, advanced regularization techniques within the CNN are integrated, enhancing the model's performance on new, unseen data. In particular, this approach to generating synthetic data has been instrumental in improving the performance of the CNN. The employment of the Adam optimizer and the mean squared error loss function aligns the model for efficient learning, ensuring accurate and reliable predictions. By introducing this CNN architecture, a robust, precise, and adaptable tool for rapid, unsupervised spectroscopic analysis is provided, demonstrating the potential of deep learning combined with synthetic data generation for advanced spectroscopy applications.