Machine Learning-Based Prediction of Crystal Structure Proportions in Ag-TiO₂ Nanofibers for Enhanced Photocatalytic Activity

In the past decade, mechasim learning has suraged a revolution in the field of materials science and bringing about advancements in predicting materials properties, materials discovery, and materials characterization. Due to the remarkable activity, stability, and non-toxic nature, TiO₂ has emerged as a highly promising photocatalyst for pollutant removal and self-cleaning paints. The diverse crystal structures inherent in TiO₂ play a cricial role in modulating in its energy band structure, thereby enhancing its photocatalytic activity. However, the precise determination of optimal conditions for multi-phase TiO₂ poses a challenge, as it necessitates a thorough and time-consuming analysis of the quantitative crystal structure.<br/> To overcome this difficulty, we embarked on developing a robust model for predicting the proportion of crystal structures in multi-phase TiO₂. Prior to build a reliable model, we assembled a collection of Raman spectra encompassing various proportions of anatase and rutile phases for serving as database of machine learning. Over the past decade, our research team has focused on investigating metal-doped TiO₂ nanofibers (NFs). Among the studied compositions, silver doped TiO₂ NFs (Ag-TiO₂ NFs) exhibits the highest photocatalytic performance. Excepet for the interaction between inceidnet light and self-precipitated sliver nanoparticles, the phase composition of TiO₂ photocatalysts is also an influential factor for their photocatalytic activity. Although the anantase-to-rutile phase transition occurs at 900 to 1,100 °C in pristine TiO₂ NFs, the senerio is toltally different as a dopant induced into TiO₂ NFs. How to precisely control metal dopant and anatase-to-rutile phase transition in TiO₂ NFs still remains an ongoing challenge.<br/> Herein, we employed the "Variational Autoencoder" machine learning model to perform unsupervised learning and extract meaningful characteristics from the Raman spectra. After thorough inspection and verification, we successfully applied the model to predict the proportion of crystal structures in Ag-TiO₂ nanofibers with different calcination temperatures. By employing innovative and efficient methods, our research offers valuable insights into the correlation between quantitative crystal structure and photocatalytic activity. Notably, our model achieved a mean absolute error of 2.33% and a root mean square error of 0.0291, demonstrating its high predictive accuracy.<br/> By leveraging machine learning techniques, our study provides a powerful tool for predicting crystal structure proportions in multi-phase TiO₂, specifically focusing on Ag-TiO₂ NFs. This enables researchers and practitioners to efficiently analyze the relationship between crystal structure and photocatalytic activity, paving the way for the design and optimization of advanced TiO₂-based materials with enhanced performance in pollutant removal and self-cleaning applications.