Research Information System
Materials Science in Semiconductor Processing, vol.192, 2025 (SCI-Expanded)
The ability to adjust the phase composition in titanium dioxide (TiO2) structures is crucial for customizing their properties to fit various applications. However, traditional approaches struggle to accurately forecast and regulate the balance between amorphous and crystalline phases within these materials. Here, we introduced an innovative method, utilizing machine learning (ML) techniques, to predict and classify the ratio of amorphous to crystalline phases in TiO2 nanocomposites based on thermogravimetric analysis (TGA) data. Non-isothermal TGA experiments were conducted at heating rates of 1 °C/min, 5 °C/min, 10 °C/min, and 20 °C/min to obtain dataset. Various ML algorithms including Adaboost, Decision Trees (DT) Regression, Gaussian Process Regression (GPR), k-Nearest Neighbor Regression (KNN), Linear Regression (LR), Multi-Layer Perceptron (MLP), Random Forest Regression (RF), Support Vector Machine Regression (SVM) and XGBoost (XGB) were employed. The performances of models were evaluated by the R-squared (R2), root mean square error (RMSE) and mean absolute error (MAE) metrics for training and test data. Among these, GPR, KNN, RF, and XGB emerged as the top-performing algorithms, with GPR achieving an exceptional R2 value of 0.999 and the lowest error rates (RMSE: 2 × 10−4, MAE: 2.4 × 10−5). Thus, GPR was identified as the most successful regression model. As for classification part, the XGB algorithm achieved the highest accuracy of 99.9 % with DT, RF, and XGB also excelling in True Positive Rate (TPR) and False Positive Rate (FPR) metrics. These findings highlight the potential of ML techniques in optimizing phase composition prediction and classification for TiO2 nanocomposites, thereby reducing timescales, cost, and rigorous calculations.