Open Access

Abdelmalek Zahzam

• This thesis investigates the use of Artificial Intelligence (AI) models for predicting the performance characteristics of compressors, with a focus on centrifugal configurations. Using a dataset of 27 compressor samples manually extracted from academic literature, and further expanded through data augmentation. The study evaluates the predictive capabilities of three regression models: RandomThis thesis investigates the use of Artificial Intelligence (AI) models for predicting the performance characteristics of compressors, with a focus on centrifugal configurations. Using a dataset of 27 compressor samples manually extracted from academic literature, and further expanded through data augmentation. The study evaluates the predictive capabilities of three regression models: Random Forest, XGBoost, and Multi-Layer Perceptron (MLP). The models were trained to predict multiple compressor attributes, including choke and surge points for pressure ratio and mass flow rates and other geometric features, based on input design parameters. Advanced data preprocessing techniques, such as multiple imputation and feature scaling, were applied to improve model performance. A data augmentation strategy was also implemented to address the dataset’s limited size. Model evaluation was conducted using Mean Absolute Error (MAE) and R-squared (R²) metrics. The results indicate that XGBoost consistently outperformed the other models across most target features, especially after data augmentation, achieving R² scores above 0.9 for several outputs. Random Forest also delivered robust results, particularly in predicting isentropic efficiency and blade geometry. MLP showed poor performance across all scenarios, highlighting the challenges of applying deep learning to small, structured datasets. The study concludes that tree-based ensemble methods, particularly XGBoost, are well-suited for compressor performance prediction. The work also emphasizes the importance of data augmentation and preprocessing in enhancing model generalization. Despite limitations due to data scarcity, the findings demonstrate the viability of AI models as effective tools for compressor analysis and design optimization.…