Automated morphological classification of LMC-like galaxies through machine learning

Sammanfattning

Understanding the morphological evolution of galaxies requires efficient and accu- rate classification methods, particularly in large-scale simulations and observations. This thesis presents a deep learning approach for the automated morphological classification of Large Magellanic Cloud-like galaxies, using high-resolution N-body simulations from the KRATOS suite. The study focuses on detecting structural features, such as bars and spiral arms, within the evolving stellar density maps of Large Magellanic Cloud-like galaxies undergoing gravitational interactions with the Milky Way and the Small Magellanic Cloud. Using convolutional neural networks, the project developed binary classifiers for bar and spiral morphologies. The network was trained on 2D face-on galaxy density maps, ensuring alignment and centring based on stellar angular momentum. Despite challenges such as class imbalance and subtle morphological features, the network achieved 92% accuracy on unseen test data and generalised well to an unseen simulation (K20), where it maintained 89% classification accuracy. Gradient-weighted Class Activation Mapping visualisations confirmed that the model consistently identified relevant morphological regions, while receiver operating characteristic and precision-recall curves indicated fair performance, more so in spiral classification. The CNN model experienced limitations due to a small and imbalanced dataset, which could be addressed by using an expanded dataset and higher-resolution data. Finally, this work demonstrated that deep learning techniques can enhance the efficiency and consistency of galaxy morphology classification in simulations.