Titulo Estágio
Evaluating Machine Learning Architectures on Orbital Particle Classification using THOR-SR
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
DEI
Enquadramento
Understanding high-energy emissions in space is a key objective in modern astrophysics, enabling the investigation of extreme phenomena such as black holes, neutron stars, cosmic rays, and gamma-ray bursts. A central challenge in this domain lies in the accurate identification and classification of particles—such as protons, electrons, photons, and alpha particles—under highly dynamic and radiation-intense orbital conditions.
THOR-SR (TGF and High-Energy Astrophysics Observatory for Gamma-Rays) is a scientific experiment developed as part of the European Space Agency’s Space Rider mission. It is led by the Laboratório de Instrumentação e Física Experimental de Partículas (LIP) and designed to study high-energy astrophysical events and the radiation environment in Low Earth Orbit (LEO). Mounted on a reusable spacecraft capable of long-duration missions, THOR-SR offers a unique opportunity to perform continuous and high-fidelity data acquisition in space.
Preliminary work has demonstrated the feasibility of using convolutional neural networks (CNNs) to classify particles based on laboratory or simulated sensor data. However, with real orbital data now becoming available from THOR-SR, there is an opportunity to perform a rigorous evaluation of modern machine learning (ML) architectures under realistic conditions—including resource constraints, model robustness, and potential deployment in onboard systems.
This thesis aims to investigate and benchmark a variety of ML architectures for the classification of orbital particle data collected by THOR-SR, with a focus on transfer learning, model efficiency, and embedded applicability. The work contributes toward the advancement of intelligent, space-ready particle classification systems in upcoming ESA missions.
Objetivo
The thesis is structured around the following key objectives:
1. THOR-SR Data Analysis and Characterization: Analyze particle data collected by THOR-SR (starting with previous data already available in the project), identifying relevant patterns and classification challenges in a space environment. The impact of noise in the overall is also to be explored.
2. Evaluation of ML Architectures: Compare various machine learning models (e.g., CNNs, Transformers, lightweight architectures like MobileNet and EfficientNet) in terms of accuracy, generalization, and computational efficiency.
3. Transfer Learning Exploration: Investigate the use of pre-trained models (from simulated or laboratory datasets) and assess their adaptation to THOR-SR real-world orbital data.
4. Resource-Aware Model Optimization: Evaluate memory footprint, inference time, and model compression techniques (e.g., pruning, quantization) for potential deployment on spaceborne computing platforms.
5. Design of an End-to-End Particle Classification Pipeline: Propose and validate a robust and scalable ML-based pipeline tailored to the needs of orbital particle detection and classification missions.
Plano de Trabalhos - Semestre 1
Literature Review
Review of the current state of the art in high-energy astrophysics, particle detection, ML classification techniques, transfer learning, and embedded AI systems for space applications.
[13/10/2025 to 09/11/2025] Exploratory Data Analysis and Preprocessing
Examine the structure and characteristics of THOR-SR data. Perform cleaning, normalization, and feature extraction to prepare datasets for training and evaluation.
[10/11/2025 to 07/12/2025] Baseline Model Development
Implement baseline ML models (e.g., simple CNNs, MLPs) and conduct initial experiments using real or simulated particle datasets. Record performance benchmarks.
[08/12/2025 to --/01/2026] Thesis Proposal Writing
Prepare and submit the full thesis proposal, including motivation, goals, methodology, and initial experimental setup.
Plano de Trabalhos - Semestre 2
Advanced ML Modeling and Transfer Learning
Train and fine-tune more advanced ML models, including transfer learning from existing models trained on synthetic or laboratory data. Assess improvements in accuracy and adaptability.
[02/03/2026 to 19/04/2026] Model Optimization for Deployment
Apply model optimization techniques such as pruning, quantization, and knowledge distillation. Evaluate runtime efficiency and hardware feasibility for onboard processing in future missions.
[20/04/2026 to 10/05/2026] Experimental Validation and Analysis
Conduct large-scale validation across diverse data segments. Analyze trade-offs between classification performance and computational overhead.
[11/05/2026 to --/06/2026] Thesis Writing
Document the work in the final thesis manuscript, including methodology, results, discussion, and proposed contributions to THOR-SR and future space missions. Prepare for defense and dissemination (e.g., scientific publication).
Condições
This work occurs within the context of the THOR-SR (PEA 4000141332) project, funded by the European Space Agency (ESA). The work is to be executed at the laboratories of the CISUC’s Software and Systems Engineering (SSE) Group and in collaboration with Laboratório de Instrumentação e Física Experimental de Partículas (LIP) Coimbra.
Observações
Tese co-orientada por Rui Curado Silva (LIP) e Gabriel Falcão (DEEC)
Orientador
Joao Campos
jrcampos@dei.uc.pt 📩