This thesis explores the use of deep learning, specifically Physics-Informed Neural Networks (PINNs), to model the transient thermal behavior of PCBs in space applications. A Convolutional LSTM architecture is trained to predict the temperature distribution over time on a 13 × 13 grid, representing a simplified PCB. By incorporating physical constraints into the loss function, the model significantly improves generalization under limited-data scenarios. With as few as 50 training cases, the proposed PINN reduces the Mean Absolute Error (MAE) by more than 96% compared to a purely data-driven baseline. The impact of the temporal resolution (Δt) and training set size is studied in detail. Results show that intermediate values of Δt, such as 10 s, provide the best trade-off between accuracy and stability. The use of physical loss terms also leads to narrower confidence intervals and improved robustness, particularly in low-data regimes. Finally, training cost and execution time are evaluated, confirming that although PINNs introduce overhead during training, their inference speed remains comparable to classical solvers. The proposed approach demonstrates the potential of PINNs for thermal modelling in embedded systems, especially when simulation data is scarce. Several future directions are discussed, including generalization to variable time steps, flexible heater configurations, and integration into onboard systems.
This thesis explores the use of deep learning, specifically Physics-Informed Neural Networks (PINNs), to model the transient thermal behavior of PCBs in space applications. A Convolutional LSTM architecture is trained to predict the temperature distribution over time on a 13 × 13 grid, representing a simplified PCB. By incorporating physical constraints into the loss function, the model significantly improves generalization under limited-data scenarios. With as few as 50 training cases, the proposed PINN reduces the Mean Absolute Error (MAE) by more than 96% compared to a purely data-driven baseline. The impact of the temporal resolution (Δt) and training set size is studied in detail. Results show that intermediate values of Δt, such as 10 s, provide the best trade-off between accuracy and stability. The use of physical loss terms also leads to narrower confidence intervals and improved robustness, particularly in low-data regimes. Finally, training cost and execution time are evaluated, confirming that although PINNs introduce overhead during training, their inference speed remains comparable to classical solvers. The proposed approach demonstrates the potential of PINNs for thermal modelling in embedded systems, especially when simulation data is scarce. Several future directions are discussed, including generalization to variable time steps, flexible heater configurations, and integration into onboard systems. Read More
Related posts:
Including AI experiments onboard the UPMSat-3 satellite mission
AAIA: an efficient aggregation scheme against inverting attack for federated learning
Gated recurrent unit neural networks for automatic modulation classification with resource-constrained end-devices
Deep Learning for Botanical Insight: Automated Identification of Tree Species from Anatomical Image Features
Convolutional Neural Networks Adapted for Regression Tasks: Predicting the Orientation of Straight Arrows on Marked Road Pavement Using Deep Learning and Rectified Orthophotography
Long short-term memory prediction of user’s locomotion in virtual reality