On this page, you will find all of my past and current projects. For deep-learning projects, the models were trained remotely on a Nvidia GPU provided by Google Colab. Written mainly in Python or C++, each project repertory contains:

  • codes in Jupyter format commented.
  • a pre-trained model that can be used for inference or fine-tuning.
  • a short report with figures and results.



Projects in progress

  • StarPINN: a physics-informed neural networks for fast stellar structure prediction

    This project leverages Physics-Informed Neural Networks (PINNs) to solve partial differential equations (PDEs) governing stellar structure. The goal is to train a neural network using four fundamental equations that describe the internal physics of stars. Once trained, the model predicts key structural properties such as core temperature, pressure field, mass, and chemical composition, for well-known stars, including Vega, Betelgeuse, Sirius, and the Sun.

    The PINN model is coded using TensorFlow, a framework that allows for building custom architectures, using specialized loss functions and performing custom training with Adam followed by BFGS (from TensorFlow Probability) as the optimizer.

🛠️    Python, TensorFlow, TensorFlow Probability, Nvidia Cuda, Adaptive sampling, Fourier feature embedding, Loss balancing

  • WinGNN: a graph neural networks for aerodynamic performance prediction

    For this project, the idea is to implement a graph neural network (GNN) model with an encoder-processor-decoder architecture and train it using results from mesh-based simulations of a wing in an airflow. The simulations are generated using OpenVSP, a medium-fidelity solver that applies the vortex lattice method to compute the scalar pressure coefficient field across the wing mesh. This field is used to derive aerodynamic curves (lift, drag and moment coefficients) and draw conclusions about the wing profile’s performance. Simulations are automated to cover a wide range of parameters such as Mach number, angle of attack, sweep angle and for different airfoil profile (NACA0012, NACA2412, and NACA23012)

    The goal of this project is to compare the results generated by the conventional OpenVSP software with the model’s predictions for new cases not used in training. The limitations of each model as well as the computational resource costs are also studied.

🛠️    Python, PyTorch, PyTorch geometric, PyTorch scatter, OpenVSP, Hyperparameter optimization



Past projects

    This project focuses on binary sentiment classification of movie reviews from the IMDb Reviews dataset, using three distinct models: a multilayer perceptron, a Word2Vec-based model and a recurrent neural network (RNN). Each model is designed to classify reviews as positive or negative, leveraging advanced techniques in natural language processing. The word embeddings generated by each model are visualized using dimensionality reduction techniques (PCA, t-SNE), providing an intuitive representation of the semantic space.

    Models are assessed using confusion matrices to analyze classification accuracy and ROC curves to evaluate the trade-off between true positive and false positive rates. The project compares the performance of MLP, Word2Vec, and RNN in capturing sentiment from text data.

🛠️    Python, TensorFlow, Nvidia Cuda, RNN, word2vec

    A custom convolutional neural network (CNN) was developed from scratch and trained to recognize beer brands using images of bottles from the Beer dataset. The model’s architecture and hyperparameters were optimized to maximize recognition accuracy.

    The custom CNN is compared against MobileNetV2, a pre-trained model fine-tuned using transfer learning. The comparison highlights the trade-offs between custom-built models and transfer learning approaches in terms of accuracy, training efficiency, and computational cost.

🛠️    Python, TensorFlow, Nvidia Cuda, CNN, Transfert learning, Hyperparameter optimization

  • Brainor: a semantic segmentation model for tumor localization

    This project leverages the Brain Tumor Image database to develop a semantic segmentation model for detecting brain tumors. The model is based on an encoder-decoder U-Net architecture, which classifies each pixel in a medical image as either tumor or healthy tissue. The resulting segmentation is visualized as a binary mask, clearly delineating the tumor region from the surrounding healthy areas.

    The model successfully localizes the tumor’s global position with high accuracy but the precise shape and boundaries remain approximate (moderate Dice coefficient). Extending the training phase with additional epochs could refine the segmentation quality, leading to a higher Dice coefficient and more accurate tumor delineation.

🛠️    C++, OpenCV, Nvidia Cuda