Gaurav Khanal
MSc Data Science & Artificial Intelligence
Scientific machine learning, physics-informed modeling, inverse problems, and topology- and geometry-aware AI for physical and biomedical systems.
Education
MSc in Data Science and Artificial Intelligence
Université Côte d’Azur · Sophia Antipolis, France
EFELIA Bourse d’Excellence, 2025
Sep 2024–Aug 2026
BA in Economics and Accounting
Westminster College · Fulton, Missouri, USA
Aug 2009–May 2013
Current Professional Experience
Research Intern, Predictive Constitutive Modeling of Atrial Appendage Tissue
Mines Saint-Étienne · Centre Ingénierie et Santé · SAINBIOSE Lab, Inserm U1059
Apr 2026–Present
Developing mechanics-informed machine learning models to infer specimen-specific force response and constitutive behavior of atrial appendage tissue from projected 2D thickness maps and 3D surface mesh scans. Current work includes 3D morphology and anisotropy descriptors, point-cloud and mesh representation learning, topology-aware shape descriptors, and hybrid constitutive layers for force-response prediction.
Selected Research Projects
Beyond Energy MAE: Input-Gradient Geometry in Molecular GNNs (In progress)
Designing a controlled energy-only comparison of SchNet and NequIP on 50,500 QM7-X molecular families, using held-out reference forces, numerical \(\mathrm{SE}(3)\) symmetry audits, and Wilson B-matrix/Shapley attribution to test whether comparable energy MAE can conceal differences in learned energy-surface geometry.
Physics-Informed Image Registration for Hyperelastic Material Identification
Developed a unified PINN framework coupling differentiable image registration with Ogden hyperelastic parameter identification, replacing the classical DIC to displacement extraction to constitutive fitting pipeline.
Deep Learning for Full Waveform Inversion
Built a hybrid Swin V2 Transformer and enhanced U-Net framework for reconstructing subsurface velocity fields from multi-source seismic shot gathers, trained at OpenFWI scale with PyTorch DDP/AMP on H100 GPUs.
PERSIST: Topology-Guided Multiscale Domain Identification
Designed an unsupervised topology-guided framework for spatial transcriptomics using persistent homology and discrete exterior calculus based multiscale stability fields.
Graph Signal Processing for Characterization of Multipolar Electrograms in AF
Developed a geometry-aware graph signal processing framework for multipolar intracardiac electrograms, using joint time-graph spectral analysis and graph total variation to detect rotor-proximal spatial discordance.
NeurIPS 2025 EEG Foundation Challenge
Built DDP/HPC-ready pipelines for both challenge tracks, combining Barlow Twins self-supervision and an EEGNeX backbone for cross-task transfer with masked-autoencoder pretraining, subject-level multiple-instance learning, MixStyle, and domain-adversarial training for cross-subject externalizing-factor prediction on HBN-EEG.
Research Interests
- Scientific machine learning and AI for science
- Physics-informed modeling and inverse problems
- Geometric and topological deep learning
- Information geometry, optimal transport, and learning dynamics
- Medical image registration and biomedical signal analysis
Technical Skills
Earlier Professional Experience
Before returning to graduate study in data science and AI, I worked in consulting across the United States and Australia, including roles at Ernst & Young, Protiviti, and PricewaterhouseCoopers.