Computational physicist — physics-informed machine learning for orbital dynamics and physical systems.
MSc in Physics (Theoretical Physics and Astrophysics), University of Turin — 110 cum laude. My thesis applied physics-informed neural networks to multi-regime satellite orbit propagation, reproducing a Cowell reference propagator to within 1% and generalising to satellites unseen in training (0.03%–0.67% relative error); a paper on that work is in preparation.
My research question is a sharper version of the same one: can a learned model recover physics rather than interpolate data? In one project, a Fourier Neural Operator trained on a band of parameter space that deliberately excludes the magnetic stabilization threshold of a Kelvin–Helmholtz instability reconstructs the growth-rate curve inside that unseen band and assigns the correct stability verdict to 97% of held-out runs. In my latest one, a PPO agent learns an orbital rendezvous from scratch and is held against the classical controllers (LQR, two-impulse transfer) on starting points it never saw.
I build simulators from first principles and validate them against reference data — not notebooks, but tested packages with CI, C++ kernels where speed matters, and physics-validation suites.
| Project | What it is |
|---|---|
| Orbital_Rendezvous | A PPO agent that learns to dock a chaser spacecraft in low Earth orbit, in the Clohessy–Wiltshire relative frame propagated exactly. It docks 199 times out of 200 on unseen starts, sooner and cheaper than the fastest LQR controller that never crashes; includes a live training window, fuel-budget and docking-port studies, and an honest account of where the classical V-bar procedure still wins. |
| MHD Neural Operator | A Fourier Neural Operator emulating a magnetized Kelvin–Helmholtz instability, tested on a held-out band of parameter space around the magnetic stabilization threshold. Includes the from-scratch pseudo-spectral MHD solver that generates the ground truth — and a documented account of the two loss functions that failed first. |
| Tokamak | End-to-end fusion reactor simulator: transport PDEs, Grad–Shafranov equilibrium, feedback control, ML surrogates. Validated against ITER parameters with 93 physics tests, CI, a pybind11 C++ kernel and a Streamlit dashboard. |
| Three_Body | Sun–Earth–Jupiter system in C++ (RK4, RKCK), with energy-stability and chaotic-dynamics analysis. |
| Stellar_Radius_Estimation | Stellar radii from multi-band photometry, with Monte Carlo uncertainty propagation and formal consistency tests against reference measurements. |
| Warp_Drive | Numerical study of Alcubierre and Van Den Broeck warp-bubble spacetimes: exotic energy budget, horizons, null-geodesic ray tracing of the sky seen from inside the bubble, the throat of the Van Den Broeck pocket with its energy floor and quantum-inequality checks, and an acoustic analogue in a Bose–Einstein condensate. |
| F1-strategy-engine | Race strategy simulator: Monte Carlo analysis, ML tyre-degradation models, live safety-car re-optimisation. |
| Particle_EM | Charged-particle motion in prescribed EM fields — RK4 and Boris integrators, C++. |
| DM_direct_detection | WIMP direct-detection rates under the Standard Halo Model, annual modulation, Xenon vs NaI targets. |
| Lane_Emden_Solver | Lane–Emden equation for polytropic stellar models, with Chandrasekhar mass estimation. |
Orbit propagation and determination · reinforcement learning for guidance and control · space situational awareness and debris · thermospheric density and satellite drag · physics-informed neural networks and neural operators · scientific machine learning
Python (PyTorch, TensorFlow, Stable-Baselines3, Gymnasium, NumPy/SciPy) · C++ · LaTeX · Git, GitHub Actions
Open to PhD and R&D positions in machine learning for astrodynamics and space systems.