I'm a PhD researcher at the University of Surrey working on Machine Learning for Dynamical Systems in astrophysics. I build differentiable surrogates for expensive physical simulators, inverse methods for physical-parameter recovery, and research tooling for scientific ML.
First-author work accepted to MNRAS; Honourable Mention at the Tesseract Hackathon 2025 (Pasteur Labs).
Neural-ODE Surrogates for N-body Simulations · Accepted to MNRAS
Parameter-conditioned Neural ODEs as surrogates for N-body galaxy-merger simulations — median fractional error ~10⁻², zero-shot interpolation/extrapolation across unseen parameters, at a fraction of direct simulation cost. End-to-end autodiff through the surrogate recovers physically meaningful parameter sensitivities.
Differentiable Inverse Methods for PDE Parameter Recovery · Honourable Mention, Tesseract Hackathon 2025
Burgers viscosity recovery packaged as three swappable Tesseract components — a differentiable spectral solver, a JAX/PyTorch PINN, and an amortized flow-matching posterior (FMPE). Recovers viscosity three ways through one JAX/Optax loop; posterior validated with SBC and TARP coverage. Ongoing collaboration with Pasteur Labs.
Zero-Shot Parametric Neural Operators for Chaotic Dynamics
Extended CFO (ICLR 2026) to a parametric setting: a single PyTorch model that generalises zero-shot across Lorenz systems to unseen parameters without retraining, holding accuracy under sparse, irregularly-sampled observations where autoregressive baselines degrade.
Semantic Search for Scientific Literature
Deployed Streamlit app that ingests configurable arXiv categories, embeds titles/abstracts with SPECTER, and retrieves cross-domain literature by meaning rather than keyword — built on Sentence Transformers and FAISS.
- Scientific Machine Learning — Neural ODEs, PINNs, operator learning (FNO, DeepONet), PDE surrogates
- Generative & Foundation Models for Science — flow matching, scientific foundation models
- LLM & Retrieval Systems — semantic search and retrieval over scientific literature, with an eye toward agentic tooling for research
- Interpretability & Discovery — symbolic regression, equation discovery from data
Domain knowledge: N-body simulations, galaxy mergers, supermassive black hole binaries
