Skip to content
View julian-8897's full-sized avatar
:octocat:
Working from home
:octocat:
Working from home

Block or report julian-8897

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
julian-8897/README.md

Hi, I'm Julian

I'm a research engineer in AI for Science and a final-year PhD researcher at the University of Surrey. I build machine learning methods and tools for scientific discovery, spanning LLM agents, differentiable simulation and neural surrogates for complex dynamical systems.

LinkedIn Email

LLM agents for scientific ML

  • Prior or Feedback? What an LLM Uses When Adapting Neural Operators (code): Studied LLM-guided adaptation of Fourier neural operators across PDE regimes. The LLM outperformed random search and TPE in nearly all matched evaluations, while controlled interventions revealed how experimental feedback influences its decisions. Reproducible experiment records and verification tools are available. Equal contribution.

    NeurIPS 2026 AI4Science Workshop

  • Research Brief Agent: Built a tool-using LLM agent that retrieves scientific papers, synthesises technical briefs and verifies citations against source evidence. Uses FastAPI, Qdrant and automated evaluations in CI to assess reliability.

Differentiable simulation

  • Differentiable hybrid closure for 2D turbulence (walkthrough): Developed a neural turbulence closure trained end-to-end through a differentiable spectral solver using PyTorch and JAX. Reduced long-horizon vorticity error by 71.7% relative to the coarse solver and 15.7% relative to conventional supervised training. Also trained with the solver and network running as separate Tesseract services. Tesseract Hackathon 2026.

  • Inverse problems and simulation-based inference for PDEs: Built a differentiable framework for recovering PDE parameters from sparse, noisy observations, combining solver-based inference, JAX/PyTorch physics-informed neural networks and flow-matching posterior estimation. Includes uncertainty calibration and interchangeable inference methods. Presented at the Tesseract Forum 2026.

    Tesseract Hackathon 2025 Honourable Mention

ML for astrophysics

  • Neural ODE surrogates for black hole binary evolution: Developed a neural ODE surrogate for the orbital evolution of supermassive black hole binaries, trained on N-body simulations of galaxy mergers. Reproduces merger timescales at a fraction of the computational cost of direct simulation. First author.

    MNRAS 2026

Stack

Python PyTorch JAX NumPy SciPy Docker HPC (SLURM)

Pinned Loading

  1. budgeted-search-ai4science budgeted-search-ai4science Public

    Python 2

  2. tesseract-hybrid-closure tesseract-hybrid-closure Public

    Differentiable hybrid closure for 2D turbulence: a PyTorch CNN trained a-posteriori through a JAX/Exponax spectral solver, composed across two served Tesseracts with reverse-mode gradients.

    Python 1

  3. tesseract-pinn-inverse-burgers tesseract-pinn-inverse-burgers Public

    Inverse parameter estimation for the viscous Burgers equation: a differentiable solver-adjoint, PINNs (JAX/PyTorch), and an amortized flow-matching posterior, packaged as swappable Tesseract compon…

    Python 4

  4. research-brief-agent research-brief-agent Public

    Streaming FastAPI agent that turns research questions into cited technical briefs over arXiv literature using Qdrant retrieval, full-text reading, and Anthropic/OpenAI-compatible tool use.

    Python 1

  5. hyperbolic-latent-vae hyperbolic-latent-vae Public

    Variational Autoencoder with non-euclidean (hyperbolic) latent space

    Python 15 2

  6. conv-vae-pytorch conv-vae-pytorch Public

    Convolutional Variational Autoencoder in Pytorch benchmarked on CelebA Dataset

    Jupyter Notebook 20 2