Research software for noisy data and stiff equations.
I'm Georgios Vourvachakis (Vakis), an independent researcher working across probabilistic machine learning, stochastic simulation, and numerical methods. I completed my MSc in Materials Science & Engineering (Computational Materials Science) at the University of Crete, with research at IACM and IESL, FORTH.
- TimeCausalVQVAE — Extending Time-Causal VAE baselines with causal VQ/RVQ tokenizers, autoregressive token priors, and path-risk diagnostics for market generation. Python package.
- exponax-spde — Stochastic Allen–Cahn solvers built on Exponax, with exponential Euler–Maruyama integration, Q-Wiener noise, ensemble utilities, and numerical validation in JAX.
Surrogate Modelling for Damage Threshold Estimation in Ultrashort Laser-Metal Interactions
Developed linear and nonlinear two-temperature-model solvers, including an IMEX Heun–Crank–Nicolson scheme with a strong Gauss–Seidel predictor, supported by stiffness and convergence studies. Built machine-learning surrogates for damage-threshold estimation.
Read the thesis · Thesis and presentation source
I'm particularly interested in neural differential equations, stochastic flows, and generative modeling for financial time series.
🎨 Interpretation
A cubist representation of Microsoft Clippy observing a hierarchical U-Net encoder-decoder model compressing state spaces. Generated with Nano Banana 2.1.
- Architecture Topology: The upper quadrant traces hierarchical compression, a central trapezoidal bottleneck, and symmetric skip connections.
- Stochastic Integration: The lower quadrant simulates continuous time-series forecasting via rough SDE trajectories and path-wise confidence bands exhibiting gradual variance reduction (CIR-type volatility decay).
- Spatial Mesh & Palette: Rendered in the Seaborn
makocolor space over a subtle Voronoi and finite element discretization mesh.