I build numerical and machine-learning models, and I care most about how they perform in speed and fidelity on GPUs and HPC clusters. Nine years of high-performance computing, four years of machine learning.
Most recently a postdoc at TU Ilmenau, building reservoir-computing frameworks to improve prediction of chaotic dynamical systems. That work included a GPU-batched implementation that computes a reservoir's information processing capacity nearly 20 times faster than the CPU version. Before that, a PhD in rotating fluid turbulence built on shell models and direct numerical simulations, with detours into astrophysics and quantum optics.
Beyond accuracy, what makes a model usable is whether its results reproduce, whether its evaluation holds up, what it costs to run, and whether its predictions can be explained. The projects below are where that shows up in practice. See Resume for the full research and employment history.
Scheduling maintenance too early wastes good parts; too late risks unplanned downtime. Benchmarked a gradient-boosting baseline against a classical Echo State Network and NG-RC (next-generation reservoir computing) on NASA's hardest aircraft-engine degradation dataset (C-MAPSS FD004) to predict remaining useful life.
Result: reservoir computing is a simple, fast, and competitive alternative here. A classical Echo State Network uses a fixed random reservoir and a linear readout, with no backpropagation. It trains in 2.47s on a desktop CPU and outscores four of the five recently published results it is compared against, trailing only a GPU-trained TCN. NG-RC trains in 0.41s and is fully deterministic; I found no prior published application of it to this problem.
A classical Echo State Network trains in 2.47s on a desktop CPU and beats a tuned gradient-boosting baseline by nearly three orders of magnitude on NASA's hardest turbofan benchmark, outscoring four of the five recently published results it is compared against.