Henok Tenaw Moges

AI for Dynamical Systems

Scientific Machine Learning  ·  Nonlinear Dynamics  ·  Spatio-Temporal Graph Learning

Scholar GitHub LinkedIn Email
Henok Tenaw Moges

“When does physical structure help learning in dynamical systems; and when does it fail?”

I am a Research Fellow in AI Systems at the University of Cape Town. My work sits at the intersection of machine learning, numerical analysis, and nonlinear dynamics, with a focus on how forecasting models behave under instability, coupling, and chaos.

A central thread of my research concerns spatio-temporal graph neural networks (STGNNs): when do spatial modules genuinely help, and how do graph-based architectures perform across different dynamical regimes? I study these questions through controlled benchmark design, architectural analysis, and principled evaluation of predictability limits in complex systems.

My recent work includes ChaosNetBench, a synthetic benchmark and evaluation framework built on a lattice of coupled standard maps with tunable chaos and coupling, designed to compare STGNN architectures across controlled dynamical regimes. My broader work spans efficient STGNN design, spatial attribution analysis, chaos indicators, and anomalous diffusion in Hamiltonian systems.

Spatio-Temporal Graph Learning
Scientific Machine Learning
Nonlinear Dynamics & Chaos
Controlled Benchmark Design
Efficient Architecture Design
Predictability in Complex Systems

Current Roles

Research Fellow — AI Systems
Artificial Intelligence Research Unit (AIRU), University of Cape Town
Work Package Lead — AI & Modelling
Mi-Hy EU Project
AI-driven modelling, simulation, and digital twin design for microbial hydroponics systems

Selected Projects

ChaosNetBench

Synthetic benchmark dataset and evaluation framework for studying STGNNs under controlled chaotic lattice dynamics. Built on coupled standard maps with tunable local chaos (K), coupling strength (ε), and system size (N). Evaluation across 13 architectures reveals regime-dependent differences between STGNN and non-graph forecasting models.

arXiv Dataset Code

Spatial Attribution for Long-Term Forecasting

Attribution study quantifying how much spatial modules contribute to long-term multivariate time series forecasting. The analysis compares spatial and non-spatial components across standard benchmarks and controlled coupling settings. Accepted at SAICSIT 2026; proceedings in press.

Code

Lite-STGNN

Lightweight STGNN for long-term multivariate forecasting. The model combines decomposition-based temporal modeling with learnable sparse graph structure. Published in the ICAART 2026 proceedings.

Paper Code

Selected Publications

Full list on Google Scholar.

Talks & Presentations

Academic Service