Speeding up realistic ecohydrological modeling with machine learning emulation
Congratulations to Octavia Crompton whose final PhD thesis product "Sensitivity of dryland vegetation patterns to storm characteristics" has been accepted for publication in Ecohydrology! Tavia's study overcomes a computational challenge in making ecohydrological models more realistic -- namely that the processes of interest span a really wide range of timescales - from minutes for runoff within storms, to decades, for plants to grow and disperse and spatial patterns to emerge.
It's almost impossible to accommodate processes at all these timescales without breaking computational models (Tavia's "truth" simulations, which did this, required 6 months of super computer time!). Emulation of physical processes with a simpler model (in this case a machine learning model) can greatly speed up the fast-timescale processes which have to be simulated many times in a model that spans decades/centuries. Tavia's paper demonstrates that this kind of emulation makes it possible to ask new questions about how fast timescale processes lead to long timescale outcomes. For example, the figures on the left show how different storm properties lead to different vegetation patterns (green bands are vegetated, white are not) in arid ecosystems, even if the same total amount of rainfall occurs in each case.