This rambling account may not do justice to the original research, but I found it just another example of how naive ML correlations find it impossible/difficult

This rambling account may not do justice to the original research, but I found it just another example of how naive ML correlations find it impossible/difficult/laborious to re-discover basic platitudes of a specialist field. If you discretize features from a continuous domain, like space-time chunking of "afternoon rain", then it will degenerate into a hybrid CNN-RNN ML system. Wet soil in the morning will depend on "overnight rain" and "morning dew". Then "overnight rain" might suppress afternoon rain, if it drains moisture from the atmosphere etc. etc. There will be various spatial and temporal influences expressed as Bayesian priors. One fact that I found surprising when I first heard it, but within 10 seconds realized was obvious, is that water is lighter than air (N2 O2 v. H2O). I imagined that clouds need convection currents from hot land causing temperature gradients, for moisture to rise into cooler layers, then condense and rain. But water (MW 18 g/mol) is lighter than air (MW 29 g/mol), so it just rises and rains (unless there is a temperature inversion anti-convection barrier). I imagine many such 'obvious' facts of fluid dynamics that elude CS ML hammer-nail practitioners.