This is the "fully-grown" dragon I've been talking about. Finally online with Open Access. A new format of "differentiable" models that is, simultaneously:
process-based, learnable, regionalized, highly-efficient, high-performance, well-generalized, and multiphysics.Cover page
My most exciting work so far on systematic integration of deep learning and physical model is published in WRR.
dx.doi.org/10.1029/2022WR032… To summary, now we have a regionalized physical model with similar performance to pure data-driven LSTM, achieving NSE median 0.732 in CAMELS.