Physically-constrained ML projections: Difference between revisions

Hybrid modelling<ref>{{Cite journal|last=Reichstein|first=Markus|last2=Camps-Valls|first2=Gustau|last3=Stevens|first3=Bjorn|last4=Jung|first4=Martin|last5=Denzler|first5=Joachim|last6=Carvalhais|first6=Nuno|last7=Prabhat|date=2019-02|title=Deep learning and process understanding for data-driven Earth system science|url=https://www.nature.com/articles/s41586-019-0912-1.|journal=Nature|language=en|volume=566|issue=7743|pages=195–204|doi=10.1038/s41586-019-0912-1|issn=1476-4687}}</ref>, by incorporating physical-constraints into data-driven ML or deep learning models is a promising field of leveraging the large amounts of data available from observational products, while making use of physical constraints present in the climate system, to ensure robust projections and extrapolating well outside of the training [data<ref>{{Cite journal|last=Zhao|first=Wen Li|last2=Gentine|first2=Pierre|last3=Reichstein|first3=Markus|last4=Zhang|first4=Yao|last5=Zhou|first5=Sha|last6=Wen|first6=Yeqiang|last7=Lin|first7=Changjie|last8=Li|first8=Xi|last9=Qiu|first9=Guo Yu|date=2019|title=Physics-Constrained Machine Learning of Evapotranspiration|url=https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2019GL085291|journal=Geophysical data]Research Letters|language=en|volume=46|issue=24|pages=14496–14507|doi=10.1029/2019GL085291|issn=1944-8007}}</ref>. The output from physics-driven GCM climate models can be used for a "perfect model test" of the ML models, before the ML model is applied to make projections based on the observations<ref>{{Cite journal|last=Schlund|first=Manuel|last2=Eyring|first2=Veronika|last3=Camps‐Valls|first3=Gustau|last4=Friedlingstein|first4=Pierre|last5=Gentine|first5=Pierre|last6=Reichstein|first6=Markus|date=2020|title=Constraining Uncertainty in Projected Gross Primary Production With Machine Learning|url=https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2019JG005619|journal=Journal of Geophysical Research: Biogeosciences|language=en|volume=125|issue=11|pages=e2019JG005619|doi=10.1029/2019JG005619|issn=2169-8961}}</ref>.