In the last ten years I focused on building a reliable system for doing hydrology by computer. This systems learns from the implementation of the process-based GEOtop and is based on the framework developed by ARS/USDA called OMS3. The new system is called GEOframe .
I have been concerned with the fact that many results claimed on the basis of computer simulations were, in fact, not properly reliable and verifiable, do to lack of software engineering, description of the internals of the tools and availability to researchers. Moreover, also the goodness experimental science is quite dependent on the capability and reliability of models. In my experience with the model GEOtop so far, when there have been discrepancies between data and the model, most of the time the model was right and the experiment imprecise. Sometimes though was the model GEOtop (or other that we used) wrong and we worked to improve it. Models have to be robust, reliable, realistic and their results reproducible (R4). The new system GEOframe, which is built on these premises, is not a model though. It can fit several modelling solutions and it is actually is agnostic with respect to the methods. It is designed to offer a platform to compare different modelling strategies, lumped modelling, process-grid-based modelling and whatever but avoiding to redo any time the unnecessary.
We used some ML technique in the process of calibration and we implemented also a ANN framework (in OMS3) but that part is underdeveloped at that moment. (If we expand to much, we bleed).
The older part of the GEOframe system contains lumped based types of models. However, we have recently implemented a solver of Richards 1D [0] and 2D (paper in writing) and 3D (software in deployment). The latter implement a new algorithm for integration of non linear PDE systems which always converges and can naturally switch between groundwater, vadose zone and surface water. Soil can be hot, warm and frozen without problems (the latter is in deployment). Besides this I worked underground in having a better estimation of evaporation and transpiration. I published very little with respect the amount of work I did on these subjects, but disentangling the theory, the misconceptions, the scale issue was (is) not very easy, and took its time. I have a first (not completely satisfactory paper, from my point of view, on this topic, just published on Water [1] , but better ones are going to be written in the Fall 2021 and in the 2022. Finally I worked on disentagling the theory of travel time residence time. We have some paper on it, since 2016 [2,3,4,5] which are also connected to a new way to categorize and, before of it, representing lumped-semi-distributed models [6] in order to be able to produce some quantitative reasoning about models structure. I did not pursued very much applied work but with GEOframe growing we were able to produce some nice applications on the Posina catchment [6] (~110 km2), Blue Nile (~175000 km2) [7]. These applications, could be the basis for a comparison of traditional and ML methods. We have also ongoing the modelling of the largest river basin in Italy, the Po river (~75000km2) and the Nera catchment (closed at some hundreds of square kilometers), being very interesting because affected by karst. Those latter two catchments could be possible candidates for applying ML techniques and doing performance comparisons, once we have collected the appropriate data. Notably in the last work (actually since the implementation of GEOtop) working on the catchment meant for me working on the water budget of which the discharge is just one element completed by evaporation, transpiration and heat turbulent transfer. I viewed the use of the energy budget necessary to describe irrigation needs by crops and vegetation in this changing climate era, and in general as a tool to support a more complete view of the hydrological cycle.
