Driven by Francesco and Olaf, one step in surrogate modelling

 This is one of the paper to which I was more spectator than a real player. However, my name appears among the contributors because I was somewhat crucial for the project to succeed. The idea that bother many is that hydrological models aren too complicate and we need more simple models that get almost the same results. This is the rational for this work that was part of Francesco Serafin Ph.D. Thesis. 

The main ideas was to get into OMS3/CSIP framework an artificial neural network (ANN) system that could surrogate the hydrological models. Surrogate means that it can reproduce most of the dynamic features of the model training the ANN and then making using it. The roadmap is simple but the realization involves several steps that the paper delineates.

So, here it is our paper and its abstract: Serafin, Francesco, Olaf David, Jack R. Carlson, Timothy R. Green, and Riccardo Rigon. 2021. “Bridging Technology Transfer Boundaries: Integrated Cloud Services Deliver Results of Nonlinear Process Models as Surrogate Model Ensembles.” Environmental Modelling and Software[R], no. 105231 (October): 105231.

Environmental models are often essential to implement projects in planning, consulting and regulatory institutions. Research models are often poorly suited to such applications due to their complexity, data requirements, operational boundaries, and factors such as institutional capacities. This contribution enhances a modeling framework to help mitigate research model complexity, streamline data and parameter setup, reduce runtime, and improve model infrastructure efficiency. Using a surrogate modeling approach, we capture the intrinsic knowledge of a conceptual or process-based model into an ensemble of artificial neural networks. The enhanced modeling framework interacts with machine learning libraries to derive surrogate models for each model service. This process is secured using blockchain technology. After describing the methods and implementation, we present an example wherein hydrologic peak discharge provided by the curve number model is emulated with a surrogate model ensemble. The ensemble median values outperformed any individual surrogate model fit to the curve number model.