In Phase1 of our convergence accelerator project, we developed and prototyped the concept for a web-based machine learning (ML) platform for hydrologic scenarios called HydroGEN. Our approach combines powerful physics-based simulations with ML and observations to provide customizable scenarios of watershed behavior from the bedrock through the treetops. We envision a platform where, without any prior modeling experience, water managers and planners can directly manipulate state-of-the-art tools to explore scenarios that matter to them. The primary outcome of our Phase 1 project was our successful Phase 2 proposal and pitch. All of our work from Phase 1 is being used as the starting point for Phase 2 where we are building out the national platform we proposed. 

In Phase 1, we brought together a diverse team of hydrologists, water managers, ML specialists, data scientists, software developers, and user experience specialists to develop and prototype this concept. We conducted 23 user interviews with water managers from the local to the national level, environmental consultants, forest managers and conservation groups. Our first set of interviews was designed to clarify potential user’s workflow needs, pain-points and understanding of machine learning (ML) concepts. Using this information, we developed our first conceptual designs for HydroGEN through a series of low fidelity prototyping activities. We used our initial wireframes for a second set of user interviews where we asked participants to “walk through” the process.  We gathered feedback on what was most valuable to them, and what would be needed to establish trust in our outputs. Finally, we refined and extended our initial design based on the feedback we received in round two. We shared the refined wireframes in a third series of user interviews to confirm that the product would be useful and to understand what we might be missing. 

In parallel with user interviews we also developed several prototype applications  to test our workflows and demonstrate proof-of-concept for our scientific approach. We prototyped our ML emulators on three problems of increasing complexity.  We started with a simple 2D domain  to generate our first proof-of-concept. This also served as an easy toy problem for the entire team to learn from and was converted into a stand-alone interactive educational app called SandTank-ML. Next, we prototyped a more complex 3D domain of two idealized hillslopes joined by a river (the so-called Tilted-V catchment used frequently in hydrologic benchmarking and intercomparison). Finally, in collaboration with our partners at Reclamation, we selected the Taylor River Watershed for our first real-world use case. The Taylor prototype illustrates that we are able to simulate extreme scenarios and provide spatially complete coverage for a range of hydrologic variables from the bedrock to the treetops, regardless of the availability of historical data

Last Modified: 06/29/2022
Modified by: Laura Condon