Jobs:Job-00258
Apply before: 21 December 2020
|
Posting:
Position: Postdoctoral Fellow
Apply before: 21 December 2020
Apply online:
|
The subfield of machine learning called “deep learning”, which uses neural networks with many layers of neurons, has shown great promise in many applications including water temperature prediction. However, deep learning is infamously fallible when predicting in out-of-bounds conditions. Very recently, hybrid models that blend deep learning with process knowledge have produced the most accurate predictions yet – with more robustness to out-of-bounds conditions – for lake water temperatures1,2 and stream temperature and discharge3. These early explorations of process-guided deep learning (PGDL) have opened the door to a new generation of predictive models, but many questions about the capabilities and limitations of such models remain unanswered, and many promising techniques for PGDL modeling remain untested.
Description of the Research Opportunity
We seek a postdoctoral fellow to advance process-guided deep learning for water temperature projections under scenarios of climate change and land use change. The study areas for the project are the Delaware River Basin (DRB), a USGS priority basin that supplies water to 16 million people, and the Neversink Watershed within DRB, where recreationally prized trout and endangered mussels are known to be temperature sensitive. A successful proposal will likely include some combination of model evaluation, development, and application. Above all, we encourage innovation and fundamental contributions to the sciences of water temperature prediction and process-guided deep learning. Specific angles and approaches of the candidate's proposal might include, but need not be limited to:
- Assess the reliability of PGDL-based temperature projections for out-of-bounds conditions such as very different climates or land uses.
- Assess model performance with respect to DRB-stakeholder-relevant metrics, e.g., accuracy in predicting exceedance of a temperature threshold.
- Assess the accuracy of one or more PGDL variants over different scenarios of basin size, process model resolution, and observation resolution.
- Experiment with uses of process-model outputs as inputs, observations, or a hybrid.
- Identify and incorporate new physical constraints or loss terms into the PGDL model.
- Refine hyperparameter tuning methods to improve model accuracy and avoid overfitting.
- Generate projections of water temperatures under scenarios of climate and/or land use change, with optional emphasis on extreme events such as drought and heat waves.
- Investigate projections to identify spatial and temporal patterns, locations of highly vulnerable human or aquatic communities, or locations where management intervention would be most cost effective.
- Apply machine learning interpretation techniques to derive new insights about the processes that control stream temperature patterns.
This position will be part of a DOE-funded collaboration titled “ExaSheds: Advancing Watershed System Science using Machine Learning-Assisted Simulation.” As such, resources available to the postdoc will include:
- Compiled, spatially referenced observations of water temperature throughout the DRB.
- Model inputs and temperature predictions from USGS’s PRMS-SNTemp model.
- Prepared model driver data sufficient to run DOE’s high-resolution model, Amanzi-ATS, in the * Neversink Watershed (available partway through postdoc).
- Personalized training at ORNL, and assistance from USGS staff, to run Amanzi-ATS in the Neversink Watershed using USGS’s Denali computing cluster.
Interested applicants are strongly encouraged to contact the Research Advisor early in the application process to discuss project ideas.
