The National Science Foundation-funded Learning the Earth with Artificial intelligence and Physics (LEAP) Science and Technology Center (STC),, a large multi-institutional center effort meant to improve climate projections using novel artificial intelligence for better climate adaptation, invites applications for three Postdoctoral Research Scientist positions in the fields of climate science and data science.

In your application materials, please indicate to which project you are applying, or rank your preference if you are interested in applying for more than one. Applications will be reviewed on a rolling basis, but applications received by April 15th are guaranteed to receive consideration.

• Project 1: Attributing Climate Model Divergence with Hierarchical Representation Transformers

PI: David John Gagne; co-Is: Michael Pritchard, Carl Vondrick

Location: NSF NCAR, Boulder, Colorado

• The goal of this project is to develop an automated approach to identify systemic differences between climate simulations across different spatial scales and attribute those differences to particular model components. One of the most challenging and time-consuming aspects of climate model development is determining how adding new components to the modeling system changed the model climate. While some changes can be quite obvious and significant, others are much more subtle and are harder to attribute to a particular interaction within the model. In some instances, climate scientists have spent years tracking down the source of these changes, delaying the deployment of new models or interfering with the results from expensive runs. For this project, we intend to utilize recent advances in encoding multiscale data and learning relationships with newer transformer models and machine learning tasks that compare two items to accomplish this task. The transformer models will be trained to distinguish different forcing scenarios, and then we will use explainable AI (XAI) attribution methods to identify the model inputs that most affected the predictions. We will examine how the choice of encoding and the scale of the XAI perturbation affect the resulting attribution and how closely the attribution is associated with changes in forcing mechanisms. If successful, this approach could speed up the debugging and iteration process of incorporating new model components (ML or physics-based) into CESM and other Earth System Models, enabling a leap ahead in model development.

• The postdoc will be supervised by PI Gagne (primary supervisor; NSF NCAR/LEAP) and will benefit from mentoring by Profs. Pritchard and Vondrick as well as the NSF NCAR Machine Integration and Learning for Earth Systems (MILES) group. It is expected that the postdoc has a prior background in machine learning with significant experience with a major deep learning framework along with the ability to set independent goals, work with interdisciplinary teams, and communicate clearly. Prior climate science experience and/or a willingness to learn about climate science is strongly encouraged.

• Project 2: The Metrics Reloaded: Improved similarity assessment for climate maps

PI: Viviana Acquaviva; co-Is: Sara Shamekh, Duncan Watson-Parris

Location: Columbia University, New York, NY

Recent improvements in climate modeling and machine learning methods give us more opportunities to reduce uncertainties in future climate projections, which is crucial to planning for adaptation and mitigation measures. One important consideration is assessment of output similarity between models and data, or among different models. Various classic measure of error, such as the mean square/absolute error (MSE/MAE) of differences between cell values in a gridded map, have been used as a summary statistic of relevance, but they might not be suitable to capture the complexities of maps, where different spatial and temporal scales may be at play. We propose to develop, test, and validate improved metrics for climate models, beginning from assessing differences in static maps. We formulate a plan to use improved metrics beyond model skill assessment, as a foundation for dimensionality reduction, model comparison, equation discovery, and visualization purposes.

The postdoc will be supervised by PI Acquaviva (primary supervisor; CUNY/LEAP) and also benefit from mentoring from Profs. Shamekh (NYU) and Watson-Parris (UCSD). It is expected that the postdoc will have a genuine curiosity and interest for data exploration and analysis, the ability to set independent goals and communicate clearly, and a background in either climate science or data science.

• Project 3: Deep convection emulation and role of cloud organization

PI: Pierre Gentine; co-Is: Stephan M Mandt, Mike Pritchard

Location: Columbia University, New York, NY

Deep convection is one of the major sources of uncertainties in climate models and especially to constrain the hydrologic cycle. The postdoc will work on the development of a new deep convective algorithm based on deep learning, in the Community Earth System Model (CESM). The algorithm will include the role of convective aggregation and evaluate its impact on regional to local scale. The postdoc will also develop new theoretical developments to understand the stability of hybrid (physics and machine learning) models. The work will use high-resolution storm resolving simulations as well as observational products to define a state-of-the-art convective parameterization. Some previous experience with either convective parameterizations or analysis of deep convection would be ideal, yet not required.

The postdoc will be supervised by PI Pierre Gentine and collaborate with Profs. Mike Protichard and Stephan M Mandt at UC Irvine.

Commitment to Diversity

One of LEAP's goals is to increase the diversity in climate science and data science. We welcome and encourage applications from individuals of all backgrounds and identities. We are committed to building a diverse and inclusive community and believe that a variety of perspectives and experiences is essential to advancing our research and mission.

Minimum Qualifications

• A PhD. in Data Science, Computer Science, Physics, Earth System Science or a directly related discipline is required by the start of the appointment.

• Strong programming skills.

• Excellent command of the English language (oral and written).

Preferred Qualifications

• Fluency in Python.

• Advanced experience with machine learning/deep learning algorithms and libraries.

• Experience in statistical/mathematical analyses of model output and/or observational datasets.

• Strong communication skills.

• Experience collaborating on interdisciplinary teams.

Columbia University is an Equal Opportunity Employer / Disability / Veteran

Pay Transparency Disclosure

The salary of the finalist selected for this role will be set based on a variety of factors, including but not limited to departmental budgets, qualifications, experience, education, licenses, specialty, and training. The above hiring range represents the University's good faith and reasonable estimate of the range of possible compensation at the time of posting.

Equal Opportunity Employer / Disability / Veteran

Columbia University is committed to the hiring of qualified local residents.

Minimum Salary: 31200.00 Maximum Salary: 31200.00 Salary Unit: Yearly