Two papers from the VCM team have recently been published in peer-reviewed scientific journals.  

The first paper describes VCM's 'nudge-to-observations' machine-learning methodology for improving weather forecasts and the simulation of precipitation patterns by computer models of the global atmosphere:  

Watt-Meyer, O., et al, 2021: Correcting weather and climate models by machine learning nudged historical simulations. Geophys. Res. Lett., 48, e2021GL092555.

Read the first paper

The second paper describes VCM's 'Python wrapper'. This enables the VCM team to use Python to run and modify the U.S. global weather forecast model, FV3GFS, which is written in Fortran:

McGibbon, J., et al, 2021:  fv3gfs-wrapper: a Python wrapper of the FV3GFS atmospheric model. Geophys. Model Dev.. 14, 4401-4409.

Oliver Fuhrer, Jeremy McGibbon and Johann Dahm presented the DSL team’s work in a seminar at Argonne National Laboratory. The talk focused on how a domain-specific compiler in Python is implemented and how the internal mechanics of translating Python code written by a climate scientist gets lowered through the toolchain into efficient code tailored for a specific target hardware. Learn more at https://www.evs.anl.gov/events/2021/06-09-fuhrer-mcgibbon-seminar.cfm
 

Oliver Elbert presented the DSL team's work at NOAA's General Modeling Meeting and Fair, a conference for NOAA employees and collaborators to share the latest developments in modeling and forecasting technology. Oliver showcased the design and motivations for our domain specific language, and demonstrating the code to a group of NOAA climate scientists.

Lucas Harris from NOAA / GFDL presented a Unified Forecast System (UFS) Webinar talk co-authered with Oliver Fuhrer highlighting status and results from the efforts of VCM to port the FV3 dynamical core using a domain-specific language (DSL) in Python. The title of the talk was FV3 on GPUs: Two approaches to accelerate global cloud resolving modeling.

Watch the talk

Rhea George presented a talk titled Separating physics and performance: Using Python to implement fast and maintainable climate models on VCM DSL activities at the virtual UCAR Software Engineering Assembly (SEA) Improving Scientifc Software Conference 2021 which is a venue targeted at an international audience of atmospheric scientists interested in best practices for software engineering.

VCM is committed to open-source development and we have been preparing and documenting our Github repositories to share with the broader FV3GFS and climate modeling communities.  Here is an entry point to this living foundation of both our DSL and ML software engineering.

Later this year VCM will transfer from Vulcan into the Allen Institute for Artificial Intelligence (AI2) in Seattle, along with the rest of Vulcan’s products teams. AI2 has a strong intellectual culture of top-quality product oriented research working collaboratively toward societal impact that aligns well with the climate modeling team's approach.

Chris Bretherton presented a NOAA Unified Forecast System (UFS) Webinar talk on behalf of the VCM ML group titled Improving weather forecast skill and rainfall climatology of FV3GFS using machine learning.

Oli Watt-Meyer and Noah Brenowitz presented a VCM weather-climate briefing to Vulcan Products titled La Niña and Its Impacts on the Pacific Northwest.

Chris Bretherton presented a talk on VCM ML activities in a virtual talk series run by U. Illinois and Texas A&M on Equilibrium Climate Sensitivity, which reaches an international audience of climate scientists.

Langwen Huang from ETH Zurich successfully finished his thesis in the VCM DSL group on “Improving the performance of a domain-specific language compiler”.

Jeremy McGibbon, Noah D. Brenowitz, Mark Cheeseman, Spencer K. Clark, Johann Dahm, Eddie Davis, Oliver D. Elbert, Rhea C. George, Lucas M. Harris, Brian Henn, Anna Kwa, W. Andre Perkins, Oliver Watt-Meyer, Tobias Wicky, Christopher S. Bretherton, and Oliver Fuhrer, 2021: fv3gfs-wrapper: a Python wrapper of the FV3GFS atmospheric model, submitted to the journal Geoscientific Model Development.  

VCM organized an FV3GFS Python Wrapper workshop for GFDL, led by Jeremy McGibbon and Spencer Clark with help from Lucas Harris of GFDL. Ten GFDL and Princeton scientists enjoyed hands-on tutorial exercises demonstrating wrapper capabilities including running a machine-learning augmented version of FV3GFS from an iPython notebook on their personal computers. 

It describes the ML group’s corrective machine learning methodology to make weather forecasts and mean precipitation patterns simulated by a global atmospheric model better match an observational analysis: Watt-Meyer, O., N. Brenowitz, S. K. Clark, B. Henn, A. Kwa, J. McGibbon, W. A. Perkins, and C. S.  Bretherton, 2021:  Correcting weather and climate models by machine learning nudged historical simulations.  Submitted to Geophysical Research Letters. 

Mikael Stellio from ETH Zurich successfully finished his thesis in the VCM DSL group on “Application of a DSL to Weather and Climate Modeling” where he ported GFDL’s microphysics parameterization to GPUs using the GT4Py DSL.

At the American Meteorological Society (AMS) Annual Meeting, VCM gave six presentations (led by Andre Perkins, Jeremy McGibbon, Johann Dahm, Rhea George, Tobias Wicky, and Oliver Fuhrer). Oliver Fuhrer gave a motivational talk about why VCM is working on DSLs for weather and climate. Jeremy McGibbon and the VCM ML group presented a talk at the American Meteorological Society Symposium on Artificial Intelligence: Portable Python-wrapped FV3GFS atmospheric model.

Watch Jeremy McGibbon's talk

Oli Watt-Meyer and the VCM ML team presented a talk titled Bias correcting weather models with machine learning.

Noah Brenowitz and the VCM ML team presented a talk titled Machine Learning Climate Model Dynamics: Offline versus Online Performance.

At the American Geophysical Union (AGU) Fall Meeting, the VCM ML group gave six presentations (led by Anna Kwa, Brian Henn, Oli Watt Meyer, Spencer Clark, Jeremy McGibbon, and Chris Bretherton). Noah Brenowitz and Chris Bretherton were also co-chairs of an AGU session on Machine Learning for Weather and Climate. Oliver Fuhrer was co-chair of an AGU session on addressing challenges for next generation Earth system models.

Spencer Clark and the VCM ML team gave a talk titled Challenges associated with training a machine learning based moist physics parameterization by coarse-graining in a model with topography.

VCM co-organized a virtual workshop together with Lucas Harris from NOAA with over 25 participants from NOAA, NASA and DoE. The workshop introduced the GT4Py DSL with hands-on tutorials running in Jupyter notebooks on the Orion HPC cluster.

Christopher Kung, a collaborator of VCM DSL from NASA’s Global Modeling and Assimilation Office (GMAO), presented the GT4Py DSL project in a virtual exhibit at the Supercomputing Conference 2020.

With the move, the VCM DSL team goes entirely open for both development and release of their software products.