1 of 14

Artificial Intelligence

Pre-Summit Intensive

Maria J. Molina (U. of Maryland) and Kelvin Droegemeier (UIUC)

Participants: Francis (Frank) Alexander (Argonne National Laboratory), Dee A Bates (UIUC), Christopher S. Bretherton (Allen AI Institute), Matthew Chantry (ECMWF), Hristo Chipilski (FSU), Peter Dueben (ECMWF), Dale Durran (UW), Pedram Hassanzadeh (U. of Chicago), Daniel S Katz (UIUC), Volodymyr Kindratenko (UIUC), Christian Lessig Otto-von-Guericke (Universität Magdeburg), Ruby Leung (PNNL), John Shalf (LBNL), Maike Sonnewald (UC Davis), Duncan Watson-Parris (UCSD), and Oliver Watt-Meyer (Allen AI Institute).

2 of 14

2) km-scale & AI Community Views

1) What We Know & (Generally) Agree On

3) Opportunities & Research Questions

4) Potential Issues on the Horizon

Images: Adobe Stock

3 of 14

1) What We Know and (Generally) Agree On

Community organization

is needed

Europe’s Destination Earth (DestinE) is organizing the community.

AI-based NWP models already are skillful (e.g., extremes, ensembles).

Daily simulations at 1-km (not fully coupled ESM) show benefits (e.g., extremes).

Keeping the dynamical core intact and replacing physical parametrizations with ML has shown success.

A fully AI-based model is possible (i.e., bypassing data assimilation).

4 of 14

1) What We Know and (Generally) Agree On

Addressing stakeholder needs in addition to pursuing scientific questions is important.

Researchers are practitioners as well. Ease of use is important for a computational system.

5 of 14

1) What We Know and (Generally) Agree On

AI in a km-scale climate computer

Observations at similar scale are needed.

Efficiency of parameterizations is not a priority – more potential lies in improving their accuracy.

Replacing the dynamical core with AI does not appear to offer most benefit currently.

From an ethics and equity perspective – where do we lack observations? AI can help with “optimal” sensor placement.

6 of 14

1) What We Know and (Generally) Agree On

AI in a km-scale climate computer

Observations at similar scale are needed.

Efficiency of parameterizations is not a priority – more potential lies in improving their accuracy.

Replacing the dynamical core with AI does not appear to offer most benefit currently.

From an ethics and equity perspective – where do we lack observations? AI can help with “optimal” sensor placement.

7 of 14

2) Views Between the km-Scale and AI Communities

AI can downscale to km-scale.

Users care about local scales.

km-scale can capture extremes.

AI-based models can predict extremes.

Physics-informed frameworks within AI are important.

Data-driven NWP at times exceeds the skill of physics-based models.

8 of 14

Do physics-based scale dependencies also apply in fully AI models?

Can AI models capture the fastest growing modes in ensembles?

3) Opportunities and Open Research Questions

What is the “optimal” number of ensemble members?

Do AI-based and physics-based models have different predictability?

Why can some AI-based models extrapolate to extremes?

9 of 14

Pair AI-based ensembles with established ensemble assimilation systems.

AI for more efficient calibration of existing parameterizations.

AI for interaction, exploration, & interpolation of km-scale simulations.

3) Opportunities and Open Research Questions

Empirical evidence of the need for physics-informed AI is needed.

AI-based explicit generation of thunderstorms in downscaling.

10 of 14

Concrete decisions regarding ML/AI might be superseded.

4) Potential Issues on the Horizon

11 of 14

Timescale(s) of relevance for tasks should be established.

4) Potential Issues on the Horizon

12 of 14

Careful consideration needs to be given to model validation.

4) Potential Issues on the Horizon

13 of 14

Data storage and access to the broader community is an important consideration.

4) Potential Issues on the Horizon

14 of 14

Artificial Intelligence Pre-Summit Intensive

Maria J. Molina (U. of Maryland) and Kelvin Droegemeier (UIUC)

Participants: Francis (Frank) Alexander (Argonne National Laboratory), Dee A Bates (UIUC), Christopher S. Bretherton (Allen AI Institute), Matthew Chantry (ECMWF), Hristo Chipilski (FSU), Peter Dueben (ECMWF), Dale Durran (UW), Pedram Hassanzadeh (U. of Chicago), Daniel S Katz (UIUC), Volodymyr Kindratenko (UIUC), Christian Lessig Otto-von-Guericke (Universität Magdeburg), Ruby Leung (PNNL), John Shalf (LBNL), Maike Sonnewald (UC Davis), Duncan Watson-Parris (UCSD), and Oliver Watt-Meyer (Allen AI Institute).