HiRO-ACE

The HiRO-ACE framework enables efficient generation of 3 km precipitation fields over decades of simulated climate and arbitrary regions of the globe. HiRO (High Resolution Output) is a diffusion model which generates downscaled fields at 3 km resolution from 100 km resolution inputs. The HiRO checkpoint included in this model generates 6-hourly averaged surface precipitation rates at 3 km resolution. The Ai2 Climate Emulator (ACE) is a family of models designed to simulate atmospheric variability from the time scale of days to centuries. For usage with the HiRO downscaling model, we include a checkpoint for ACE2S. Compared to previous ACE models, ACE2S uses an updated training procedure and can generate stochastic predictions. For more details, please see the accompanying HiRO-ACE paper linked below.

Quick links

• 📃 Paper
• 💻 Code
• 💬 Docs
• 📂 All ACE Models

Inference quickstart

1. Download this repository. Optionally, you can just download a subset of the forcing_data and initial_conditions for the period you are interested in.

2. Install code dependencies with pip install fme.

3. Update paths in the ACE inference config file ace2s_inference_config_global.yaml. Specifically, update experiment_dir, checkpoint_path, initial_condition.path and forcing_loader.dataset.path.

4. Update paths in the HiRO downscaling inference config file hiro_downscaling_ace2s_pnw_output.yaml. Specifically, update experiment_dir, model.checkpoint_path, and data.coarse. The directory path in data.coarse should point to the same ACE inference experiment_dir from step 3. Optionally, if you wish to change the region and/or time selection of the area(s) to downscale you may edit those in the downscaling config (see downscaling inference docs for more details). An example of global downscaling is also provided in hiro_downscaling_ace2s_global_output.yaml.

5. Run the script run-hiro-ace.sh.

Strengths and weaknesses

ACE2S

The strengths of ACE2S are:

• stochastic (generative) emulator of 100km coarsened X-SHiELD
• improved small scale variability compared to deterministic ACE2 (especially surface precipitation)
• exact conservation of global dry air mass and moisture
• low time-mean biases for most variables relative to the 10-year training data available
• compatible with HiRO model with no observed degradation

Some known weaknesses of ACE2S are:

• trained only on 9 years of X-SHiELD (2014-2022)
• not expected to generalize outside of the limited forcing conditions used during training (e.g., past/future SSTs and CO2)
• may not produce statistically independent and representative samples of the base X-SHiELD model when run for significantly longer than 9 years
• not suitable for weather forecasting (e.g., not trained on reanalysis data only climate model output)
• model was not tested as rigorously as ACE2-ERA5 (e.g., cannot comment on stratospheric variability, etc)
• overestimation of tropical cyclone generation compared to X-SHiELD
• some aspects of the pre-training and fine-tuning methodologies did not have ablations
• exactly pre-training and fine-tuning methodologies are subject to potential significant changes for future versions

HiRO

The strengths of HiRO are:

• Quickly generates surface precipitation rates almost anywhere on the globe analogous to X-SHiELD 3 km output
• Reproduces surface precipitation rate distribution of X-SHiELD out to the 99.99th percentile
• Low time-mean biases against X-SHiELD
• Recovers time-mean characteristics of precipitation related to topography
• Stochastic for easy ensemble generation of small-scale variability

Some known weaknesses of HiRO are:

• Not expected to generalize to climates outside of the 10-year training dataset
• Only produces precipitation features representative of X-SHiELD data, not observations
• May not produce physically consistent precipitation features; different precipitation regimes (convective, stratiform, orographic) have not been rigorously analyzed separately
• Does not tend to produce the strongest isolated convection over tropical ocean regions of X-SHiELD (e.g., representative of 99.9999th percentile of X-SHiELD outputs)
• Not trained or tested outside of 66S -- 70N, and cannot run outside of 88S or 88N
• Downscaling regions larger than the 16x16 1-degree patch size used for training will have discontinuities for patch overlaps of 0 or some blending artifacting from averaging in the overlap region for overlaps > 0

License

This model is licensed under Apache 2.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.