pvankatwyk/iseflow-datasets

ISEFlow Training Datasets

Training data for ISEFlow — a hybrid normalizing flow + deep ensemble emulator for ice sheet sea-level projections with full uncertainty quantification.

Model weights: pvankatwyk/ISEFlow GitHub: Brown-SciML/ise PyPI: ise-py

Developed at Brown University by Peter Van Katwyk.

---

Dataset Description

These datasets contain sector-averaged ISMIP6 climate forcings and ice sheet model projection outputs, processed and feature-engineered for use with ISEFlow. Both the Antarctic Ice Sheet (AIS) and Greenland Ice Sheet (GrIS) are covered, with train/validation/test splits and scikit-learn scalers included for full reproducibility.

Key Numbers

	AIS	GrIS
Sectors	18 (8 km resolution)	6 Rignot basins (5 km resolution)
AOGCMs	8 CMIP5	7 CMIP5
Projection period	2015–2100 (86 timesteps)	2015–2100 (86 timesteps)
Total rows	291,024	85,140
Train / Val / Test	191,436 / 41,022 / 41,022	54,610 / 11,696 / 11,782

---

Repository Structure

v1.1.0/
├── AIS/
│   ├── dataset.csv            # Full merged dataset (forcings + projections + ISM metadata)
│   ├── forcings.csv           # Sector-averaged atmospheric + ocean forcings
│   ├── projections.csv        # ISMIP6 projection outputs (IVAF → SLE)
│   ├── AIS_atmospheric.csv    # Atmospheric forcings only
│   ├── AIS_oceanic.csv        # Ocean forcings only
│   ├── train.csv              # Training split (feature-engineered, lag variables included)
│   ├── val.csv                # Validation split
│   ├── test.csv               # Test split
│   ├── ids.json               # Train/val/test simulation IDs for reproducibility
│   ├── scalers/
│   │   ├── scaler_X.pkl       # sklearn scaler fitted on training inputs
│   │   └── scaler_y.pkl       # sklearn scaler fitted on training targets
│   ├── process_sectors.py     # Reference script: raw ISMIP6 → sector-averaged CSV
│   └── feature_engineer.py   # Reference script: dataset.csv → train/val/test splits
└── GrIS/
    └── (same structure as AIS)

---

Input Features

AIS Atmospheric Forcings (anomalies vs. 1995–2014 sector climatology)

Variable	Description	Units
pr_anomaly	Precipitation anomaly	kg m⁻² s⁻¹
evspsbl_anomaly	Evaporation/sublimation anomaly	kg m⁻² s⁻¹
smb_anomaly	Surface mass balance anomaly	kg m⁻² s⁻¹
ts_anomaly	Surface temperature anomaly	K

AIS Ocean Forcings (absolute values)

Variable	Description	Units
thermal_forcing	Ocean thermal forcing	°C
salinity	Ocean salinity	PSU
temperature	Ocean temperature	°C

GrIS Atmospheric Forcings (anomalies vs. 1960–1989 MAR baseline)

Variable	Description
aSMB	Surface mass balance anomaly
aST	Surface temperature anomaly

GrIS Ocean Forcings (absolute values)

Variable	Description
thermal_forcing	Ocean thermal forcing
basin_runoff	Basin-integrated runoff

Each forcing variable is also included as lag features *.lag1 through *.lag5 in the train/val/test splits.

ISM Characteristics (one-hot encoded in train/val/test)

The train/val/test splits include one-hot encoded ice sheet model (ISM) configuration metadata merged from the ISMIP6 model characteristics table:

• AIS: numerics, stress balance, resolution, initialization method, melt parameterization, ice front treatment, ocean forcing type, ocean sensitivity, ice shelf fracture, melt param type
• GrIS: numerics, ice flow model, initialization method, SMB product, velocity/bed/surface datasets, geothermal heat flux source, resolution (min/max), ocean forcing type, ocean sensitivity, ice shelf fracture

Target

sle — sea-level equivalent contribution (mm), derived from IVAF (ice volume above flotation) change.

---

AOGCMs

AIS: ccsm4_rcp85, csiro-mk3.6_rcp85, hadgem2-es_rcp85, ipsl-cm5-mr_rcp26, ipsl-cm5-mr_rcp85, miroc-esm-chem_rcp85, noresm1-m_rcp26, noresm1-m_rcp85

GrIS: access1.3_rcp85, csiro-mk3.6_rcp85, hadgem2-es_rcp85, ipsl-cm5-mr_rcp85, miroc5_rcp26, miroc5_rcp85, noresm1-m_rcp85

---

Data Sources

• Ice sheet projections: ISMIP6 Phase 2 outputs (AIS) and ISMIP6 GrIS outputs; processed from Zenodo archives
• Atmospheric forcings: CMIP5 AOGCMs downscaled via ISMIP6 protocol, sector-averaged using ise.data.ISMIP6.process
• Ocean forcings: ISMIP6 ocean thermal forcing, salinity, and temperature (AIS); basin runoff (GrIS)
• ISM metadata: ISMIP6 model characteristics table

---

Feature Engineering

The train/val/test splits are generated from dataset.csv via ise.data.ISMIP6.feature_engineer.FeatureEngineer:

1. mrro_anomaly (runoff) dropped for AIS v1.1.0
2. ISM model characteristics merged and one-hot encoded
3. Inputs scaled with RobustScaler; targets scaled with StandardScaler
4. Lag variables added (lag=5 timesteps)
5. Outliers removed: SLE outside the [0.5%, 99.5%] quantile range excluded
6. Split 70/15/15 by simulation ID (not by row), preventing data leakage across splits

The ids.json file records the exact simulation IDs in each split for reproducibility.

---

Reproducing the Splits

from ise.data.ISMIP6.feature_engineer import FeatureEngineer
import pandas as pd

# AIS
fe = FeatureEngineer(
    ice_sheet='AIS',
    data=pd.read_csv('v1.1.0/AIS/dataset.csv'),
    fill_mrro_nans=None,
    split_dataset=False,
    output_directory=None,
)
fe.data = fe.data.drop(columns='mrro_anomaly')
fe.add_model_characteristics()
fe.scale_data(save_dir='v1.1.0/AIS/scalers/')
fe.add_lag_variables(lag=5)
fe.drop_outliers('quantile', 'sle', quantiles=[0.005, 0.995])
fe.split_data(train_size=0.7, val_size=0.15, test_size=0.15, output_directory='v1.1.0/AIS/')

---

Dataset Versions

Version	Ice Sheets	Notes
v1.1.0 (current)	AIS + GrIS	mrro_anomaly removed from AIS inputs; GrIS added

---

Citation

If you use these datasets, please cite:

@software{vankatwyk2026ise,
  author       = {Van Katwyk, Peter},
  title        = {{ISE}: {Ice Sheet Emulator}},
  year         = {2026},
  version      = {2.0.0},
  publisher    = {GitHub},
  url          = {https://github.com/Brown-SciML/ise},
}

And the associated paper:

@article{vankatwyk_iseflow,
  title   = {ISEFlow: A Flow-Based Neural Network Emulator for Improved Sea Level Projections and Uncertainty Quantification},
  author  = {Van Katwyk, Peter},
  year    = {2025},
}

Please also cite the underlying ISMIP6 data sources as appropriate.

---

License

MIT License. See LICENSE.md.

---

Contact

Peter Van Katwyk — pvankatwyk@gmail.com Brown University, Department of Earth, Environmental and Planetary Sciences