# CSU Synthetic Attribution Benchmark for Explainable AI in Geoscience

## Facts

- Provider: Mamalakis, Ebert-Uphoff and Barnes (Colorado State University, CIRA)
- License: CC-BY-4.0 (https://radiantearth.blob.core.windows.net/mlhub/csu-synthetic-attribution/README.md)
- Format: JSON (in sampled files)
- Stored at: Azure Blob Storage, https://radiantearth.blob.core.windows.net/mlhub/csu-synthetic-attribution
- Access: public, readable without an account
- DOI: https://doi.org/10.34911/rdnt.8snx6c
- Category: Machine Learning & AI
- Source: https://doi.org/10.1017/eds.2022.7
- Page: https://scigantic.com/archive/ce91c20b-09f1-57e1-a9f7-95a5034f14cc
- Keywords: Machine Learning & AI, benchmark, JSON, Azure Blob Storage

## About

A synthetic benchmark for testing explainable AI (XAI) attribution methods on a geoscience-style regression problem, built at Colorado State University and the Cooperative Institute for Research in the Atmosphere. The inputs X are a series of 2D global fields drawn from a multivariate normal distribution, standing in for monthly sea surface temperature anomalies on a 1 degree grid (89.5N to 89.5S, 0.5E to 359.5E). The output Y is a scalar series produced by a known nonlinear function F of X, standing in for a regional climate index such as seasonal temperature or precipitation. Because F is known, the ground-truth attribution of every output value to every input grid cell is supplied alongside X and Y, so the faithfulness of saliency, gradient, Integrated Gradients, LRP, SHAP and other heat-map methods can be scored exactly after training a neural network. Measured 2026-11-15: one NetCDF file, synth_exm_data.nc (10.4 GB), plus a STAC collection JSON and a PDF documentation file; open it with xarray. Released under CC BY 4.0; cite Mamalakis, Ebert-Uphoff and Barnes (2022), Environmental Data Science, doi:10.1017/eds.2022.7, and the Radiant MLHub dataset DOI.

## Load the data

```python
import adlfs
fs = adlfs.AzureBlobFileSystem(account_name='radiantearth', anon=True)
fs.ls('mlhub/csu-synthetic-attribution/')[:10]
```

Lists the files with anonymous access. Runs on any machine, with no cloud account.

## Data structure

File types in sampled files: JSON

Example file: `.source/metadata.json`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/ce91c20b-09f1-57e1-a9f7-95a5034f14cc
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/ce91c20b-09f1-57e1-a9f7-95a5034f14cc
- Over MCP: `get_schema_card` with id `ce91c20b-09f1-57e1-a9f7-95a5034f14cc` at https://api.scigantic.com/mcp

## Related topics

- [JSON datasets](https://scigantic.com/topics/json)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Machine Learning & AI datasets](https://scigantic.com/topics/machine-learning-and-ai)
