CREDIT Documentation

Welcome to the documentation for CREDIT, the NSF NCAR Community Research Earth Digital Intelligence Twin project. CREDIT is an open foundational research platform for building machine learning Earth system prediction emulators. It is developed and maintained primarily by the NSF NCAR Machine Integration and Learning for Earth Systems (MILES) group along with significant contributions from other NSF NCAR scientists and engineers, interns, visitors, and collaborators across the world.

CREDIT enables users to train, run, and evaluate AI-based numerical weather and climate models. This documentation will guide you through installation, configuration, training, inference, evaluation, and extending the system with custom datasets and models. CREDIT’s new Generation 2 restructuring and a more intuitive CLI make it easier than ever to train your own emulator.

New here? Begin with Get Started — it gets you from zero to a running training job in under 10 minutes.

What you’ll find here:

• How to install CREDIT

• How to set up and train a model

• How to run inference and evaluate results

• How to contribute datasets, models, and enhancements

If you encounter problems or have suggestions, please open an issue on our GitHub repository. Contributions are welcome!

Introduction

• Get Started
  • 1. Install CREDIT
  • 2. Generate a config
  • 3. Fit the scalers with credit preprocess
  • 4. Start a training job
  • 5. Monitor progress
  • 6. Visualize a prediction
  • 7. Get help from the AI assistant
  • Common problems
• CLI
  • The typical workflow at a glance
  • credit begin — create a starter config
  • credit init — copy a built-in template
  • credit check — validate a config
  • credit preprocess — fit normalization scalers
  • credit train — train a model
  • credit rollout — generate forecasts
  • credit realtime — operational-style single forecast
  • credit submit — submit a job to the scheduler
  • credit convert — upgrade a v1 config to v2
  • credit plot — quick visualization of a prediction
  • credit metrics — WeatherBench2-style evaluation
  • credit ask — AI assistant
  • Quick reference

Generation 2 Components

• Overview
  • Gen 1 vs Gen 2: which one am I using?
  • Datasets
  • Preblocks
  • Models
  • Postblocks
  • Interfaces
  • Custom Objects
• Datasets
  • CREDIT Data Schema
  • Dataset Types
  • Writing Your Own Dataset
• Preblocks
  • General Preblock Structure
  • Preblock Types
  • Example: processing the ARCOERA5Dataset
  • Writing Your Own Preblock
• Postblocks
  • General Postblock Structure
  • Reconstruction
  • Inverse Transforms
  • Diagnostics
  • Vertical Interpolation
  • Conservation Fixers
  • Filtering, Masking, and Advection
  • Example: the WXFormer prediction + target-twin chain
  • Writing Your Own Postblock
• Models
  • How variables become model channels
  • WXFormer
  • CAMulator
  • Swin Transformer
  • U-Net
  • Choosing a model
• Loss Functions
  • Why a per-variable loss
  • Required postblocks
  • The loss section
  • Variable weighting
  • Latitude weighting
  • Per-variable loss overrides
  • Per-variable logging
  • Using a plain registry loss
  • Adding a custom loss
  • Complete example
• Verification Metrics
  • Quick start
  • Required postblocks
  • Built-in metrics
  • Output format
  • Variable weighting
  • Full option reference
  • Anomaly metrics: ACC and forecast activity
  • Adding a custom metric
  • Summary
• Custom Objects
  • When to use custom objects
  • The six object types
  • Adding a custom object from your own package
  • Where each type is referenced
  • Full example — all six types
  • Validate before you train
  • Gotchas
  • Contributing a new built-in type

Training and Inference

• Training a Model
  • Training locally
  • Start Training on Casper or Derecho
  • How many jobs do I need?
  • Available configs
  • What does a healthy training run look like?
  • Trainer configuration
  • Gen2 parallelism: FSDP2, tensor parallel, and domain parallel
  • Job submission
  • Resuming training
• Monitoring with TensorBoard
  • Enabling TensorBoard
  • What gets logged
  • Viewing logs locally
  • Viewing logs from Casper or Derecho (SSH port forwarding)
  • Comparing multiple runs
  • Resuming a run
  • Installation
• Running Inference
  • Gen2 (Current) Inference
  • Gen1 (Legacy) Rollouts
• Forecast API Server
  • Quick start
  • Endpoints
  • Interactive docs
  • Configuration
  • Deployment notes
• AI Agent
  • Quick start
  • Two modes, one command
  • Installation and setup
  • Usage reference
  • Example sessions
  • What the agent can access
  • Tips for best results
  • Cost
  • Troubleshooting
• Evaluation and Metrics
• Ensemble Training
  • Training Non-Deterministic Model Ensembles
  • Noise-Injection Ensembles
  • Technical Implementation Summary
  • Diffusion-Based Ensembles
• Ensemble Inference
  • Configuration
  • Ensemble Generation Methods
  • Execution
  • Metrics and Output
  • Technical Implementation

Contributing

• Contributing to CREDIT
  • Introduction
  • Where to start
  • Reporting bugs
  • Requesting new features
  • Improving Documentation
  • Development workflow overview
  • Get set up to develop on your local machine
  • Understanding the CREDIT Codebase
  • Feature Integration Process
  • Make your changes
  • Documentation
  • Contribute your code
• Backwards Compatibility

API Reference

• Overview

Generation 1

• Installing CREDIT from source
• What’s in the Configuration File?
• CREDIT Configuration Guide
• Prepare New Dataset
• Supported Model Architectures
• Post Blocks
• Losses (Gen 1)
• Training (Gen 1)
• Weather Forecasting vs Downscaling Climate Projections
• Training ML Weather and Climate Models
• Data Pipelines: ERA5_* vs DownscalingDataset
• DownscalingDataset How-To
• Summary
• RAL GWC regional model

---

Indices and Tables

• Index

• Module Index

• Search Page