RuNNer tutorials overview¶

This document summarizes what RuNNer can do, based on the current tutorials for training (2G/3G/4G), precomputation, and prediction, including committee support.

Operating modes¶

RuNNer workflows are controlled via runner_mode:

• runner_mode train
  Train models (2G, 3G, 4G), including committees, and (in 3G) arbitrary atomic-property models.

• runner_mode predict
  Run inference to predict energies and optional derived quantities (forces, stress, etc.) using trained models.

• runner_mode precompute
  Precompute descriptor features (and optionally derivatives) for analysis and/or faster training, and precompute Coulomb matrices needed for 4G workflows.

Supported model generations¶

• 2G-HDNNP: Second-generation HDNNPs are local (short-range) atom-centered neural networks.
• 3G-HDNNP: Third-generation workflows extend 2G by learning environment-dependent atomic properties (notably charges), which are used to compute and subtract long-range contributions during short-range training.
• 4G-HDNNP: Fourth-generation workflows compute globally equilibrated charges via charge equilibration (QEq) rather than directly predicting charges. The model predicts electronegativity (chi) and hardness (hardness), then solves a QEq system to obtain charges and electrostatic energies. This allows it to capture non-local charge transfer effects.