Getting started

A host builds a band, steps until the projected force is under force_tol, and never lets this crate own the loop.

1. Depend on the published crate

[dependencies]
rgsaddle = "0.2"
rgmin = "0.3"
ndarray = "0.17"

2. Implement BandSurface

One eval fills energies and gradients for the rows it receives: the whole band on the first call, the interior images afterwards (the endpoints never move, so their energies are cached). The double-well test surface is the shortest honest example:

impl BandSurface for DoubleWell {
    fn eval(
        &self,
        positions: ArrayView2<f64>,
        energies: &mut Array1<f64>,
        gradients: &mut Array2<f64>,
    ) -> Result<(), SaddleError> {
        // V = (x² − 1)² + 2y² + 2z²
        Ok(())
    }
}

3. Step

let config = BandConfig {
    force_tol: 1e-3,
    max_move: 0.1,
    ..BandConfig::default()
};
let mut session = BandSession::new(config, initial)?;
loop {
    let report = session.step(&surface)?;
    if report.status == BandStatus::Converged {
        break;
    }
}

Default stepper is FIRE, as in eOn’s velocity NEB stepper. L-BFGS runs under rgmin’s Accept::Step: the projected band force is not conservative, so a step is taken without an energy test.

4. Reset when the surface changes

BandSession::reset is the model-update boundary. Quasi-Newton history taken on one surface epoch must not survive onto the next.

See tests/double_well.rs for the numbers the crate actually converges.