run_neb

run_neb(program, theta_1, theta_2, *, n_pivots=12, n_steps=50, learning_rate=0.1, gradient_method=GradientMethod.PARAMETER_SHIFT, eps=0.001, convergence_tol=None, reporter=None)[source]

Find a minimum-energy path between two parameter vectors via NEB.

Places n_pivots images (including the two fixed endpoints) along a straight line from theta_1 to theta_2 and iteratively relaxes the interior images. At each step the gradient perpendicular to the chain tangent is computed and the images are updated via gradient descent, then redistributed uniformly along the path.

Warning

This function is experimental. Convergence is sensitive to learning_rate, eps, and n_pivots.

Parameters:

  • program (_SupportsVizScan) – Variational program whose loss landscape is explored.

  • theta_1 (TypeAliasType) – First endpoint (fixed).

  • theta_2 (TypeAliasType) – Second endpoint (fixed).

  • n_pivots (int) – Total number of images including the two endpoints. Must be at least 3.

  • n_steps (int) – Maximum number of NEB relaxation iterations.

  • learning_rate (float) – Step size for gradient descent on interior images.

  • gradient_method (GradientMethod) – Strategy for computing gradients. GradientMethod.PARAMETER_SHIFT (default) uses the parameter-shift rule. GradientMethod.FINITE_DIFFERENCE uses centered finite differences with step size eps.

  • eps (float) – Finite-difference step size. Only used when gradient_method is GradientMethod.FINITE_DIFFERENCE.

  • convergence_tol (float | None) – If set, stop early when the maximum pivot displacement in a step falls below this threshold.

  • reporter (ProgressReporter | None) – Optional ProgressReporter.

Returns:

Relaxed path, energies, cumulative distances, and the full iteration history.

Return type:

NEBResult

Raises: