LBFGS

Local minimisation routines. Global optimisation, single and double-ended transition state searches all rely on local minimisation

topsearch.minimisation.lbfgs.minimise(func_grad, initial_position: NDArray[Any, Any], bounds: list, conv_crit: float = 1e-06, history_size: int = 5, n_steps: int = 200, args: list = None) → tuple[NDArray[Any, Any], float, dict]

Wrapper for the scipy box-constrained LBFGS implementation. Takes in a function, gradient and initial position and performs local minimisation subject to the specified bounds

topsearch.minimisation.lbfgs.minimise_ase(potential, initial_position: NDArray[Any, Any], numbers: NDArray[Any, Any], conv_crit: float = 0.001, n_steps: int = 200, args: list = None, output_level=0) → tuple[NDArray[Any, Any], float, dict]

Wrapper for ASE implementation of LBFGS.

Parameters:

  • potential (MachineLearningPotential or any other Potential type with an ase calculator associated)

  • initial_position (1D NDArray with positions to be optimised)

  • numbers (atomic numbers)

  • conv_crit (float, maximum force to define convergence. Note that this differs from scipy version of function)

  • n_steps (int, maximum number of steps)

  • args (list, not used currently, there for compatibility)

Returns:

  • min_coords (minimised coordinates as 1D array)

  • f_val (energy value)

  • results_dict (info on the success or otherwise of optimisation)