Skip to content


Repository files navigation

Fewest Switches Surface Hopping Build Status Documentation Status

Python implementation of Tully's Fewest Switches Surface Hopping (FSSH) for model problems including a propagator and an implementation of Tully's model problems described in Tully, J.C. J. Chem. Phys. (1990) 93 1061. The current implementation works for diabatic as well as ab initio models, with two or more electronic states, and with one or more dimensional potentials.


  • mudslide package that contains
    • implementation of all surface hopping methods
      • TrajectorySH - FSSH
      • TrajectoryCum - FSSH with a cumulative point of view
      • Ehrenfest - Ehrenfest dynamics
    • collection of 1D models
      • TullySimpleAvoidedCrossing
      • TullyDualAvoidedCrossing
      • TullyExtendedCouplingReflection
      • SuperExchange
      • SubotnikModelX
      • SubotnikModelS
      • ShinMetiu
    • some 2D models
      • Subotnik2D
  • mudslide script that runs simple model trajectories
  • mudslide-surface script that prints 1D surface and couplings


  • numpy
  • scipy (for Shin-Metiu model)
  • pyyaml
  • turboparse


Mudslide has switched to a proper python package structure, which means to work properly it now needs to be "installed". The most straightforward way to do this is

cd /path/to/mudslide
pip install --user -e .

which install into your user installation dir. You can find out your user installation directory with the command

python -m site --user-base

To set up your PATH and PYTHONPATH to be able to use both the command line scripts and the python package, use

export PATH=$(python -m site --user-base)/bin:$PATH
export PYTHONPATH=$(python -m site --user-base):$PYTHONPATH

Trajectory Surface Hopping

Sets of simulations are run using the BatchedTraj class. A BatchedTraj object must be instantiated by passing a model object (handles electronic PESs and couplings), and a traj_gen generator that generates new initial conditions. Some simple canned examples are provided for traj_gen. All other options are passed as keyword arguments to the constructor. The compute() function of the BatchedTraj object returns a TraceManager object that contains all the results, but functionally behaves like a python dictionary. Custom TraceManagers can also be provided. For example:

import mudslide

simple_model = mudslide.models.TullySimpleAvoidedCrossing()

# Generates trajectories always with starting position -5, starting momentum 10.0, on ground state
traj_gen = mudslide.TrajGenConst(-5.0, 10.0, 0)

simulator = mudslide.BatchedTraj(simple_model, traj_gen, mudslide.TrajectorySH, samples = 4, bounds=[[-4],[4]])
results = simulator.compute()
outcomes = results.outcomes

print("Probability of reflection on the ground state:    %12.4f" % outcomes[0,0])
print("Probability of transmission on the ground state:  %12.4f" % outcomes[0,1])
print("Probability of reflection on the excited state:   %12.4f" % outcomes[1,0])
print("Probability of transmission on the excited state: %12.4f" % outcomes[1,1])

will run 4 scattering simulations with a particle starting in the ground state (0) at x=-5.0 a.u. and traveling with an initial momentum of 10.0 a.u.


  • initial_state - specify the initial electronic state
    • 0 (default) - start on the ground state
  • mass - particle mass (default: 2000 a.u.)
  • initial_time - starting value of time variable (default: 0.0 a.u.)
  • dt - timestep (default: abs(0.05 / velocity)))
  • total_time - total simulation length (default: 2 * abs(position/velocity))
  • samples - number of trajectories to run (default: 2000)
  • seed - random seed for trajectories (defaults however numpy does)
  • propagator - method used to propagate electronic wavefunction
    • "exponential" (default) - apply exponentiated Hamiltonian via diagonalization
    • "ode" - scipy's ODE integrator
  • nprocs - number of processes over which to parallelize trajectories (default: 1)
  • outcome_type - how to count statistics at the end of a trajectory
    • "state" (default) - use the state attribute of the simulation only
    • "populations" - use the diagonals of the density matrix
  • trace_every - save snapshot data every nth step (i.e., when nsteps%trace_every==0)


A primary goal of mudslide is to include the flexibility for users to provide their own model classes. To interface with mudslide, a model class should derive from mudslide.ElectronicModel_ and should implement

  • a compute() function that computes energies, forces, and derivative couplings
  • an nstates() function that returns the number of electronic states in the model
  • an ndim() function that returns the number of classical (vibrational) degrees of freedom in the model

compute() function

The compute() function needs to have the following signature:

def compute(self, X: ArrayLike, couplings: Any = None, gradients: Any None, reference: Any = None) -> None

The X input to the compute() function is an array of the positions. All other inputs are ignored for now but will eventually be used to allow the trajectory to enumerate precisely which quantities are desired at each call. At the end of the compute() function, the object must store

  • self.hamiltonian - An nstates x nstates array of the Hamiltonian (nstates is the number of electronic states)
  • self.force - An nstates x ndim array of the force on each PES (ndim is the number of classical degrees of freedom)
  • self.derivative_coupling - An nstates x nstates x ndim array where self.derivative_coupling[i,j,:] contains <i|d/dR|j>.

See the file mudslide/ for an example of a standalone ab initio model.

Diabatic Models

For diabatic models, such as Tully's scattering models, there is a helper class, mudslide.DiabaticModel_ to simplify construction of model classes.

To use it, have your model derive from mudslide.DiabaticModel_ and then you just need to implement two helper functions:

  • V(self, x) --- returns (nstates,nstates)-shaped numpy array containing the diabatic Hamiltonian matrix at nuclear position x
  • dV(self, x) --- returns (nstates,nstates,ndim)-shaped numpy array containing the gradient of the diabatic Hamiltonian matrix at nuclear position x

And further define two class variables:

  • nstates_ --- the number of electronic states in the model
  • ndim_ --- the number nuclear degrees of freedom in the model

In all cases, the input x ought to be a numpy array with ndim elements.

See the file mudslide/ for examples, which includes, among other models, the original Tully models. Additional models included are:

  • SuperExchange - Oleg Prezhdo's three-state super exchange model from Wang, Trivedi, Prezhdo JCTC (2014) doi:10.1021/ct5003835
  • SubotnikModelX - 'Model X' from Subotnik JPCA (2011) doi: 10.1021/jp206557h
  • SubotnikModelS - 'Model S' from Subotnik JPCA (2011) doi: 10.1021/jp206557h
  • Subotnik2D - 2D model from Subotnik JPCA (2011) doi: 10.1021/jp206557h
  • Shin-Metiu

Trajectory generator

For batch runs, one must tell BatchedTraj how to decide on new initial conditions and how to decide when a trajectory has finished. The basic requirements for each of those is simple.

The structure of these classes is somewhat strange because of the limitations of multiprocessing in python. To make use of multiprocessing, every object must be able to be pickled, meaning that multiprocessing inherits all the same limitations. As a result, when using multiprocessing, the trajectory generator class must be fully defined in the default namespace.

Generating initial conditions

This should be a generator function that accepts a number of samples and returns a dictionary with starting conditions filled in, e.g., the yield statement should be something like

yield { "position" : x0, "momentum" : p0, "initial_state" : 0 }

See TrajGenConst for an example.


Scans of the surfaces can be printed using the mudslide-surface command that is included for the installation. For usage, make sure your PATH includes the installation directory for scripts (e.g., by running export PATH=$(python -m site --user-base)/bin:$PATH) and run

mudslide-surface -h


Python implementation of Tully's Fewest Switches Surface Hopping








No packages published