CurrentCorrelationFunction.py

# **************************************************************************
#
# MDANSE: Molecular Dynamics Analysis for Neutron Scattering Experiments
#
# @file      Src/Framework/Jobs/CurrentCorrelationFunction.py
# @brief     Implements module/class/test CurrentCorrelationFunction
#
# @homepage  https://www.isis.stfc.ac.uk/Pages/MDANSEproject.aspx
# @license   GNU General Public License v3 or higher (see LICENSE)
# @copyright Institut Laue Langevin 2013-now
# @copyright ISIS Neutron and Muon Source, STFC, UKRI 2021-now
# @authors   Scientific Computing Group at ILL (see AUTHORS)
#
# **************************************************************************

import collections
import itertools
from math import ceil
import os
from tempfile import gettempdir, tempdir

import netCDF4
import numpy

from MDANSE import REGISTRY
from MDANSE.Framework.Jobs.IJob import IJob
from MDANSE.Mathematics.Arithmetic import weight
from MDANSE.Mathematics.Signal import correlation, differentiate, normalize, get_spectrum
from MDANSE.MolecularDynamics.Trajectory import read_atoms_trajectory


class CurrentCorrelationFunction(IJob):
    """
    Computes the current correlation function for a set of atoms.
    The transverse and longitudinal current correlation functions are typically used to study the propagation of excitations in disordered systems.
    The longitudinal current is directly related to density fluctuations and the transverse current is linked to propagating 'shear modes'.
    
    For more information, see e.g. 'J.-P. Hansen and I. R. McDonald, Theory of Simple Liquids (3rd ed., Elsevier), chapter 7.4: Correlations
    in space and time)' 
    """
    
    label = "Current Correlation Function"

    category = ('Analysis','Scattering',)
    
    ancestor = ["mmtk_trajectory","molecular_viewer"]

    settings = collections.OrderedDict()
    settings['trajectory'] = ('mmtk_trajectory',{})
    settings['frames'] = ('frames', {'dependencies':{'trajectory':'trajectory'}})
    settings['instrument_resolution'] = ('instrument_resolution',{'dependencies':{'trajectory':'trajectory','frames' : 'frames'}})
    settings['interpolation_order'] = ('interpolation_order', {'label': "velocities",
                                                               'dependencies': {'trajectory': 'trajectory'},
                                                               'default': 'no interpolation'})
    settings['interpolation_mode'] = ('single_choice', {'choices': ['one-time in-memory interpolation',
                                                                    'repeated interpolation',
                                                                    'one-time disk interpolation',
                                                                    'automatic'],
                                                        'default': 'automatic'})
    settings['number_of_preloaded_fames'] = ('integer', {'default': 50, 'mini': -1, 'exclude': (0,)})
    settings['q_vectors'] = ('q_vectors',{'dependencies':{'trajectory':'trajectory'}})
    settings['atom_selection'] = ('atom_selection',{'dependencies':{'trajectory':'trajectory'}})
    settings['normalize'] = ('boolean', {'default':False})
    settings['atom_transmutation'] = ('atom_transmutation',{'dependencies':{'trajectory':'trajectory','atom_selection':'atom_selection'}})
    settings['weights'] = ('weights', {'default':'b_coherent',"dependencies":{'trajectory':'trajectory','atom_selection':'atom_selection', 'atom_transmutation':'atom_transmutation'}})
    settings['output_files'] = ('output_files', {'formats':["hdf","netcdf","ascii"]})
    settings['running_mode'] = ('running_mode',{})

    def initialize(self):
        """
        Initialize the input parameters and analysis self variables
        """

        self.numberOfSteps = self.configuration['q_vectors']['n_shells']

        nQShells = self.configuration["q_vectors"]["n_shells"]
        
        self._nFrames = self.configuration['frames']['number']
        
        self._instrResolution = self.configuration["instrument_resolution"]
        
        self._nOmegas = self._instrResolution['n_omegas']
                
        self._outputData.add("q","line", numpy.array(self.configuration["q_vectors"]["shells"]), units="1/nm") 

        self._outputData.add("time","line", self.configuration['frames']['duration'], units='ps')
        self._outputData.add("time_window","line", self._instrResolution["time_window"], units="au") 

        self._outputData.add("omega","line", self._instrResolution["omega"],units='rad/ps')
        self._outputData.add("omega_window","line", self._instrResolution["omega_window"], axis="omega", units="au") 

        self._elements = self.configuration['atom_selection']['unique_names']
        self._elementsPairs = sorted(itertools.combinations_with_replacement(self._elements,2))
        
        self._indexesPerElement = self.configuration['atom_selection'].get_indexes()

        for pair in self._elementsPairs:
            self._outputData.add("j(q,t)_long_%s%s"  % pair,"surface", (nQShells,self._nFrames), axis="q|time", units="au")                                                 
            self._outputData.add("j(q,t)_trans_%s%s" % pair,"surface", (nQShells,self._nFrames), axis="q|time", units="au")                                                 
            self._outputData.add("J(q,f)_long_%s%s"  % pair,"surface", (nQShells,self._nOmegas), axis="q|omega", units="au") 
            self._outputData.add("J(q,f)_trans_%s%s" % pair,"surface", (nQShells,self._nOmegas), axis="q|omega", units="au") 

        self._outputData.add("j(q,t)_long_total","surface", (nQShells,self._nFrames), axis="q|time"    , units="au")                                                 
        self._outputData.add("J(q,f)_long_total","surface", (nQShells,self._nOmegas), axis="q|omega", units="au") 
        self._outputData.add("j(q,t)_trans_total","surface", (nQShells,self._nFrames), axis="q|time"    , units="au")                                                 
        self._outputData.add("J(q,f)_trans_total","surface", (nQShells,self._nOmegas), axis="q|omega", units="au")

        traj = self.configuration['trajectory']['instance']
        nAtoms = traj.universe.numberOfAtoms()
        nFrames = self.configuration['frames']['n_frames']

        # Interpolate velocities of all atoms throughout the entire trajectory
        self._order = self.configuration["interpolation_order"]["value"]
        self._mode = self.configuration['interpolation_mode']['index']
        self._preload = self.configuration['number_of_preloaded_fames']['value']

        if self._order != "no interpolation" and self._mode == 3:
            from psutil import virtual_memory
            max_memory = virtual_memory().total

            if max_memory > nFrames * nAtoms * 3 * 8:
                self._mode = 0
            else:
                most_atoms = 0
                for idxs in self._indexesPerElement.values():
                    if len(idxs) > most_atoms:
                        most_atoms = len(idxs)

                if max_memory > nFrames * most_atoms * 3 * 8:
                    self._mode = 1
                else:
                    self._mode = 2

        if self._order != "no interpolation" and self._mode == 0:
            self._velocities = numpy.empty((nAtoms,nFrames,3),dtype=float)
            # Loop over the selected indexes and fill only this part of the 
            # self._velocities array, the rest, which is useless, remaining unset. 
            for idx in self.configuration['atom_selection']['flatten_indexes']:
                atomicTraj = read_atoms_trajectory(traj,
                                                    [idx],
                                                    first=self.configuration['frames']['first'],
                                                    last=self.configuration['frames']['last']+1,
                                                    step=self.configuration['frames']['step'],
                                                    variable=self.configuration['interpolation_order']["variable"])
                for axis in range(3):
                    self._velocities[idx,:,axis] = differentiate(atomicTraj[:, axis], order=self._order,
                                                                 dt=self.configuration['frames']['time_step'])

        # An alternative interpolation method which saves the velocities to an HDF5-style .nc file to save on memory
        elif self._order != "no interpolation" and self._mode == 2:
            if not hasattr(self, '_name'):
                self._name = '_'.join([self._type, IJob.define_unique_name()])

            with netCDF4.Dataset(os.path.join(gettempdir(), 'mdanse_' + self.name + '.nc'), 'w') as velocities:
                velocities.createDimension('particles', nAtoms+1)
                velocities.createDimension('time', nFrames)
                velocities.createDimension('axis', 3)

                if self._preload < 25:
                    # Chunking into 10x10 grid has shown to be faster at low self._preload values
                    velocities.createVariable('velocities', 'f8', ('particles', 'time', 'axis'),
                                              chunksizes=(ceil((nAtoms + 1) / 10), ceil(nFrames / 10), 3))
                else:
                    # At high self._preload values, chunks consisting of all atoms x preload number of frames are faster
                    velocities.createVariable('velocities', 'f8', ('particles', 'time', 'axis'),
                                              chunksizes=((nAtoms + 1), self._preload, 3))

                vels = numpy.empty((nFrames, 3))
                for idx in self.configuration['atom_selection']['flatten_indexes']:
                    atomicTraj = read_atoms_trajectory(traj,
                                                       [idx],
                                                       first=self.configuration['frames']['first'],
                                                       last=self.configuration['frames']['last'] + 1,
                                                       step=self.configuration['frames']['step'],
                                                       variable=self.configuration['interpolation_order']["variable"])
                    for axis in range(3):
                        vels[:, axis] = differentiate(atomicTraj[:, axis], order=self._order,
                                                      dt=self.configuration['frames']['time_step'])
                    velocities['velocities'][idx, :, :] = vels
            self._netcdf = netCDF4.Dataset(os.path.join(tempdir, 'mdanse_' + self.name + '.nc'), 'r')
            self._velocities = self._netcdf['velocities']

    def run_step(self, index):
        """
        Runs a single step of the job.\n
 
        :Parameters:
            #. index (int): The index of the step.
        :Returns:
            #. index (int): The index of the step. 
            #. rho (numpy.array): The exponential part of I(q,t)
        """

        shell = self.configuration["q_vectors"]["shells"][index]
        
        if not shell in self.configuration["q_vectors"]["value"]:
            return index, None

        traj = self.configuration['trajectory']['instance']

        qVectors = self.configuration["q_vectors"]["value"][shell]["q_vectors"]

        nQVectors = qVectors.shape[1]

        rho = {}
        rho_loop = {}
        rhoLong = {}
        rhoTrans = {}
        rhoLong_loop = {}
        rhoTrans_loop = {}
        for element in self._elements:
            rho[element] = numpy.zeros((self._nFrames, 3, nQVectors), dtype = numpy.complex64)
            rho_loop[element] = numpy.zeros((self._nFrames, 3, nQVectors), dtype = numpy.complex64)
            rhoLong_loop[element] = numpy.zeros((self._nFrames, 3, nQVectors), dtype = numpy.complex64)
            rhoTrans_loop[element] = numpy.zeros((self._nFrames, 3, nQVectors), dtype = numpy.complex64)

        # Certain interpolation strategies are faster when looping occurs primarily over elements
        if self._order != 'no interpolation' and (self._mode == 1 or (self._mode == 2 and self._preload == -1)):
            for element, idxs in self._indexesPerElement.items():
                nFrames = self.configuration['frames']['n_frames']
                all_velocities = numpy.empty((len(idxs), nFrames, 3), dtype=float)

                if self._mode == 1:
                    # Looping over elements is the only way to make repeated interpolation feasible at all
                    for i, __ in enumerate(idxs):
                        atomicTraj = read_atoms_trajectory(traj,
                                                           [i],
                                                           first=self.configuration['frames']['first'],
                                                           last=self.configuration['frames']['last'] + 1,
                                                           step=self.configuration['frames']['step'],
                                                           variable=self.configuration['interpolation_order'][
                                                               "variable"])
                        for axis in range(3):
                            all_velocities[i, :, axis] = differentiate(atomicTraj[:, axis], order=self._order,
                                                                       dt=self.configuration['frames']['time_step'])
                elif self._mode == 2:
                    # Loading the whole trajectory of atoms of 1 element into memory can be faster than above, and
                    # is much faster than accessing the disk repeatedly, but it takes more memory too.
                    all_velocities[:, :, :] = self._velocities[idxs, :, :]

                for i, frame in enumerate(self.configuration['frames']['value']):
                    coordinates = traj.configuration[frame].array[idxs, :]
                    velocities = numpy.transpose(all_velocities[:, i, :])[:, :, numpy.newaxis]
                    tmp = numpy.exp(1j * numpy.dot(coordinates, qVectors))[numpy.newaxis, :, :]
                    rho[element][i, :, :] = numpy.add.reduce(velocities * tmp, 1)

                Q2 = numpy.sum(qVectors ** 2, axis=0)

                qj = numpy.sum(rho[element] * qVectors, axis=1)
                rhoLong[element] = (qj[:, numpy.newaxis, :] * qVectors[numpy.newaxis, :, :]) / Q2
                rhoTrans[element] = rho[element] - rhoLong[element]

        # No interpolation and some interpolation approaches are faster when looping occurs primarily over frames
        else:
            # loop over the trajectory time steps
            for i, frame in enumerate(self.configuration['frames']['value']):
                conf = traj.configuration[frame]

                if self._order == 'no interpolation':
                    vel = traj.velocities[frame].array
                elif self._mode == 0:
                    vel = self._velocities[:, i, :]
                else:
                    # When accessing disk repeatedly (minimum memory usage), it is faster to load a whole frame at once
                    # as it requires fewer I/O. This does take more memory, but only as much as if there were no
                    # interpolation, since that also loads one whole frame at a time.
                    if i % self._preload == 0:
                        try:
                            preloaded_velocities = self._velocities[:, [j+i for j in range(self._preload)], :]
                        except IndexError:
                            preloaded_velocities = vel = self._velocities[:, i:, :]
                    vel = preloaded_velocities[:, i%self._preload, :]

                for element, idxs in self._indexesPerElement.items():
                    selectedCoordinates = conf.array[idxs,:]
                    selectedVelocities = vel[idxs,:]
                    selectedVelocities = numpy.transpose(selectedVelocities)[:,:,numpy.newaxis]
                    tmp = numpy.exp(1j*numpy.dot(selectedCoordinates, qVectors))[numpy.newaxis,:,:]
                    rho[element][i,:,:] = numpy.add.reduce(selectedVelocities*tmp,1)

            Q2 = numpy.sum(qVectors ** 2, axis=0)

            for element in self._elements:
                qj = numpy.sum(rho[element] * qVectors, axis=1)
                rhoLong[element] = (qj[:, numpy.newaxis, :] * qVectors[numpy.newaxis, :, :]) / Q2
                rhoTrans[element] = rho[element] - rhoLong[element]

        return index, (rhoLong, rhoTrans)

    def combine(self, index, x):
        """
        Combines returned results of run_step.\n
        :Parameters:
            #. index (int): The index of the step.\n
            #. x (any): The returned result(s) of run_step
        """

        if x is None:
            return
        
        jLong, jTrans = x
        
        for at1,at2 in self._elementsPairs:
            
            corrLong = numpy.zeros((self._nFrames,),dtype=numpy.float64)
            corrTrans = numpy.zeros((self._nFrames,),dtype=numpy.float64)
            
            for i in range(3):
                corrLong += correlation(jLong[at1][:,i,:],jLong[at2][:,i,:], axis=0, average=1)
                corrTrans += correlation(jTrans[at1][:,i,:],jTrans[at2][:,i,:], axis=0, average=1)
                            
            self._outputData["j(q,t)_long_%s%s" % (at1,at2)][index,:] += corrLong
            self._outputData["j(q,t)_trans_%s%s" % (at1,at2)][index,:] += corrTrans

    def finalize(self):
        """
        Finalizes the calculations (e.g. averaging the total term, output files creations ...)
        """
        try:
            self._netcdf.close()
        except (AttributeError, RuntimeError):
            pass

        try:
            os.remove(os.path.join(tempdir,  'mdanse_' + self.name + '.nc'))
        except (OSError, AttributeError):
            # OSError catches file not existing, and AttributeError caches gettempdir() having not been called
            pass

        try:
            del self._velocities
        except AttributeError:
            pass

        nAtomsPerElement = self.configuration['atom_selection'].get_natoms()
        for pair in self._elementsPairs:
            at1,at2 = pair
            ni = nAtomsPerElement[at1]
            nj = nAtomsPerElement[at2]
            self._outputData["j(q,t)_long_%s%s" % pair][:] /= ni*nj
            self._outputData["j(q,t)_trans_%s%s" % pair][:] /= ni*nj
            self._outputData["J(q,f)_long_%s%s" % pair][:] = get_spectrum(self._outputData["j(q,t)_long_%s%s" % pair],
                                                                          self.configuration["instrument_resolution"]["time_window"],
                                                                          self.configuration["instrument_resolution"]["time_step"],
                                                                          axis=1)        
            self._outputData["J(q,f)_trans_%s%s" % pair][:] = get_spectrum(self._outputData["j(q,t)_trans_%s%s" % pair],
                                                                           self.configuration["instrument_resolution"]["time_window"],
                                                                           self.configuration["instrument_resolution"]["time_step"],
                                                                           axis=1)        

        if self.configuration['normalize']["value"]:
            for pair in self._elementsPairs:
                self._outputData["j(q,t)_long_%s%s" % pair] = normalize(self._outputData["j(q,t)_long_%s%s" % pair],axis=1)
                self._outputData["j(q,t)_trans_%s%s" % pair] = normalize(self._outputData["j(q,t)_trans_%s%s" % pair],axis=1)

        jqtLongTotal = weight(self.configuration["weights"].get_weights(),self._outputData,nAtomsPerElement,2,"j(q,t)_long_%s%s")
        self._outputData["j(q,t)_long_total"][:] = jqtLongTotal

        jqtTransTotal = weight(self.configuration["weights"].get_weights(),self._outputData,nAtomsPerElement,2,"j(q,t)_trans_%s%s")
        self._outputData["j(q,t)_trans_total"][:] = jqtTransTotal
        
        sqfLongTotal = weight(self.configuration["weights"].get_weights(),self._outputData,nAtomsPerElement,2,"J(q,f)_long_%s%s")
        self._outputData["J(q,f)_long_total"][:] = sqfLongTotal

        sqfTransTotal = weight(self.configuration["weights"].get_weights(),self._outputData,nAtomsPerElement,2,"J(q,f)_trans_%s%s")
        self._outputData["J(q,f)_trans_total"][:] = sqfTransTotal
    
        self._outputData.write(self.configuration['output_files']['root'], self.configuration['output_files']['formats'], self._info)
        
        self.configuration['trajectory']['instance'].close()


REGISTRY['ccf'] = CurrentCorrelationFunction