lammps/src/ML-PACE/compute_pace.cpp

// clang-format off
/* ----------------------------------------------------------------------
   LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
   https://www.lammps.org/, Sandia National Laboratories
   LAMMPS Development team: developers@lammps.org
   Copyright (2003) Sandia Corporation.  Under the terms of Contract
   DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
   certain rights in this software.  This software is distributed under
   the GNU General Public License.
   See the README file in the top-level LAMMPS directory.
------------------------------------------------------------------------- */

#include "compute_pace.h"

#include "ace-evaluator/ace_c_basis.h"
#include "ace-evaluator/ace_evaluator.h"
#include "ace-evaluator/ace_types.h"

#include "atom.h"
#include "comm.h"
#include "error.h"
#include "force.h"
#include "memory.h"
#include "modify.h"
#include "neigh_list.h"
#include "neighbor.h"
#include "pair.h"
#include "update.h"

namespace LAMMPS_NS {
struct ACECimpl {
  ACECimpl() : basis_set(nullptr), ace(nullptr) {}
  ~ACECimpl()
  {
    delete basis_set;
    delete ace;
  }
  ACECTildeBasisSet *basis_set;
  ACECTildeEvaluator *ace;
};
}    // namespace LAMMPS_NS

using namespace LAMMPS_NS;

enum { SCALAR, VECTOR, ARRAY };
ComputePACE::ComputePACE(LAMMPS *lmp, int narg, char **arg) :
    Compute(lmp, narg, arg), cutsq(nullptr), list(nullptr), pace(nullptr), paceall(nullptr),
    pace_peratom(nullptr), map(nullptr), c_pe(nullptr), c_virial(nullptr), acecimpl(nullptr)
{
  array_flag = 1;
  extarray = 0;
  bikflag = 0;
  dgradflag = 0;

  int ntypes = atom->ntypes;
  int nargmin = 4;

  acecimpl = new ACECimpl;
  if (narg < nargmin) error->all(FLERR,"Illegal compute pace command");

  bikflag = utils::inumeric(FLERR, arg[4], false, lmp);
  dgradflag = utils::inumeric(FLERR, arg[5], false, lmp);
  if (dgradflag && !bikflag)
    error->all(FLERR,"Illegal compute pace command: dgradflag=1 requires bikflag=1");

  memory->create(map,ntypes+1,"pace:map");

  //read in file with CG coefficients or c_tilde coefficients

  auto potential_file_name = utils::get_potential_file_path(arg[3]);
  delete acecimpl->basis_set;
  acecimpl->basis_set = new ACECTildeBasisSet(potential_file_name);
  cutmax = acecimpl->basis_set->cutoffmax;

  //# of rank 1, rank > 1 functions

  int n_r1, n_rp = 0;
  n_r1 = acecimpl->basis_set->total_basis_size_rank1[0];
  n_rp = acecimpl->basis_set->total_basis_size[0];

  int ncoeff = n_r1 + n_rp;
  nvalues = ncoeff;

  ndims_force = 3;
  ndims_virial = 6;
  bik_rows = 1;
  yoffset = nvalues;
  zoffset = 2*nvalues;
  natoms = atom->natoms;
  if (bikflag) bik_rows = natoms;
    dgrad_rows = ndims_force*natoms;
  size_array_rows = bik_rows+dgrad_rows + ndims_virial;
  if (dgradflag) {
    size_array_rows = bik_rows + 3*natoms*natoms + 1;
    size_array_cols = nvalues + 3;
    if (comm->me == 0)
      error->warning(FLERR,"dgradflag=1 creates a N^2 array, beware of large systems.");
  } else size_array_cols = nvalues*atom->ntypes + 1;
  lastcol = size_array_cols-1;

  ndims_peratom = ndims_force;
  size_peratom = ndims_peratom*nvalues*atom->ntypes;

  nmax = 0;
}

/* ---------------------------------------------------------------------- */

ComputePACE::~ComputePACE()
{
  modify->delete_compute(id_virial);

  delete acecimpl;
  memory->destroy(pace);
  memory->destroy(paceall);
  memory->destroy(cutsq);
  memory->destroy(pace_peratom);
  memory->destroy(map);
}

/* ---------------------------------------------------------------------- */

void ComputePACE::init()
{
  if (force->pair == nullptr)
    error->all(FLERR,"Compute pace requires a pair style be defined");

  if (cutmax > force->pair->cutforce)
    error->all(FLERR,"Compute pace cutoff is longer than pairwise cutoff");

  // need an occasional full neighbor list
  neighbor->add_request(this, NeighConst::REQ_FULL | NeighConst::REQ_OCCASIONAL);

  if (modify->get_compute_by_style("pace").size() > 1 && comm->me == 0)
    error->warning(FLERR,"More than one compute pace");

  // allocate memory for global array
  memory->create(pace,size_array_rows,size_array_cols, "pace:pace");
  memory->create(paceall,size_array_rows,size_array_cols, "pace:paceall");
  array = paceall;

  // find compute for reference energy

  c_pe = modify->get_compute_by_id("thermo_pe");
  if (!c_pe) error->all(FLERR,"Compute thermo_pe does not exist.");

  // add compute for reference virial tensor

  id_virial = id + std::string("_press");
  c_virial = modify->add_compute(id_virial + " all pressure NULL virial");
}

/* ---------------------------------------------------------------------- */

void ComputePACE::init_list(int /*id*/, NeighList *ptr)
{
  list = ptr;
}

/* ---------------------------------------------------------------------- */

void ComputePACE::compute_array()
{
  int ntotal = atom->nlocal + atom->nghost;
  invoked_array = update->ntimestep;

  // grow pace_peratom array if necessary

  if (atom->nmax > nmax) {
    memory->destroy(pace_peratom);
    nmax = atom->nmax;
    memory->create(pace_peratom,nmax,size_peratom,"pace:pace_peratom");
  }

  // clear global array

  for (int irow = 0; irow < size_array_rows; irow++){
    for (int icoeff = 0; icoeff < size_array_cols; icoeff++){
      pace[irow][icoeff] = 0.0;
    }
  }

  // clear local peratom array

  for (int i = 0; i < ntotal; i++){
    for (int icoeff = 0; icoeff < size_peratom; icoeff++) {
      pace_peratom[i][icoeff] = 0.0;
    }
  }

  // invoke full neighbor list (will copy or build if necessary)

  neighbor->build_one(list);

  const int inum = list->inum;
  const int* const ilist = list->ilist;
  const int* const numneigh = list->numneigh;
  int** const firstneigh = list->firstneigh;
  int * const type = atom->type;

  //determine the maximum number of neighbours
  int max_jnum = -1;
  int nei = 0;
  int jtmp =0;
  for (int iitmp = 0; iitmp < list->inum; iitmp++) {
    int itmp = ilist[iitmp];
    jtmp = numneigh[itmp];
    nei = nei + jtmp;
    if (jtmp > max_jnum){
      max_jnum = jtmp;
    }
  }

  // compute pace derivatives for each atom in group
  // use full neighbor list to count atoms less than cutoff

  const int* const mask = atom->mask;
  const int ntypes = atom->ntypes;

  for (int ii = 0; ii < inum; ii++) {
    int irow = 0;
    if (bikflag) irow = atom->tag[ilist[ii] & NEIGHMASK]-1;
    const int i = ilist[ii];
    if (mask[i] & groupbit) {
      const int itype = type[i];
      const int* const jlist = firstneigh[i];
      const int jnum = numneigh[i];
      const int typeoffset_local = ndims_peratom*nvalues*(itype-1);
      const int typeoffset_global = nvalues*(itype-1);

      delete acecimpl->ace;
      acecimpl->ace = new ACECTildeEvaluator(*acecimpl->basis_set);
      acecimpl->ace->compute_projections = true;
      acecimpl->ace->compute_b_grad = true;
      int n_r1, n_rp = 0;
      n_r1 = acecimpl->basis_set->total_basis_size_rank1[0];
      n_rp = acecimpl->basis_set->total_basis_size[0];

      int ncoeff = n_r1 + n_rp;
      acecimpl->ace->element_type_mapping.init(ntypes+1);
      for (int ik = 1; ik <= ntypes; ik++) {
        for(int mu = 0; mu < acecimpl->basis_set->nelements; mu++){
          if (mu != -1) {
            if (mu == ik - 1) {
              map[ik] = mu;
              acecimpl->ace->element_type_mapping(ik) = mu;
            }
          }
        }
      }

      if (dgradflag) {

        // dBi/dRi tags

        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 0][0] = atom->tag[i]-1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 0][1] = atom->tag[i]-1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 0][2] = 0;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 1][0] = atom->tag[i]-1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 1][1] = atom->tag[i]-1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 1][2] = 1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 2][0] = atom->tag[i]-1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 2][1] = atom->tag[i]-1;
        pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 2][2] = 2;

        // dBi/dRj tags

        for (int j=0; j<natoms; j++) {
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 0][0] = atom->tag[i]-1;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 0][1] = j;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 0][2] = 0;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 1][0] = atom->tag[i]-1;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 1][1] = j;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 1][2] = 1;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 2][0] = atom->tag[i]-1;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 2][1] = j;
          pace[bik_rows + ((j)*3*natoms) + 3*(atom->tag[i]-1) + 2][2] = 2;
        }
      }

      // resize the neighbor cache after setting the basis
      acecimpl->ace->resize_neighbours_cache(max_jnum);
      acecimpl->ace->compute_atom(i, atom->x, atom->type, list->numneigh[i], list->firstneigh[i]);
      Array1D<DOUBLE_TYPE> Bs = acecimpl->ace->projections;

      for (int jj = 0; jj < jnum; jj++) {
        const int j = jlist[jj];
        //replace mapping of jj to j
        if (!dgradflag) {
          double *pacedi = pace_peratom[i]+typeoffset_local;
          double *pacedj = pace_peratom[j]+typeoffset_local;

          //force array in (func_ind,neighbour_ind,xyz_ind) format
          // dimension: (n_descriptors,max_jnum,3)
          //example to access entries for neighbour jj after running compute_atom for atom i:
          for (int func_ind =0; func_ind < n_r1 + n_rp; func_ind++){
            DOUBLE_TYPE fx_dB = acecimpl->ace->neighbours_dB(func_ind,jj,0);
            DOUBLE_TYPE fy_dB = acecimpl->ace->neighbours_dB(func_ind,jj,1);
            DOUBLE_TYPE fz_dB = acecimpl->ace->neighbours_dB(func_ind,jj,2);
            pacedi[func_ind] += fx_dB;
            pacedi[func_ind+yoffset] += fy_dB;
            pacedi[func_ind+zoffset] += fz_dB;
            pacedj[func_ind] -= fx_dB;
            pacedj[func_ind+yoffset] -= fy_dB;
            pacedj[func_ind+zoffset] -= fz_dB;
            }
         } else {
            for (int iicoeff = 0; iicoeff < ncoeff; iicoeff++) {

              // add to pace array for this proc
              // dBi/dRj
              DOUBLE_TYPE fx_dB = acecimpl->ace->neighbours_dB(iicoeff,jj,0);
              DOUBLE_TYPE fy_dB = acecimpl->ace->neighbours_dB(iicoeff,jj,1);
              DOUBLE_TYPE fz_dB = acecimpl->ace->neighbours_dB(iicoeff,jj,2);
              pace[bik_rows + ((atom->tag[j]-1)*3*natoms) + 3*(atom->tag[i]-1) + 0][iicoeff+3] -= fx_dB;
              pace[bik_rows + ((atom->tag[j]-1)*3*natoms) + 3*(atom->tag[i]-1) + 1][iicoeff+3] -= fy_dB;
              pace[bik_rows + ((atom->tag[j]-1)*3*natoms) + 3*(atom->tag[i]-1) + 2][iicoeff+3] -= fz_dB;

              // dBi/dRi
              pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 0][iicoeff+3] += fx_dB;
              pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 1][iicoeff+3] += fy_dB;
              pace[bik_rows + ((atom->tag[i]-1)*3*natoms) + 3*(atom->tag[i]-1) + 2][iicoeff+3] += fz_dB;
            }
          }
        } // loop over jj inside
      if (!dgradflag) {

        int k = typeoffset_global;

        for (int icoeff = 0; icoeff < ncoeff; icoeff++){
          pace[irow][k++] += Bs(icoeff);
        }
      } else {
        int k = 3;
        for (int icoeff = 0; icoeff < ncoeff; icoeff++){
          pace[irow][k++] += Bs(icoeff);
        }
      }
    } //group bit
  } // for ii loop
  // accumulate force contributions to global array
  if (!dgradflag){
    for (int itype = 0; itype < atom->ntypes; itype++) {
      const int typeoffset_local = ndims_peratom*nvalues*itype;
      const int typeoffset_global = nvalues*itype;
      for (int icoeff = 0; icoeff < nvalues; icoeff++) {
        for (int i = 0; i < ntotal; i++) {
          double *pacedi = pace_peratom[i]+typeoffset_local;
          int iglobal = atom->tag[i];
          int irow = 3*(iglobal-1)+1;
          pace[irow++][icoeff+typeoffset_global] += pacedi[icoeff];
          pace[irow++][icoeff+typeoffset_global] += pacedi[icoeff+yoffset];
          pace[irow][icoeff+typeoffset_global] += pacedi[icoeff+zoffset];
        }
      }
    }
  }

  if (!dgradflag) {
    // accumulate forces to global array
    for (int i = 0; i < atom->nlocal; i++) {
      int iglobal = atom->tag[i];
      int irow = 3*(iglobal-1)+1;
      pace[irow++][lastcol] = atom->f[i][0];
      pace[irow++][lastcol] = atom->f[i][1];
      pace[irow][lastcol] = atom->f[i][2];
    }
  } else {

    // for dgradflag=1, put forces at first 3 columns of bik rows

    for (int i=0; i<atom->nlocal; i++) {
      int iglobal = atom->tag[i];
      pace[iglobal-1][0+0] = atom->f[i][0];
      pace[iglobal-1][0+1] = atom->f[i][1];
      pace[iglobal-1][0+2] = atom->f[i][2];
    }
  }

  dbdotr_compute();

  // sum up over all processes
  MPI_Allreduce(&pace[0][0],&paceall[0][0],size_array_rows*size_array_cols,MPI_DOUBLE,MPI_SUM,world);

  // assign energy to last column

  if (!dgradflag) {
    for (int i = 0; i < bik_rows; i++) paceall[i][lastcol] = 0;
    int irow = 0;
    double reference_energy = c_pe->compute_scalar();
    paceall[irow][lastcol] = reference_energy;
  } else {

    // assign reference energy right after the dgrad rows, first column

    int irow = bik_rows + 3*natoms*natoms;
    double reference_energy = c_pe->compute_scalar();
    paceall[irow][0] = reference_energy;
  }

  // assign virial stress to last column
  // switch to Voigt notation

  if (!dgradflag) {
    c_virial->compute_vector();
    int irow = 3*natoms+bik_rows;
    paceall[irow++][lastcol] = c_virial->vector[0];
    paceall[irow++][lastcol] = c_virial->vector[1];
    paceall[irow++][lastcol] = c_virial->vector[2];
    paceall[irow++][lastcol] = c_virial->vector[5];
    paceall[irow++][lastcol] = c_virial->vector[4];
    paceall[irow++][lastcol] = c_virial->vector[3];
  }
}

/* ----------------------------------------------------------------------
   compute global virial contributions via summing r_i.dB^j/dr_i over
   own & ghost atoms
------------------------------------------------------------------------- */
void ComputePACE::dbdotr_compute()
{

  if (dgradflag) return;

  double **x = atom->x;
  int irow0 = bik_rows+ndims_force*natoms;

  // sum over ace contributions to forces
  // on all particles including ghosts

  int nall = atom->nlocal + atom->nghost;
  for (int i = 0; i < nall; i++)
    for (int itype = 0; itype < atom->ntypes; itype++) {
      const int typeoffset_local = ndims_peratom*nvalues*itype;
      const int typeoffset_global = nvalues*itype;
      double *pacedi = pace_peratom[i]+typeoffset_local;
      for (int icoeff = 0; icoeff < nvalues; icoeff++) {
        double dbdx = pacedi[icoeff];
        double dbdy = pacedi[icoeff+yoffset];
        double dbdz = pacedi[icoeff+zoffset];
        int irow = irow0;
        pace[irow++][icoeff+typeoffset_global] += dbdx*x[i][0];
        pace[irow++][icoeff+typeoffset_global] += dbdy*x[i][1];
        pace[irow++][icoeff+typeoffset_global] += dbdz*x[i][2];
        pace[irow++][icoeff+typeoffset_global] += dbdz*x[i][1];
        pace[irow++][icoeff+typeoffset_global] += dbdz*x[i][0];
        pace[irow++][icoeff+typeoffset_global] += dbdy*x[i][0];
      }
    }
}

/* ----------------------------------------------------------------------
   memory usage
------------------------------------------------------------------------- */

double ComputePACE::memory_usage()
{

  double bytes = (double)size_array_rows*size_array_cols*sizeof(double); // pace
  bytes += (double)size_array_rows*size_array_cols*sizeof(double);       // paceall
  bytes += (double)nmax*size_peratom * sizeof(double);                   // pace_peratom
  int n = atom->ntypes+1;
  bytes += (double)n*sizeof(int);        // map

  return bytes;
}