Inertialess tracers are broken

[jngrad]

In the LB particle coupling code, real particles are folded onto the local domain to eliminate edge cases, such as when a particle is located at pos=[-1e-30,-1e-30,-1e-30] in unfolded coordinates. Here is the corresponding code:

espresso/src/core/lb/particle_coupling.cpp

Lines 179 to 180 in 4285bb6

	auto const halo_pos = positions_in_halo(m_box_geo.folded_position(p.pos()),
	m_box_geo, m_local_box, agrid);

The LB inertialess tracers do not take this precaution:

espresso/src/core/virtual_sites/lb_tracers.cpp

Lines 63 to 65 in 4285bb6

	for (auto const &pos :
	positions_in_halo(p.pos(), box_geo, local_box, agrid)) {
	add_md_force(lb, pos * to_lb_units, p.force(), time_step);

In addition, the inertialess tracers code converts positions from MD units to LB units, even though add_md_force() expects positions in MD units since it calls LB::Solver::add_force_density(), which converts positions from MD units to LB units:

espresso/src/core/lb/Solver.cpp

Lines 198 to 200 in 4285bb6

	if (not ptr->add_force_at_pos(pos / ptr->get_agrid(), force_density)) {
	throw std::runtime_error("Cannot apply force to LB");
	}

When using agrid values other than unity, inertialess tracers are converted twice. The confusion is probably due to the fact that LB::Solver methods don't convert all quantities. For example, forces aren't converted, but positions and velocities are.

Here is a MWE:

import espressomd
import espressomd.lb
import espressomd.propagation

system = espressomd.System(box_l=3 * [6.0])
system.time_step = 0.01
system.cell_system.skin = 0.4
lbf = espressomd.lb.LBFluidWalberla(
    tau=0.01, agrid=0.5, density=1., kinematic_viscosity=1.,
    ext_force_density=[0.02, 0.04, 0.06], single_precision=False)
system.lb = lbf
system.thermostat.set_lb(LB_fluid=lbf, seed=3, gamma=1.)
p = system.part.add(pos=3 * [-1e-30], v=[-1, 0, 0],
                    propagation=espressomd.propagation.Propagation.TRANS_LB_TRACER)
system.integrator.run(1)
print(p.v)

Output:

RuntimeError: Cannot apply force to LB

Expected output:

[0.0003 0.0006 0.0009]

The virtual_sites_tracers.py test fails when agrid=0.5 in virtual_sites_tracers_common.py.

This bug most likely affects all waLBerla versions of ESPResSo. Here is the same bug in the original commit that introduced waLBerla on the python branch:

espresso/src/core/virtual_sites/VirtualSitesInertialessTracers.cpp

Lines 74 to 76 in 3fd1709

	for (auto pos : positions_in_halo(p.pos(), box_geo)) {
	add_md_force(pos * to_lb_units, -p.force(), time_step);
	}