lbboundaries reset after lbf.load_checkpoint()

[christophlohrmann]

After loading lb populations from checkpoint, boundary flags disappear and the fluid accelerates indefinitely

import numpy as np
import tempfile
import espressomd
import espressomd.lb
import espressomd.lbboundaries
import espressomd.shapes
import espressomd.checkpointing


def assert_has_boundaries(lbf):
    if np.sum(lbf[:,:,:].boundary) == 0:
        raise AssertionError("no boundary flags set in lb fluid")


system = espressomd.System(box_l = 3*[10])
system.time_step = 1
lbf = espressomd.lb.LBFluid(agrid = 1, tau=1, visc = 0.2, dens=1)
system.actors.add(lbf)
bound_shape = espressomd.shapes.Sphere(center = 3*[5], radius = 2)
lbbound = espressomd.lbboundaries.LBBoundary(shape = bound_shape)
system.lbboundaries.add(lbbound)
assert_has_boundaries(lbf)

checkpoint = espressomd.checkpointing.Checkpoint(
    checkpoint_id="test",
    checkpoint_path="./",
)
checkpoint.register("system")
checkpoint.save()
lbf.save_checkpoint('./test/lbf.cp', binary=True)
system.actors.clear()
system.lbboundaries.clear()

checkpoint.load()
assert_has_boundaries(lbf)

lbf.load_checkpoint('./test/lbf.cp', binary=True)
assert_has_boundaries(lbf)

[jngrad]

Thank you for this MWE!

When reloading LB populations, any LB boundary information is overwritten with zeros. There is functionality to regenerate the LB boundary information, but the callback call is missing. Here is the bugfix for 4.2.0 and 4.3-dev:

diff --git a/src/core/grid_based_algorithms/lb.cpp b/src/core/grid_based_algorithms/lb.cpp
index 504f3088f..a6e37cb58 100644
--- a/src/core/grid_based_algorithms/lb.cpp
+++ b/src/core/grid_based_algorithms/lb.cpp
@@ -33,4 +33,5 @@
 #include "communication.hpp"
 #include "errorhandling.hpp"
+#include "event.hpp"
 #include "grid.hpp"
 #include "grid_based_algorithms/lb_boundaries.hpp"
@@ -208,3 +209,4 @@ void lb_initialize_fields(std::vector<LB_FluidNode> &lb_fields,
 #endif // LB_BOUNDARIES
   }
+  on_lbboundary_change();
 }

Could you please confirm that your simulations are fixed with this change? From my side, both the testsuite and your MWE are successful.

This regression was introduced in 4.2.0, back when serialization of object lists was temporarily removed from ESPResSo. The checkpointing test worked around this issue by checking LB boundaries before the LB populations were reloaded. I'll rewrite the test to check LB boundary information both before and after the LB populations are reloaded.

[christophlohrmann]

works like a charm