[sbryngelson]

Four commits: the code and goldens of the first version, the merge of master (after #2206, the dual-pass hypoelastic force), and a rework of the correction with its regenerated goldens (below).

Summary

The force and torque an immersed boundary receives (ib_force_wrt, two-way coupling) are summed from the fluid-body face fluxes. Those faces are fixed in the grid. When the body moves (moving_ibm > 0), the lab-frame momentum flux rho u_i u_n through a face is not the flux through the moving wall, rho u_i (u_n - V_n), with V_n the wall's normal velocity at the face (V_b = V + omega x r). The force then depends on the body's velocity through the grid, by a term of order M^2.

s_accumulate_ib_face_fluxes now takes q_cons_vf and, for moving_ibm > 0, subtracts V_n times the momentum density of the face (the mean of its fluid cell and its body cell) from the momentum flux used for the force and torque. This is a state, not a flux. An earlier version of this PR subtracted V_b f_m with f_m the discrete face mass flux; that is zero for a rigid translation only if the staircase ghost-cell fill leaks no mass, which it does not, so it was off right after an impulsive start (1.3% in a hypoelastic HLLD case at V = 10, 35% at V = 100, 42% for plain HLLC at V = 50, at step 1) while the unmodified flux was within 1e-3 there. Pressure and viscous terms are unchanged. Fixed bodies (moving_ibm = 0) take the old path. The energy and mass fluxes of the lumped-body heat (ib_heat_wrt) are not changed.

Test changes: two new tests (IBM -> Translating Circle and IBM -> Translating Spinning Circle -> Body-Frame Force) record ib_forces.dat of a prescribed translating (and turning) circle. Five goldens move because forces change (listed under Tests). The docs paragraph on the force history says what the force now is.

AI disclosure: made with Claude Code.

Verification: Galilean twins

A body translating at U through gas at rest must feel the same force as the same body held in a uniform flow -U. Setup A/B: same grid and domain, case A has moving_ibm = 1, vel(1) = U, gas at rest; twin B has the gas at -U and the body held (moving_ibm = 0; for the spin cases moving_ibm = 1, vel = 0 and the same spin). Metric rel = rms(F_A - F_B) / rms(F_B) over all steps. "master" has no correction, "merged" is this PR with the V_b f_m correction (master with #2206 merged), "fixed" is the present change. NVIDIA A100 node, nvfortran 25.11, --gpu acc, 1 rank, all three built from source and run on the same node.

Developed flow: 2D periodic unit box, circle R = 0.1 at x = 0.8 (it crosses the periodic boundary at t = 0.4), rho = 1, p = 1/1.4 (c = 1), HLLC, WENO5, RK3, dt = 0.2 dx, to t = 1, no-slip, inviscid, 200 cells per unit length unless stated.

case	master	merged	fixed
Fx, M = 0.1	2.05e-2	1.92e-2	1.72e-2
Fx, M = 0.5	2.21e-1	4.23e-2	3.91e-2
Fx, M = 0.9	7.24e-1	5.74e-2	3.84e-2
Fx, M = 0.5, 100 cells	2.12e-1	6.00e-2	6.24e-2
Fx, M = 0.5, 400 cells	2.15e-1	6.35e-2	5.01e-2
Fx, M = 0.5, slip wall	2.53e-1	2.28e-1	5.51e-2

Translation plus spin 3 rad/s, M = 0.5, 200 cells (twin: same spin, body held):

case	master	merged	fixed
Fx, inviscid	2.23e-1	3.99e-2	5.58e-2
Fy, inviscid	2.23	4.84e-1	3.18e-1
Tz, inviscid	2.13	3.85e-1	4.42e-1
Fx, viscous (Re_D = 200)	1.73e-1	3.92e-2	3.13e-2
Fy, viscous	1.59	1.21e-1	5.51e-2
Tz, viscous	8.98e-1	2.51e-1	1.95e-1

Two-way bodies (moving_ibm = 2, mass 5 times the displaced gas, M = 0.5; the twin is also two-way, starting from rest in gas moving at -U). Force mismatch, and the velocity gained at t = 1 (v_A - U against v_B):

case	master	merged	fixed
Fx rel, inviscid, 200 cells	1.45e-1	5.40e-2	4.42e-2
Fx rel, viscous, 100 cells	1.20e-1	6.80e-2	5.64e-2
gained velocity, inviscid (A / B)	-0.1942 / -0.1909	-0.1829 / -0.1893	-0.1889 / -0.1865
gained velocity, viscous (A / B)	-0.2036 / -0.1999	-0.1896 / -0.1983	-0.1969 / -0.1951

The Mach-dependent error stays removed (M = 0.9: 72% to 3.8%), and the slip-wall error, the largest limitation of the previous version, falls from 23% to 5.5% because the state carries the slip velocity. Fixed does not beat merged in every row: Fx and Tz of the inviscid spinning case are 5.6e-2 and 4.4e-1 against 4.0e-2 and 3.9e-1, and 100 cells at M = 0.5 is equal (6.2e-2 against 6.0e-2). The floor of 2-6% that is present in all three builds at M = 0.1 is not removed.

Impulsive start: 64 x 32 periodic box, circle r = 0.2, BA59123E fluid parameters (two fluids, pressure 1e5, dt = 1e-6) on a uniform state with G = 1e6 for the hypoelastic cases, 30 steps. Relative difference of Fx of A (body at V in gas at rest) from B (body fixed in gas at -V), (A - B)/B:

case	step	master	merged	fixed
hypoelastic HLLD dual pass, V = 10	1	-3.5e-6	+1.26e-2	+6.3e-6
	29	-3.4e-3	+1.26e-2	-1.5e-3
hypoelastic HLLD dual pass, V = 100	1	-2.1e-4	+3.50e-1	+3.1e-4
	29	-8.7e-2	+4.01e-1	-1.4e-2
plain HLLC, V = 50	1	-1.9e-4	+4.18e-1	+1.4e-4
	29	+2.2e-2	+4.59e-1	+2.1e-2

Over the 30 steps the rms difference is 2.0e-3 / 1.3e-2 / 8.6e-4 (V = 10), 4.9e-2 / 3.9e-1 / 2.7e-2 (V = 100) and 5.0e-2 / 4.4e-1 / 3.2e-2 (plain, V = 50) for master / merged / fixed. At step 1 the fixed force matches the held body to the level of master (about 1e-4 to 1e-6); the later difference is not a force-correction error but the moving body crossing cells while the twin's does not. For plain HLLC at V = 50 the fixed step-29 difference equals master's (2.1e-2 against 2.2e-2) and the rms is 3.2e-2 against master's 5.0e-2.

Limitations

• The twin floor (2-6% in the developed flow, from the body moving through the grid) is not removed and does not shrink with resolution.
• The face momentum is the mean of the fluid and body cell; the body cell is the ghost-filled cell. Using the fluid cell alone gave 6.7e-2 (M = 0.5) and 9.2e-2 (100 cells), and a no-slip wall state rho V_n V_b gave 1.97e-1 on the slip wall, so the mean was kept. I did not study a more principled face state.
• The spin-case torque (inviscid) is 44% of the twin's own torque. It is a small quantity and the relative measure is harsh, but it is worse than the previous version's 39%.
• Verified in 2D, one body, one rank. No 3D twin and no multi-rank twin was run.

Fixed bodies unchanged

Fixed-body twins (moving_ibm = 0; inviscid no-slip at 100 and 400 cells and slip at 200 cells), master against this branch on the same GPU: all 10 conservative-variable files of each run are byte-identical (cmp).

Tests

./mfc.sh test --no-chemistry with the existing goldens, A100 node, nvfortran 25.11, --gpu acc --mpi:

label	passed / failed	failures
IBM (101 tests)	94 / 7	09FDDC81, 997E2B0E, 540C72F1 (goldens move); B7A370B8, 5173E637, 49893269 stopped with pre_process exit code 132 and pass on a rerun (3 / 3)
Hypoelasticity (65 tests)	64 / 1	BA59123E (golden moves)
IBM Examples with moving bodies	3 / 1	AA49A8BC mibm_sphere_head_on_collision (golden moves); mibm_cylinder_in_cross_flow, mibm_shock_cylinder, mibm_particle_cloud pass

The five goldens that move, regenerated on an MI300A node (flang 24.0.0, --gpu mp), and why:

UUID	test	change
997E2B0E	IBM -> Translating Circle	first Fx -0.605 to -0.725 (14% of the largest entry over the run): the force just after the impulsive start, where the old correction was wrong
540C72F1	IBM -> Translating Spinning Circle -> Body-Frame Force	first Fx -0.665 to -0.825 (6% over the run), same cause
09FDDC81	IBM -> Adaptive dt (two-way circle)	4 of 10 files, up to 6.2e-4 of the file maximum
BA59123E	Hypoelasticity -> HLLD -> IBM -> Two-way Circle	101 of 111 files, up to 3.2e-2 of the file maximum; first Fx 171458 to 171882 (0.25%)
AA49A8BC	Example mibm_sphere_head_on_collision	5 of 12 files, up to 1.0e-1 (energy) of the file maximum

The two-way tests (09FDDC81, BA59123E, AA49A8BC) move because the body's acceleration is the corrected force; the two new tests move because their first entries are the impulsive-start force. I did not separate how much of each two-way move comes from the start and how much from later steps.

machine	compiler	what	result
A100 node	nvfortran 25.11, --gpu acc	the five regenerated goldens	5 / 5 pass
MI300A node	flang 24.0.0, --gpu mp	the five regenerated goldens and the three moving-body Examples that do not move	8 / 8 pass

CI covers Frontier (Cray CCE); I did not run it by hand. The new force tests stay skipped under --single.

This PR was made with Claude Code (AI).

[github-actions]

Claude Code Review

Head SHA: f50674f

Files changed:

• 4
• docs/documentation/case.md
• src/simulation/m_ibm.fpp
• toolchain/mfc/test/cases.py
• toolchain/mfc/test/test.py

Findings:
No issues found. Checked m_ibm.fpp (s_accumulate_ib_face_fluxes), cases.py and test.py hunks against their surroundings.

[codecov]

Codecov Report

❌ Patch coverage is 80.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 64.90%. Comparing base (9649d50) to head (55092f4).
⚠️ Report is 7 commits behind head on master.

Files with missing lines	Patch %	Lines
src/simulation/m_ibm.fpp	80.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2208      +/-   ##
==========================================
+ Coverage   63.71%   64.90%   +1.19%     
==========================================
  Files          90      110      +20     
  Lines       24569    29485    +4916     
  Branches     3584     4467     +883     
==========================================
+ Hits        15653    19138    +3485     
- Misses       6135     6809     +674     
- Partials     2781     3538     +757     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Lines of Code

File	Lines	Diff
src/simulation/m_ibm.fpp	2880	+9

Directory	Lines	Diff
simulation	39389	+9
total	59355	+9