Sutton-Chen potential¶
SuttonChenPot evaluates
E = epsilon * [ 1/2 sum_i sum_{j != i} (a / r_ij)^n
- c * sum_i sqrt(rho_i) ]
rho_i = sum_{j != i} (a / r_ij)^m
The form is Sutton and Chen, Philosophical Magazine Letters 61, 139 (1990).
The default parameters are the reduced gold and platinum set: n = 10, m = 8,
c = 34.408, epsilon = 1, a = 1.
Doye and Wales use that set for the 10-8 cluster tables and write it on the
first line of each deposited point file.
The deposited Pt38 Oh geometry gives -10117.2454 and the Pt75 D5h geometry gives
-20431.3452, in units of epsilon, to the four digits printed in the 10-8 table.
Cagin, Dereli, Uludogan, and Tomak, arXiv:cond-mat/9709114, Table I, reprint the
Sutton-Chen rows and cite Sutton and Chen 1990 as their reference 1.
Platinum is epsilon = 1.98330e-2 eV, a = 3.92 Angstrom, c = 34.408, m = 8, n = 10.
Gold on the same row is epsilon = 1.27930e-2 eV, a = 4.08 Angstrom, with the
same c, m, and n.
Those two rows are sutton_chen_platinum and sutton_chen_gold.
The bulk nearest-neighbor distance and the cohesive well depth are
sutton_chen_scales.
The length is the static fcc minimum.
The depth is minus the energy per atom there. c is fitted to zero pressure at
the lattice parameter, so the minimum is a / sqrt(2) for every row of the Cagin
table.
The reduced 10-8 well is 295.4645 epsilon, which is 5.860 eV for platinum and
3.780 eV for gold.
The default cutoff is 40 in the coordinate unit, with no energy shift.
cutoff = 0 keeps every pair.
Cluster evaluation does not fold the cell.
Periodic evaluation uses the minimum image and needs a positive cutoff and a
cell; otherwise it throws std::invalid_argument.
Pairs come from the shared neighbour layer (nlist::updateCells), with a skin
of one fifth of the cutoff, or none in nlist::Mode::Stream.
With cutoff = 0 the layer searches without a bound: it bins nothing, stores no
list, and each pass takes the pairs straight from the coordinates, in the order
of the one bin that a search wider than the cluster gives.
A team of OMP_NUM_THREADS threads splits the atoms into as many runs, where
that one bin went to a single thread.
Every run is walked, also when the runtime starts fewer threads, so the result
depends on the count asked for and not on the threads that ran.
The forces keep their bits, and the energy, summed per run, can move in the last
bits: by up to 1.5e-14 of the energy in clusters of up to 561 atoms on 2 to 8
threads.
One pass builds the densities and a second pass writes the forces.
The density pass keeps the powers and 1/r^2 that the force of each pair takes,
and the force pass reads them back instead of raising the distance to n and
m again.
With a = 1 those are the density’s own powers; otherwise the force’s powers of
a^2 (1/r^2) are kept, since they can differ from those of a^2 / r^2 in the
last bit.
SuttonChenPot::setPairTermLimit bounds the store, 32 MiB by default, and the
force pass computes the pairs past it again, with the same bits.
rgpot_sutton_chen_energy_forces is the C entry point.
It takes an rgpot_sc_config and writes one energy and one force array.
Each thread keeps the potential of its last call in rgpot::Kept
(rgpot/KeptPotential.hpp), so a call with the same parameters and atom count
reuses its neighbour update and the room for the pair terms.
Any other call drops that potential and builds a new one, and so does a call
that throws.
Cost of a call¶
sc_call_bench (meson test --benchmark sc_call_bench) times the C entry on
free clusters with no cutoff, one thread pinned to one core, against a plain
loop over the half pairs that keeps each pair’s powers for its force.
Both give the same bits.
The medians of five runs, in microseconds per call:
Atoms |
12-6 entry |
12-6 plain loop |
9-6 entry |
9-6 plain loop |
|---|---|---|---|---|
13 |
0.749 |
0.580 |
0.848 |
0.634 |
38 |
6.434 |
4.939 |
6.686 |
5.223 |
55 |
13.690 |
10.267 |
13.739 |
10.914 |
147 |
91.082 |
72.088 |
94.165 |
75.244 |
561 |
1302.449 |
1063.068 |
1374.253 |
1122.602 |
Reduced cluster pairs¶
rgpot_lj_gradient and rgpot_morse_gradient return the energy gradient of an
unshifted cluster.
rgpot_lj_curvature_pairs fills energy, slope, and curvature for a list of
squared distances in one call.
The rgpot-pair crate exposes the same trees as lj_unit and morse_unit,
which a Rust caller can inline.
The unit Lennard-Jones branch is the direct inverse-power tree, with contraction
off, so a search sees the same rounding as that expression.
Stress¶
SuttonChenConfig::stress is off by default.
When it is set, forceImpl writes the Cauchy stress sigma = (1/V) dE/dstrain
into ForceOut::stress in row-major order.
It also sets has_stress to 1, and caps().stress is true.
The units are energy over length cubed, eV/Angstrom3for the
sutton_chen_platinum and sutton_chen_gold rows.
The sign is that of stress.hpp: a compressed crystal has negative diagonal
entries, and the pressure is minus a third of the trace.
A cell with no positive volume, such as a cluster with a null box, leaves
has_stress at 0.
The force pass writes the stress as well, because the energy depends on
positions only through pair distances.
With d = r_i - r_j and a force fscale * d on atom i, each pair adds
-fscale * d d^T to dE/dstrain.
That fscale already includes the embedding term, so the stress needs no third
loop.
Thompson, Plimpton and Mattson give the many-body virial under periodic
boundaries.
Their paper is in the Journal of Chemical Physics 131, 154107 (2009).
With the flag clear, forceImpl leaves the stress slots and has_stress as the
caller set them.
caps(), paramsKey(), the energy and the forces are then the same bits as
before the flag existed.
Setting the flag changes paramsKey(), so a cached result without stress is not
served to a caller that asked for one.
The C entry point rgpot_sutton_chen_energy_forces and rgpot_sc_config do not
carry the flag.
Periodic evaluation keeps one image of each partner. The stress, like the energy, is the whole sum only while the cutoff is under half the smallest face spacing of the cell.
Checks¶
SuttonChenPotTest compares the platinum row with the Large-scale
Atomic/Molecular Massively Parallel Simulator (LAMMPS), version 22 Jul 2025.
LAMMPS runs the potential as pair_style eam/fs, the embedded atom method (EAM)
style for Finnis-Sinclair tables.
CppCore/tests/data/sutton_chen/make_lammps_ref.py writes the table and the two
108-atom references, with a cutoff of 5.5 Angstrom.
Cell |
energy gap (eV) |
largest force gap (eV/Angstrom) |
largest stress gap (eV/Angstrom3) |
|---|---|---|---|
cubic |
3.4e-13 |
2.0e-11 |
8.3e-14 |
sheared |
2.2e-12 |
2.8e-11 |
2.0e-13 |
Tables of 12501, 25001 and 50001 points move the LAMMPS stress by at most 5.8e-13 eV/Angstrom3, so the gaps above are not set by the table. The test also takes each Voigt component as a central difference of the energy under a symmetric strain of 1e-6.
The last test takes the pressure of the perfect face-centered cubic (fcc)
platinum crystal at the lattice parameter a = 3.92 Angstrom.
Sutton and Chen fit c to zero pressure of the infinite crystal at a.
Their c = 34.408 is that zero-pressure value rounded to three decimals.
The fcc lattice sum with n = 10 and m = 8 gives c = 34.408374.
The published c therefore leaves about 0.0015 GPa in the infinite crystal at a.
With the cutoff at 0.49 of the cell edge, rgpot gives
Atoms |
Cutoff (Angstrom) |
Pressure at a (GPa) |
Pressure at 0.99 a (GPa) |
|---|---|---|---|
256 |
7.6832 |
0.0864 |
9.3308 |
864 |
11.5248 |
0.0155 |
9.2560 |
2048 |
15.3664 |
0.0053 |
9.2446 |
4000 |
19.2080 |
0.0027 |
9.2423 |
The pressure at a falls toward that 0.0015 GPa as the cutoff grows, not toward zero. The test asks it to shrink from 256 to 2048 atoms and to stay under 1e-3 of the stress at 0.99 a, about 0.009 GPa.
Cost¶
Pinned to one core, the flag adds 5.9 to 8.5 percent to a call. A jittered 864-atom cell at a cutoff of 7 Angstrom takes about 880 ns per atom without the stress and 930 with it. A 4000-atom cell with a cutoff of 19 Angstrom takes about 25800 and 27500. With the flag clear, six cases from 108 to 4000 atoms take the same time as before the flag existed, to within 4 percent.