Tunnelling splittings¶
A two-level system (TLS) in a glass is a pair of adjacent minima that the structure tunnels between at about one kelvin. Its tunnelling splitting \(\Delta_0\) and asymmetry \(\Delta\) set the TLS energy \(E = \sqrt{\Delta^2 + \Delta_0^2}\). eOn estimates \(\Delta_0\) in two ways:
NEB band (WKB) |
|
|
|---|---|---|
Path |
the minimum energy path |
the path of least imaginary-time action |
Dimensions |
one, along the mass-weighted band |
every free degree of freedom |
Cost |
the NEB itself |
one batch of forces per iteration over the beads, plus a Hessian per bead |
Output |
first frame of |
|
With mode = rate the same job estimates the thermal rate through a saddle
instead of a splitting. That path is below.
Energies are in eV, lengths in Å and masses in amu throughout, so mass-weighted lengths are in amu^0.5 Å.
WKB along a NEB band¶
Every nudged elastic band (NEB) job writes reaction_coordinate_mw, the mass-weighted arc length, on
each frame of neb.con. The first frame also carries the band’s Wentzel-Kramers-Brillouin (WKB) estimate.
Every band variant eOn runs carries them, the off-path climbing image band
[INST_GGJonsson26] among them. How eOn finds the
minima and saddles these jobs start from is set out in
Goswami [INST_Gos25].
The keys are:
Key |
Meaning |
|---|---|
|
Wells: \(\hbar\omega\) of each well along the band, from a fit of \(a s^2 + b s^3\) to the images within half the barrier |
|
Action: \(S = \hbar^{-1} \int \sqrt{2 (V(s) - E)}\, ds\) over the forbidden region |
|
Estimate: \(\Delta_0 = (\hbar\omega / \pi) e^{-S}\), with \(\omega\) the geometric mean of the wells |
|
Asymmetry: \(\Delta = V_p - V_r + (\hbar\omega_p - \hbar\omega_r)/2\), the difference of the two local ground states |
|
Energy: \(\sqrt{\Delta^2 + \Delta_0^2}\) |
|
Flag: 1 when both barriers exceed \(\hbar\omega\); below that, WKB is the wrong tool |
The profile between images is a monotone cubic, so it cannot dip below the data. The level \(E\) is the higher of the two harmonic ground states. A structure without masses, or a band whose end is flat, leaves these keys out.
\(\Delta\) is the diagonal term of the two-level Hamiltonian [INST_AHV72], the eq. S9 of Khomenko et al. [INST_KSB+20]. A tilt changes the curvature of each well as well as its depth: on a quartic double well tilted by 1 meV, the difference of the minima alone puts the energy 6 percent above the exact gap, and \(\Delta\) within 0.6 percent.
The band solved in one dimension¶
The same frame carries the exact levels of the one-dimensional problem along the band, \(-\tfrac{\hbar^2}{2} \partial_s^2 + V(s)\) in the mass-weighted coordinate, the protocol of the glass two-level-system studies [INST_KSB+20, INST_MBC+23]. The keys are:
Key |
Meaning |
|---|---|
|
Energy: \(E_2 - E_1\) of the two lowest levels, eV |
|
Asymmetry: \(E \cos 2\theta\), positive when the product well lies higher, eV |
|
Matrix element: \(\Delta_0 = E \sin 2\theta\), eV |
|
Flag: 1 when the outer walls were sampled, below |
|
Flag: 1 when the images rise to one maximum and fall, a double well and no more |
The kinetic energy is the fourth-order central difference on a uniform grid of 20 points per oscillator length \(\hbar / \sqrt{\hbar\omega}\) of the stiffer well, at most 1500 points, reaching seven such lengths past each minimum. On a quartic double well, symmetric or tilted, the gap lands within \(10^{-5}\) of a converged grid. \(\theta\) rotates the two lowest states into the pair most localised on either side of the barrier top. That rotation cancels the tail each state leaves across the barrier, so on a tilted quartic \(\Delta_0\) stays within \(10^{-3}\) of the gap of the level wells with the asymmetry thousands of times larger, where the ground state’s weight on each side alone would overshoot fourfold.
A band runs from minimum to minimum and never sees the outer walls. Past
each end the profile continues by the fit of that well, its cubic or its
parabola, whichever is stiffer; on a quartic double well that keeps the gap
within 6 percent. [Nudged Elastic Band] tunnel_wall_points above 0 samples
the real walls instead, the linear extrapolation of the reaction path of
Khomenko et al. [INST_KSB+20]: that many force calls
past each minimum along the band’s end segment continued outward, out to
seven oscillator lengths, once the band has converged. The fit stays as the
model and the residual of the samples, a cubic spline flat at the minimum,
is added to it. Eight samples bring the quartic within 1.5 percent of the
exact gap, the rest being the band’s own images.
[Nudged Elastic Band]
images = 13
tunnel_wall_points = 8
Either solution stays one-dimensional; WKB along the band is exact in one dimension up to its semiclassical error. When the path curves, the tunnelling cuts the corner, and the transverse zero-point energy changes along the way. In the two-dimensional test valley below, both effects together put the WKB band estimate a factor of 2.8 below the exact splitting. For both, run the instanton job on the pair.
The instanton job¶
The instanton is a path of \(P + 1\) beads in imaginary time \(\beta\hbar\), with its ends fixed at the two minima. It minimises the discretised Euclidean action
in mass-weighted coordinates \(q\). The splitting comes from the ratio of the off-diagonal to the diagonal imaginary-time propagator, both taken in the same steepest-descent approximation.
Here \(J\) is the Hessian of \(S\) over the interior beads. The prime leaves out its zero mode, the kink’s position in imaginary time. \(S_0 = \int |\dot q|^2 d\tau\), and \(J_\text{well}\) is the same Hessian with every bead at a minimum.
[Main]
job = instanton
[Potential]
potential = rgpot
[Instanton]
reactant_filename = reactant.con
product_filename = product.con
; start from a converged band instead of the straight line
initial_path = neb.con
beads = 256
springs = trotter
beta_hbar_omega = 30
force_tolerance = 1e-3
; one finite-difference Hessian every 4 beads, linear in between
hessian_stride = 4
springs = eco is refused. The instanton is a Trotter discretisation of
the ring.
The reactant and product must have the same ordered species, masses, cell,
and fixed-atom masks. With a frozen environment, its positions must also
match within 1e-8 Å under the minimum image. The same requirements apply
to saddle_filename in rate mode. A mismatch raises instanton: reference surface differs before reference energies or Hessians are evaluated.
For example, independently relaxing frozen atoms in saddle.con changes
the constrained potential: those coordinates cannot supply a saddle
energy for a ring whose frozen atoms stay at the reactant positions.
Generate and relax both references with the same frozen environment.
beta_hbar_omega sets the imaginary time in units of \(1/\omega\) of
the stiffer minimum along the line between them. It must be long enough for
the kink to relax into both wells. results.dat reports
instanton_mode_separation, the second eigenvalue of \(J\) over its zero
mode. Values above \(10^3\) mean the kink is isolated; small values mean
\(\beta\hbar\) is too short. With no atom fixed, the product is aligned
to the reactant first: its mass-weighted mean displacement is removed, and
for a cluster its best rotation as well.
With that reactant rotation removed, each iteration evaluates every interior bead of the kink in one call. Under
[RgpotPot] ranks_per_image, that call spreads the beads over the Car-Parrinello molecular dynamics (CPMD)
calculator groups the same way a NEB spreads its images.
instanton.con writes one frame per bead, with imaginary_time_fs. The keys are:
Key |
Meaning |
|---|---|
|
Splitting: \(\Delta_0\), eV |
|
Action: \((S - S_\text{well})/\hbar\) |
|
Energy: \(\sqrt{\Delta^2 + \Delta_0^2}\) with \(\Delta\) the zero-point asymmetry below, eV |
|
Asymmetry: \(V(\text{product}) - V(\text{reactant})\), eV |
|
Asymmetry of the local ground states: \(V_p - V_r + \tfrac{\hbar}{2}\left(\sum_k \omega_k^{p} - \sum_k \omega_k^{r}\right)\) over every vibration of the two minimum Hessians, eV |
|
Temperature: \(1/(k_B \beta)\) for the imaginary time used |
|
Separation: how well the kink’s translation separates from the other modes |
|
Symmetry: 1 when {math}`\beta |
|
Magnitude: {math}`\beta |
|
Levelled: 1 when the instanton ran on the levelled surface below |
Two minima of one energy can still hold different zero-point energies, when an isotope or the cage around a defect makes one well stiffer than the other. Then \(\Delta_0\) is the tunnelling matrix element taken with the geometric mean of the two well kernels, and the levels split by \(\sqrt{\Delta^2 + \Delta_0^2}\) with the harmonic zero-point difference in \(\Delta\) [INST_JLR20]. On a valley whose transverse stiffness differs by 10 percent between the wells, that difference splits the levels 140 times further than tunnelling does, and \(\sqrt{\Delta^2 + \Delta_0^2}\) lands within 6 percent of the exact gap.
Most two-level systems in a glass also have minima of different energy,
with the asymmetry comparable to the splitting. On such a surface no kink
joins the two minima at rest, and the propagator ratio measures
\(\Delta_0\) only while the wells lie within a small fraction of
\(k_B T\) of each other. The job therefore levels the wells first. It
subtracts \((V_p - V_r)\,\sigma(\xi)\), with \(\xi\) the projection
of \(q\) on the line from the reactant to the product (0 at one, 1 at
the other) and \(\sigma = 6\xi^5 - 15\xi^4 + 10\xi^3\). Both minima
stay stationary with their Hessians, since \(\sigma\), \(\sigma'\)
and \(\sigma''\) vanish at both ends, and the instanton of the levelled
surface gives the tunnelling matrix element. The difference of the minima
goes back into \(\Delta\), so moving the bias out of the path and into
the asymmetry, as
Eraković and Cvitaš [INST_ErakovicCvitavs22] do for the coupling
of nondegenerate well states. That holds while \(|V_p - V_r|\) is small
against the barrier, the regime of a two-level system. On a quartic double
well tilted until \(\beta |V_p - V_r| = 0.47\), the matrix element stays
within 1 percent of the level wells’ and \(\sqrt{\Delta^2 + \Delta_0^2}\)
within 2 percent of the exact gap. instanton_action is then the action on
the levelled surface, and the frames of instanton.con keep the energies of
the real one.
symmetrize = false keeps the bare surface. instanton_symmetric = 0 then
flags a pair outside the window. The job still writes the path and the
action, but no tunnel_splitting_instanton, and reports success: the flag
says why. For such a pair, set mode = rate, give the saddle and a
temperature below the crossover, and read the rate in the next sections.
Which path object¶
NEB images and ring-polymer beads are different objects. An image is a point on a path in configuration space between two minima. Its springs are fictitious. The parallel force is removed. A bead is one imaginary-time slice of a single quantum system. Its springs are physical, with stiffness fixed by the temperature and the number of beads.
The columns are:
Object |
Points |
Springs |
What it returns |
|---|---|---|---|
|
open string between two minima, started from a band when one is present |
Euclidean action, no tangent projection |
tunnelling splitting when the wells are close in energy |
|
closed ring through one saddle |
same action, stiffness set by \(T\) and \(N\) |
thermal rate: the ring below the crossover, the parabolic factor above it |
Centroid potential of mean force (PMF) |
one ring per image, centroid held on the image |
sampled, not minimised |
quantum free-energy barrier along the path |
Harmonic centroid string |
a ring at each image, optimised |
local harmonic quantum correction |
a free-energy estimate as good as that harmonic well |
The first two rows are job = instanton. The centroid potential of mean
force is a constrained path-integral molecular dynamics sample, one
thermostatted ring per image. That sample is not an optimisation, and it
does not belong in this job. A string of harmonically corrected rings is a
different calculation again, and this page does not implement it.
The rate below the crossover¶
mode = rate reads the reactant and saddle_filename (default saddle.con).
Any eOn saddle search can supply it, the Gaussian process accelerated dimer
[INST_GJonsson25, INST_GMK+25]
among them.
The instanton is a closed ring, a first-order saddle of the ring-polymer
potential, with one negative eigenvalue and one zero eigenvalue that cycles
the beads. Richardson and Althorpe [INST_RA09] give
\(B_N\) is the sum of squared steps around the ring, \(\omega_k^2\) are the eigenvalues of the mass-weighted ring Hessian, and the prime leaves out the cyclic zero mode and the rigid translations and rotations. \(Z_r\) is the harmonic ring-polymer partition function of the reactant. \(U_N\) is the ring-polymer potential, the bead potentials plus the springs.
A ring of up to 4096 coordinates (\(N f\)) uses the dense ring Hessian, with the cycle and the rigid modes lifted. Its eigenvalues give \(\det'\), the count of negative modes and the lowest eigenvalue. The block-chain determinant must agree with them; otherwise the log says so and that temperature gets no rate. A larger ring takes the determinant and the inertia from the block chain, and the lowest eigenvalue from Lanczos. A 16-bead ring of a 254-atom cell with every atom free has 12192 coordinates.
The crossover temperature is \(T_c = \hbar \omega_b / (2\pi k_B)\), with \(\omega_b\) the imaginary frequency at the saddle. At or above \(T_c\) the ring collapses onto the saddle. Above \(T_c\) the rate is the parabolic barrier factor
times the quantum harmonic transition-state theory rate from the same Hessians. The job writes the parabolic barrier factor. It does not search a ring at that temperature. \(\kappa\) tends to 1 at high temperature, which is the one-bead limit. The factor diverges as \(T\) approaches \(T_c\) from above.
[Main]
job = instanton
[Instanton]
mode = rate
reactant_filename = reactant.con
saddle_filename = saddle.con
temperature = 5
beads = 256
hessian_stride = 1
temperature is in kelvin. It must be positive. Below \(T_c\) the job
optimises the ring. Above \(T_c\) the job writes the parabolic barrier
factor and does not optimise a ring. At \(T_c\) the factor diverges
and that temperature records no rate. The optimizer is not what keeps the path unused. Each
bead still needs a force. The rate uses the fluctuation prefactor from the
bead Hessians, and above the crossover the ring collapses onto the saddle.
The default 0 means the temperature was not set, and
mode = rate then refuses to run. beads defaults to 256, the same default as the splitting. A ring
of 32 beads at 5 K does not resolve a stiff bond: the path integral starts
to converge once the bead count exceeds \(\beta \hbar \omega\) of the
stiffest mode. hessian_stride of 1 takes a Hessian on every bead. A larger
stride keeps a Hessian on every stride-th bead and interpolates linearly
between those anchors. That interpolation is an approximation: the rate
formula uses the Hessian of every bead.
half_ring defaults to enabled. On an even bead count the potential is
evaluated from one turning point to the other and copied onto the mirror.
The reaction coordinate on that half stays monotone, so the search cannot
settle on an out-and-back bounce.
A half ring that has converged, or that has stopped with the gradient
under the force tolerance while the negative-mode count is not 1, is
checked for an unstable mode that is odd under the mirror when the
interior beads still match their mirrors. Two copies of the instanton
on one ring are that mode. The search steps along it and continues on
the whole ring.
An odd count keeps every bead. energy_shift (default 0, in eV) is
subtracted from every bead potential and from the reactant and saddle
energies in the rate. An empty discretization keeps every spring equal.
A comma-separated list of one positive weight per bead divides the spring
from that bead to the next. A half ring keeps the uniform spring.
friction adds the fixed quadratic bath described under
“Fixed bath coefficients and physical units”.
The search is an index-1 Newton step on the ring Hessian while the active
coordinate count is within the Newton limit. Past that limit the ring is one
structure and the dimer follows its unstable mode, with the beads in one force
batch. Without an
initial_path, the rate job traces a steepest-descent path out of the
saddle along both signs of its unstable mode and seeds the ring from it by
the period condition below. That path is written to instanton_sd_path.con,
one frame per point, with the energy and the arc length. A later rate job
reads it as initial_path and does not repeat those force calls. On LJ13
that path costs 586 gradient calls. Cooling a cosine
ring from 0.85 of the crossover is the fallback when no path can be
built; it finds the ring only where the ring grows continuously out of
the saddle as the temperature drops. Where it does not, the search walks
to a neighbouring saddle, so a converged ring that fails the channel test
described with the results.dat keys below is refused and no rate is
written. The step climbs the mode that overlaps
the last climb and turns every other negative curvature downhill. The
rigid motions of the whole ring, rebuilt from the current beads, are left
out of the step; the imaginary-time cycle keeps its component, on its own
curvature, and is lifted to a spring-sized curvature only while that
curvature is near zero. A
converged gradient is classified with finite-difference bead Hessians
before it counts: the Bofill blocks learn curvature only along the steps
taken, so they can carry negative curvatures the surface does not have and
miss ones it does. A second negative curvature that survives is a
higher-index stationary ring, and the search steps down that mode by a
trust-sized displacement. The ring Hessian is block cyclic
tridiagonal in the beads, and every solve, determinant and inertia count
goes through a block LU of the open chain plus a low-rank Woodbury
correction for the closure, the cycle, the rigid modes and each
eigenvector-following flip: \(O(N f^3)\) for \(f\) degrees of
freedom, and the search never forms the \(Nf \times Nf\) matrix, so the same
step serves a seven-atom cluster and a 254-atom cell. The lowest ring
modes come from Lanczos on matrix-vector products. The bead curvature
blocks start from the saddle Hessian (initial_hessians = saddle, no force
calls) and follow accepted moves with a Bofill update, rebuilt from finite
differences when their count of negative curvatures grows and up to three
times when the trust radius reaches its floor. Below the crossover the
copied blocks already hold a negative curvature in the ring’s slowest
mode, so the first entry rebuilds them, and saddle costs the same force
calls as finite_difference. Known rigid motions enter an explicit quotient
and solve stabilization. A small saddle curvature alone does not identify a
rigid motion and is retained in the physical ring inertia.
initial_hessians = finite_difference takes
\(2f\) gradient calls per
bead first. Between rebuilds each block keeps the number of negative
curvatures its last finite-difference version had: a Bofill update that
opens another one has its sign flipped, and when the
ring model still shows more negative curvatures than the last exact
blocks, the blocks are rebuilt (at most four times). The search stops
when \(B_N\) falls below collapse_fraction (default 1e-4) of the
starting ring’s, since the beads have then fallen onto one point; the
job writes instanton_collapsed = 1 and no rate. The rate uses the bead Hessians chosen by hessian_final, not
that update. On the one-dimensional Eckart barrier the search converges
in 9 to 12 steps from the cosine seed and in 3 from a seed on the path.
newton_limit is the maximum number of active ring coordinates for the
dense Newton route. Its default is the 64-bit maximum, which preserves
dense Newton selection on Windows as well as on LP64. Set newton_limit = 0 to select
the existing matrix-free minimum-mode dimer search; negative values are
rejected. For a frictionless even half ring, the active count is
(P/2 + 1) * f; a full ring uses P * f. A nonlocal bath can select the
dimer independently of this limit. This setting changes the optimizer,
not the force criterion or the reference-spectrum checks. Final rate
prefactors still require the configured physical bead Hessians.
Independent reference calculations can supply the two reference Hessians:
[Instanton]
reactant_hessian_filename = reactant-hessian.dat
saddle_hessian_filename = saddle-hessian.dat
Each file is a plain, whitespace-separated square matrix over the free
atoms in their input order, three Cartesian coordinates per atom. Its
entries are mass weighted, in eV/(amu Ų). For each matrix, provide the
measured geometry as an ordinary CON file with .con appended to the
matrix filename, for example reactant-hessian.dat.con. The job checks
the complete geometry, cell, ordered species, atom IDs, masses and fixed
masks, then checks matrix dimensions, finite entries and symmetry to
roundoff. It retains every signed curvature for the reference-spectrum
checks. An absent filename requests the physical finite-difference
calculation.
The CON sidecar cannot establish an external calculator’s Hamiltonian. Supply a matrix only when the calculation has matching pseudopotentials, charge, functional, boundary conditions and converged force/displacement settings. Preserve that calculation’s provenance with the matrix. A matching file avoids repeating its reference force displacements; it does not replace final physical bead Hessians or prove reference stationarity. The job still evaluates both reference energies and forces on the selected potential and applies its rigid-mode projection to the loaded matrices.
An unexpected negative-mode count can also trigger a complete physical
Hessian refresh during optimization, including at iteration zero.
For a half ring with even bead count P, this evaluates P/2 + 1
independent bead Hessians; central differences cost 2f force geometries
per bead for f free Cartesian coordinates. hessian_stride controls the
final rate-prefactor Hessians and does not reduce this optimizer refresh.
lanczos_step = 0.0001 sets the central directional displacement for
Lanczos Hessian-vector products in the dimer search. It must be finite and
positive and has units sqrt(amu) Angstrom. It is independent of
physical_hessian_step. Choose it by measuring directional-curvature
convergence with displacement and electronic accuracy: a central gradient
difference amplifies bounded gradient error eta by at most eta/h. An SCF
residual alone does not establish that gradient-error bound.
physical_hessian_step = 0.0001 sets every central physical bead-Hessian
displacement during Newton optimization, including initial and adaptive
refreshes. Its units are sqrt(amu) Angstrom, so the Cartesian step for
an atom of mass m in amu is this value divided by sqrt(m). This default
is independent of the directional Lanczos step and of [Main] finite_difference.
It preserves the numerical displacement convention; convergence of force
differences with step size and SCF accuracy still needs measurement. Central
differencing amplifies a bounded gradient error eta by at most eta/h in norm;
an electronic SCF residual alone does not supply that gradient-error bound.
temperatures is a comma-separated list in kelvin. The search starts at
the highest and each ring starts the next, colder one. An empty list uses
temperature. bead_ladder (default off) starts a ring of at least 16
beads at a quarter of that count and doubles. hessian_final is
recomputed: the rate takes a Hessian on every stride-th bead.
For a geometry warm start of a rate calculation, set initial_ring to a
saved instanton.con. The file contains exactly beads frames of the closed
ring, without an extra copy of the first frame at the end. Every frame must
have the reference cell, ordered species, atom IDs, masses, constraints and
frozen coordinates. Cached energies and forces are ignored. The selected
potential evaluates the ring again; optimizer history is not restored.
For example, copy a saved 32-bead ring to seed-ring.con in a separate run
directory, retain the same reference and potential files, and set:
[Instanton]
mode = rate
beads = 32
initial_ring = seed-ring.con
The seed can also start a calculation at a different temperature. Without
initial_path, a supplied ring skips construction of the steepest-descent
seed path. initial_path can be supplied as well when its separate WKB path
estimate is wanted. A saved unconverged ring remains a geometry seed; it does
not establish a rate or convergence.
The optional physical_bath_filename selects the physical-bath model for
rate mode. Its coordinate mapping and callback convention must match the
selected potential. A fitted potential is identified as a fitted calculation;
its native validation and bead convergence require separate checks.
On a rate calculation, initial_path is a band over the barrier. The ring
starts on the closed orbit of that band whose period is \(\beta \hbar\),
and the same band carries a one-dimensional WKB rate.
With fixed atoms, each band frame is projected onto the reactant’s frozen
coordinates before its potential is evaluated. Stored frame energies are
reused only if every frame has a finite energy and matches the reactant’s
cell, ordered species, constraint masks and frozen coordinates. Cell entries
and frozen displacements use a tolerance of 1e-8 Angstrom; periodic frozen
displacements are compared under the minimum image. A mismatch in any frame
causes all band energies to be evaluated on the projected surface. Mobile
coordinates still supply the path and ring seed.
This geometry check does not establish external Hamiltonian provenance. Cached energies must also use the selected potential, electronic settings and pseudopotentials. When that provenance is uncertain, remove the stored energies from the input frames so the job evaluates the entire profile.
rate_instanton.dat has one row per temperature, with columns T_K,
T_c_K, beads, converged, iterations, U_N_eV, negative_modes,
ln_k_per_s, k_per_s, ln_k_htst_per_s, barrier_effective_eV,
ln_k_wkb_path_per_s, ln_k_parabolic_per_s and parabolic_factor.
Above \(T_c\) the instanton columns are empty and the last two hold
the parabolic rate. Below \(T_c\) those two are empty. At \(T_c\)
the parabolic columns are empty. A run at more than one temperature also writes
instanton_<T>K.con. The coldest temperature is written to instanton.con.
The rate calculation retains instanton_reactant_hessian.dat and
instanton_saddle_hessian.dat, with their sorted eigenvalues in
instanton_reactant_eigenvalues.dat and
instanton_saddle_eigenvalues.dat. These are the projected mass-weighted
matrices used by the rate calculation, written with enough digits to
round-trip a double. They are retained even when reference-spectrum
validation rejects the rate. The reactant must be a metastable minimum
and the saddle must have exactly one unstable vibration after the
applicable rigid modes are removed. With any frozen atoms, no whole-system
rigid modes are removed; small negative curvatures require a force and
finite-difference convergence check rather than an assumed six-mode removal.
A valid Hessian index alone does not establish a stationary reference.
instanton_reactant_forces.dat and instanton_saddle_forces.dat retain
the physical free-atom force vectors in original free-atom order, in
eV/Angstrom. Their Euclidean norms and maximum atomic norms appear in
results.dat as reference_reactant_force_norm,
reference_reactant_max_force, reference_saddle_force_norm and
reference_saddle_max_force, including a spectrum-rejected calculation.
The cached energy evaluations supply these forces without additional
potential calls. Check stationarity against the physical convergence
criterion used to prepare the minimum and saddle before interpreting a rate.
Rate reference energies and forces are recomputed on the selected potential;
evaluations embedded in the input CON files are not accepted as references.
With no atom fixed, both the reactant and the instanton omit the three
translations. A
rotation is omitted when it is a zero mode of the reactant Hessian, which a
free cluster has and a crystal does not: when its curvature r^T H r / r^T r
is at most rotation_zero_fraction (default 0.1) of the softest
vibration, the lowest eigenvalue of the
Hessian once the translations and rotations are projected out. Both sides
scale with the Hessian and neither grows with the atom count. A cluster in a large periodic cell
is told apart by that Hessian, not by the periodic flag. The springs along
those directions stay, so they cancel between the instanton and the reactant.
On the ring the omitted rotations are those of the beads themselves about
the ring’s centre of mass: a rotation moves each bead by a different
amount, so the reactant’s rotation copied to every bead is not a zero mode
of the ring. The bead Hessians keep their rotational curvature, which
balances the springs on a bead that is not a minimum, and lose only their
translations. The omitted rotations come back as the ratio of the
classical rotational partition functions: the square root of the product
of the ring’s principal moments of inertia, every bead’s atoms about the
ring’s centre of mass, over the same product for the reactant repeated on
every bead. That is the moment of inertia of the ring polymer averaged
over its beads, as the ring-polymer instanton rate takes it for a
molecule that turns freely. instanton_rotation_ratio_log reports its
logarithm. The classical comparison rate_htst_log and the parabolic rate
above \(T_c\) take the saddle’s ratio, so all three rates count the
rotations alike. A linear structure keeps the two moments perpendicular to
its axis.
results.dat reports the rate. The keys are:
Key |
Meaning |
|---|---|
|
\(k\) in s^{-1} |
|
\(\ln(k\,/\,\mathrm{s}^{-1})\) |
|
classical harmonic transition-state theory, the same logarithm |
|
\(\kappa\), above \(T_c\) |
|
\(\kappa\) times the quantum harmonic TST rate, s^{-1} |
|
\(\ln(k\,/\,\mathrm{s}^{-1})\) of that rate |
|
\(T_c\), K |
|
negative eigenvalues of the ring Hessian; a first-order saddle has 1 |
|
the eigenvalue left out |
|
ln of the ring’s rotational partition function over the reactant’s; 0 when no rotation is free |
|
the turning points along the saddle’s unstable mode, amu^0.5 Angstrom from the saddle |
|
cosine of the angle between the chord joining the turning points and the unstable mode |
|
1 when the search stopped because the beads fell onto one point (B_N below |
|
largest distance from the saddle, across the mode, at which the ring crosses the dividing plane |
A search stops at once when its beads fall onto one point, with
\(B_N\) below collapse_fraction of the starting ring’s. That ring gets no rate.
results.dat records instanton_collapsed 1 and termination_reason 1
(FAIL_MAX_ITERATIONS). The log names where along the unstable mode the
beads fell.
A converged ring is given a rate only when it belongs to the seeded saddle:
it straddles the plane through the saddle normal to the unstable mode,
crosses that plane no farther from the saddle than its own span
instanton_s_max - instanton_s_min, and its chord lies within 60 degrees
of the mode. A ring that slid to a neighbouring saddle can still straddle
the plane and have one negative mode, and then its rate would belong to a
different reaction.
Habershon et al. [INST_HMMM13] expect the sampled ring-polymer rate to lie within about a factor of two of the exact quantum rate between \(T_c\) and \(T_c/2\). That bound compares the sampled rate with the exact rate. It is not a comparison of this instanton with a free-energy profile.
The cubic metastable well, \(V = \omega_0^2 q^2/2 - g q^3/3\), is the check. Deep below the crossover its rate approaches the zero-temperature decay of Caldeira and Leggett [INST_CL83].
Path-integral quantum TST on planes¶
With pi_planes above 0, mode = rate follows the instanton with
path-integral quantum transition-state theory (PI-QTST,
Voth et al. [INST_VCM89]; review in
Voth [INST_Vot93]). A ring polymer is sampled with
its centroid held on each of a set of parallel planes. The centroid
potential of mean force along the plane coordinate gives a free-energy
barrier and a rate at every temperature in temperature or temperatures.
[Instanton]
mode = rate
saddle_filename = saddle.con
temperatures = 150, 105
pi_planes = 21
pi_beads = 32
pi_equilibration_steps = 500
pi_sampling_steps = 4000
pi_time_step = 0.5
pi_thermostat = pile
pi_direction = mode
pi_reactant_extent = 0.5
The coordinate. With \(q\) the mass-weighted displacement from the
reactant over the free atoms and \(\hat n\) a unit vector in those
coordinates, the plane coordinate is \(s = \hat n \cdot q\), in
amu^0.5 Å. The reactant sits at \(s = 0\) and the saddle at
\(s^* = \hat n \cdot q_\mathrm{saddle}\). pi_direction = mode (the
default) takes \(\hat n\) from the unstable eigenvector of the saddle’s
mass-weighted Hessian, oriented toward the saddle, so the last plane is the
dividing surface normal to the barrier mode. When that mode makes more than
60 degrees with the reactant-saddle line, the job warns and uses
pi_direction = line, the straight mass-weighted line from the reactant to
the saddle. (A saddle with no negative eigenvalue has no crossover
temperature, and the job stops before the planes.) One normal serves every
plane, so \(s\) is a linear coordinate and its mean force integrates to
its free energy with no metric correction. The planes are fixed in the reactant’s frame. Translations
of a structure with no atom fixed lie within every plane, because the
unstable mode of the projected Hessian has no rigid component, but a
rotation of a free cluster changes \(s\); fix an atom, or use a cell,
when the sampling is long enough for the cluster to turn. The job warns
when no atom is fixed in an aperiodic cell. pi_planes planes are spaced
uniformly from \(s_0 = -\,\mathtt{pi\_reactant\_extent}\; s^*\), behind
the reactant, to \(s^*\).
The sampling. On each plane the ring’s centroid starts where
initial_path crosses the plane, or on the reactant-saddle line when no
band is given. Projecting the centroid position and momentum holds it on
the plane. The ring is carried from plane to plane: its centroid moves to
the next plane and its internal modes keep their thermalised state.
The free-ring springs are propagated exactly, so pi_time_step is
limited by the physical vibrations, as for classical dynamics.
pi_equilibration_steps are discarded, then pi_sampling_steps steps of
pi_time_step fs record \(\hat a \cdot f_c\), the centroid force along
the plane’s Cartesian normal \(a = M^{1/2} \hat n\). The bead forces of
a step are one batch, so a calculator group carries the beads. pi_thermostat = pile puts PILE on the internal
modes and a Langevin thermostat of time pi_pile_tau fs on the centroid
within the plane, and pi_pile_scale (default 1) scales the critical
damping of the internal modes. Below the crossover the lowest ring modes
are soft at the barrier and overdamped at critical damping; on the Eckart
check 0.5 cuts the mean-force error at 0.7 \(T_c\) by a third. piglet reads a normal-mode GLE from pi_gle_file, in
the same format and with the same meaning as [Dynamics] path_gle_file.
pi_seed seeds the noise. PIGLET tunes the internal modes so that bead
averages converge with fewer beads; the centroid free energy assumes the
ring polymer’s own Boltzmann distribution, so treat a piglet barrier as
an approximation and check it against pile. The transmission factor’s
parents always take pile.
The free energy. The mean force on the plane at \(s\) is
in eV per amu^0.5 Å. Any vector \(u\) with \(a \cdot u = 1\),
\(M^{-1/2} \hat n\) among them, gives the same mean, because the force
within the plane averages to zero; the normal component is the estimator
the job records. \(F(s)\) is the trapezoid integral of \(F'\) from
\(s_0\). The production run is cut into ten equal blocks. The standard
error of the block means is the error of \(F'(s)\). The errors of
\(F\) and of the rate follow by linear propagation, with the planes
taken as independent. The quantum free-energy barrier is
\(\Delta F = F(s^*) - \min_{s < s^*} F(s)\), the minimum over the
planes before \(s^*\); the job warns when it is not positive, since
\(s^*\) is then no barrier on the centroid free energy. The log and results.dat
give it beside the classical barrier
\(V(\mathrm{saddle}) - V(\mathrm{reactant})\) and the instanton’s
effective barrier.
The rate. With \(s\) a unit-mass coordinate,
with \(\beta = 1/k_B T\) in eV^{-1}. The prefactor is \(\tfrac12\langle|\dot s|\rangle\), half the mean speed of a free unit mass, in amu^0.5 Å per time unit of \(\sqrt{\mathrm{amu}\,\mathrm{Å}^2/\mathrm{eV}}\) (10.18 fs). The integral, again the trapezoid rule over the planes, is the reactant’s centroid density along \(s\). The job reports \(k\) in s^{-1}. The integral stops at the first plane, so that plane must sit several \(k_B T\) above the reactant minimum of \(F\); the job warns below 5 \(k_B T\). It also warns when \(F'(s^*)\) differs from zero by more than three standard errors: the maximum of the centroid free energy then lies off the classical saddle’s plane.
The formula is classical TST on the centroid free-energy surface. With one bead, \(F\) is the classical free energy and the rate is classical TST along \(s\). Near and above the crossover it carries the tunnelling correction of a symmetric barrier. For strongly asymmetric barriers well below the crossover it fails [INST_Vot93], and the instanton gives the rate there.
Output. rate_piqtst.dat has one row per temperature and plane, with
columns T_K, s_amu05A, dF_ds_eV_per_amu05A, dF_ds_error, F_eV,
F_error_eV and spread_max_A. piqtst_planes.con holds one frame per
plane at the coldest temperature, at the production-averaged centroid. Its
readcon spreads section holds each atom’s root-mean-square bead
displacement from the centroid along x, y and z in Å. The scalars carry
\(s\), \(F'\) and \(F\) with their errors, the temperature and
the bead count. A run at more than one temperature also writes
piqtst_planes_<T>K.con. results.dat gains, for the coldest temperature:
Key |
Meaning |
|---|---|
|
\(\Delta F\), eV |
|
its standard error, eV |
|
\(k_\mathrm{PI\text{-}QTST}\), s^{-1} |
|
\(\ln(k\,/\,\mathrm{s}^{-1})\) |
|
standard error of that logarithm |
|
\(s^*\), amu^0.5 Å |
|
\(F'(s^*)\) and its error |
|
\(\beta (F(s_0) - \min F)\) |
|
the run |
Cost: pi_planes times (pi_equilibration_steps plus pi_sampling_steps
plus 1) force batches of pi_beads beads per temperature: one per step,
and one when the ring moves to a plane.
Recrossing and the RPMD rate¶
PI-QTST counts every centroid that reaches \(s^*\) moving forward as
reactive. Some of those trajectories turn back. With
pi_recrossing_parents above 0 the job measures that fraction, the
Bennett-Chandler transmission factor of ring-polymer molecular dynamics
(RPMD, Craig and Manolopoulos [INST_CM05]; the two-step scheme
of Suleimanov et al. [INST_SAG13]), and reports
[Instanton]
pi_recrossing_parents = 100
pi_recrossing_children = 20
pi_recrossing_time = 100
pi_recrossing_spacing = 50
The parents. A ring with its centroid held on the top plane \(s^*\),
the dividing surface of the scan, is thermostatted with PILE for
pi_equilibration_steps, as the scan’s pile is: Bennett-Chandler needs
the ring polymer’s own constrained distribution, which PIGLET does not
sample. Every pi_recrossing_spacing steps after that its
beads are one parent, pi_recrossing_parents in all.
The children. Each parent launches pi_recrossing_children momentum draws.
A draw takes every ring normal mode from the Maxwell-Boltzmann distribution
at \(\beta_P = \beta / P\) with the plane constraint removed, and runs
twice, with \(p\) and with \(-p\), which halves the variance. Each
child runs pi_recrossing_time fs of thermostat-free RPMD in steps of
pi_time_step: velocity Verlet on the physical forces with the free ring
propagated exactly in normal modes. Along the way the job records the
centroid coordinate \(s(t)\). The centroid velocity along the coordinate is
\(\dot s = \hat n \cdot M^{1/2} v_c\), and
with \(h\) the step function. At \(t = 0\) the side is that of
\(\dot s(0)\), the limit \(t \to 0^+\), so \(\kappa(0) = 1\).
\(\kappa(t)\) falls as children recross and levels off once they have
committed to a side. The reported \(\kappa\) is the mean of
\(\kappa(t)\) over the last quarter of pi_recrossing_time. Its
standard error is the jackknife over parents. Lengthen
pi_recrossing_time when kappa_piqtst.dat has not levelled off by the last
quarter.
What \(\kappa\) means. \(\kappa\) lies between 0 and 1. A value of 1 means no trajectory that crosses \(s^*\) forward returns, and PI-QTST is the RPMD rate. A value below 1 means the plane is not the dynamical bottleneck: either it is tilted from the barrier’s own dividing surface, or a bath coupling turns trajectories back. One bead gives the classical transmission through the plane. On a harmonic saddle whose plane normal makes an angle \(\theta\) with the unstable mode it is \(\sqrt{\cos^2\theta - \sin^2\theta\,\omega_b^2/\omega_\perp^2}\). The RPMD rate is independent of the choice of \(s^*\) in the long-time limit. The PI-QTST rate is not.
Output. kappa_piqtst.dat has the coldest temperature’s curve, columns
t_fs and kappa. A run at more than one temperature also writes
kappa_piqtst_<T>K.dat. rate_piqtst.dat gains the columns kappa,
kappa_error, ln_k_rpmd_s and ln_k_rpmd_s_error, the same on every
plane of a temperature. results.dat gains:
Key |
Meaning |
|---|---|
|
\(\kappa\) and its standard error |
|
\(k_\mathrm{RPMD}\), s^{-1} |
|
\(\ln(k_\mathrm{RPMD}\,/\,\mathrm{s}^{-1})\) |
|
its standard error, both errors in quadrature |
Option |
Default |
Meaning |
|---|---|---|
|
0 |
parent configurations; 0 is off, otherwise at least 2 |
|
20 |
momentum draws per parent, each run forward and reversed |
|
100 |
fs per child, at least four |
|
50 |
thermostatted steps between parents |
Cost per temperature: pi_equilibration_steps plus
pi_recrossing_parents times pi_recrossing_spacing constrained steps,
one force batch each, and 2 * pi_recrossing_parents * pi_recrossing_children
child runs of pi_recrossing_time / pi_time_step + 1 force batches each,
the first at the parent’s beads.
Centroid and spread¶
Both modes also write instanton_centroid.con (and
instanton_centroid_<T>K.con per temperature): one frame at the mean of
the beads, with the readcon spreads section holding each atom’s
root-mean-square displacement from that mean along x, y and z in Å. This is
the delocalised configuration as a centroid plus a per-atom spread, the same
representation a path-integral trajectory collapses to, so the atoms that
tunnel and how far they spread read off one frame. Below the crossover the
density is bimodal along the reaction path and the spread there is a width,
not a Gaussian; the beads in instanton.con keep the full path. The frame
carries spread_max, the largest entry, beside the temperature.
Checks¶
These checks use two Catch2 cases. The curved-valley case is Instanton splitting in a curved valley matches the exact gap. The corner case is The instanton cuts the corner the minimum energy path takes.
The potential is \(V = V_0 (x^2 - 1)^2 + \tfrac{K}{2} (y - C (1 - x^2))^2\) at unit mass
with \(K = 4\) eV/Ų. The exact gap comes from a fourth-order
finite-difference Hamiltonian, converged to \(10^{-5}\).
\(V_0\) / eV |
\(C\) / Å |
exact \(\Delta_0\) / eV |
instanton / exact |
WKB on the valley floor / exact |
|---|---|---|---|---|
0.12 |
0.35 |
2.520e-5 |
1.17 |
0.35 |
0.30 |
0.35 |
9.818e-8 |
1.08 |
|
0.12 |
0 |
2.040e-5 |
1.12 |
1.02 |
0.30 |
0 |
1.195e-7 |
1.07 |
The instanton’s error falls as the barrier deepens, the regime glass TLS sit in. The same cases tie the C++ path and splitting to an independent implementation of the discretisation to \(2 \times 10^{-3}\).
For the rate, The Eckart rate instanton matches the exact flux to its semiclassical error compares \(k Z_r\) through the symmetric Eckart
barrier \(V_0 / \cosh^2(x/a)\) (\(V_0 = 0.425\) eV, \(a = 0.734\) amu^0.5 Å, \(T_c = 150\) K) with the exact flux
\((2\pi\hbar)^{-1}\int P(E) e^{-\beta E} dE\) from Eckart’s transmission
probability, at \(T = 0.5\,T_c\) and \(0.35\,T_c\):
beads |
instanton / exact |
|---|---|
64 |
0.94 to 0.96 |
128 |
0.93 to 0.94 |
\(N \to \infty\) (1/N² extrapolation) |
0.928 |
In one dimension the instanton is the steepest-descent evaluation of the
WKB thermal integral, so its limit shares the uniform WKB error; the Kemble
integral along the path gives the same 0.928. The ring spectrum from the block chain matches the dense Hessian ties the chain’s determinant and
inertia to a dense eigendecomposition, A rigid mode leaves the instanton rate unchanged checks the rigid-mode bookkeeping through the
search, and The rate lifts the rotations of a free diatomic's ring
checks the rotations against the dense spectrum of a stretching bond.
References¶
P. W. Anderson, B. I. Halperin, and C. M. Varma. Anomalous low-temperature thermal properties of glasses and spin glasses. Philosophical Magazine, 25(1):1–9, 1972. doi:10.1080/14786437208229210.
A. Caldeira and A. Leggett. Quantum tunnelling in a dissipative system. Annals of Physics, 1983. doi:10.1016/0003-4916(83)90202-6.
Ian R. Craig and David E. Manolopoulos. Chemical reaction rates from ring polymer molecular dynamics. The Journal of Chemical Physics, 122(8):084106, 2005. doi:10.1063/1.1850093.
Mihael Eraković and Marko T. Cvitaš. Vibrational tunneling spectra of molecules with asymmetric wells: a combined vibrational configuration interaction and instanton approach. Journal of Chemical Theory and Computation, 18(5):2785–2802, 2022. doi:10.1021/acs.jctc.2c00124.
Rohit Goswami. Efficient exploration of chemical kinetics. October 2025. arXiv:2510.21368, doi:10.48550/arXiv.2510.21368.
Rohit Goswami, Miha Gunde, and Hannes Jónsson. Enhanced climbing image nudged elastic band method with hessian eigenmode alignment. Frontiers in Chemistry, 2026. arXiv:2601.12630, doi:10.3389/fchem.2026.1807063.
Rohit Goswami and Hannes Jónsson. Adaptive Pruning for Increased Robustness and Reduced Computational Overhead in Gaussian Process Accelerated Saddle Point Searches. ChemPhysChem, November 2025. doi:10.1002/cphc.202500730.
Rohit Goswami, Maxim Masterov, Satish Kamath, Alejandro Pena-Torres, and Hannes Jónsson. Efficient Implementation of Gaussian Process Regression Accelerated Saddle Point Searches with Application to Molecular Reactions. Journal of Chemical Theory and Computation, 21(16):7935–7943, July 2025. doi:10.1021/acs.jctc.5c00866.
Scott Habershon, David E. Manolopoulos, Thomas E. Markland, and Thomas F. Miller. Ring-polymer molecular dynamics: quantum effects in chemical dynamics from classical trajectories in an extended phase space. Annual Review of Physical Chemistry, 64(1):387–413, April 2013. doi:10.1146/annurev-physchem-040412-110122.
Elena Jahr, Gabriel Laude, and Jeremy O. Richardson. Instanton theory of tunneling in molecules with asymmetric isotopic substitutions. The Journal of Chemical Physics, 153(9):094101, 2020. doi:10.1063/5.0021831.
Dmytro Khomenko, Camille Scalliet, Ludovic Berthier, David R. Reichman, and Francesco Zamponi. Depletion of two-level systems in ultrastable computer-generated glasses. Physical Review Letters, 124(22):225901, 2020. doi:10.1103/PhysRevLett.124.225901.
Felix C. Mocanu, Ludovic Berthier, Simone Ciarella, Dmytro Khomenko, David R. Reichman, Camille Scalliet, and Francesco Zamponi. Microscopic observation of two-level systems in a metallic glass model. The Journal of Chemical Physics, 158(1):014501, 2023. doi:10.1063/5.0128820.
Jeremy O. Richardson and Stuart C. Althorpe. Ring-polymer molecular dynamics rate-theory in the deep-tunneling regime: connection with semiclassical instanton theory. The Journal of Chemical Physics, 131(21):214106, 2009. doi:10.1063/1.3267318.
Yury V. Suleimanov, Joshua W. Allen, and William H. Green. RPMDrate: bimolecular chemical reaction rates from ring polymer molecular dynamics. Computer Physics Communications, 184(3):833–840, 2013. doi:10.1016/j.cpc.2012.10.017.
Gregory A. Voth. Feynman path integral formulation of quantum mechanical transition-state theory. The Journal of Physical Chemistry, 97(32):8365–8377, 1993. doi:10.1021/j100134a002.
Gregory A. Voth, David Chandler, and William H. Miller. Rigorous formulation of quantum transition state theory and its dynamical corrections. The Journal of Chemical Physics, 91(12):7749–7760, 1989. doi:10.1063/1.457242.
Fixed bath coefficients and physical units¶
friction = implicit uses one nonnegative friction_eta coefficient.
friction = explicit uses the comma-separated friction_eta_beads, one
nonnegative coefficient per bead. These inputs define a fixed quadratic
bead operator. Their numerical values retain their input convention.
In mass-weighted coordinates \(q=\sqrt{m}\,r\), the implicit bath is
The sine factor is dimensionless. Thus friction_eta is a curvature
coefficient, in eV/(amu Angstrom²), equivalently inverse internal time
squared. It is not directly a damping frequency in inverse fs.
For a physical scalar Ohmic damping \(\gamma\) in mass-weighted coordinates,
the mapping at the requested temperature and bead count is
The internal time unit is approximately 10.1805 fs. A damping specified in
inverse fs converts as gamma_internal = gamma_per_fs * time_unit_fs.
Cartesian scalar friction divided by the atom’s mass gives the corresponding
mass-weighted damping. The one-coefficient API applies that mass-weighted
damping equally to all active coordinates.
At fixed physical damping, changing temperature or bead count requires
recomputing eta_eff. This also applies to a bead ladder: fixed raw input
coefficients do not define a study at fixed physical damping. The code
does not rescale existing input coefficients automatically.
For explicit coefficients, the implemented map is
where \(F\) is unitary and \(\Omega_{kk}=2|\sin(\pi k/P)|\). The coefficients remain fixed while differentiating this quadratic form. A coordinate-dependent friction tensor or frequency-dependent spectral density requires additional coupling data and derivatives; a list of fixed bead coefficients alone does not specify either model.
The physical-frequency mapping follows the uniform-coupling kernel in
Litman et al., Eqs. (16) and (20).
Their Eq. (4) supplies the collective-coordinate normalization when its
cyclic phase mode is a zero mode. For a nonzero bath, the rate calculation checks for one
physical negative mode and a numerically null phase before using that
factor. A resolved nonzero phase remains in the signed fluctuation
diagnostics; the absolute Gaussian weight excludes contour and rate
normalization. Its rateValid = false status must be retained by callers.