Simulating Photodissociation
in Strong Field Enabled
by the Random Phase Thermal Wavefunction Approach
Bar Ezra, Shimshon Kallush, Ronnie Kosloff
Abstract
Simulating photodissociation
becomes challenging when the number
of states involved grows large. The computational complexity is usually
reduced by restrictions that limit the model’s scope. We present
a first-principles quantum model for strong-field photodissociation
that treats rotational dynamics explicitly. The random-phase thermal
wavefunction method reduces the large computational cost. The simulation
yields experimental observables, such as the momentum angular distribution
of the photofragments. We study how these observables converge at
two field intensities, with each set at a different temperature. The
method is designed to reach conditions where exact numerical methods
become intractable; here, we validate it against exact results in
the regime where both are available. Beyond the conventional dependence
on system size and number of realizations, we find that convergence
depends on the observable’s sensitivity to coherence.
[jp6c05555_0007.webp]
[jp6c05555_0005.webp]
Introduction
Time-domain spectroscopy can be directly
modeled by solving the
time-dependent Schrödinger equation. Direct modeling for a
strong-field photodissociation process has almost no alternative,
since analytical methods are not applicable and perturbation expansions
do not converge. This work presents a first-principle simulation method
applicable to the photodissociation process of a diatomic molecule.
The simulation is guided by strong-field pulsed experiments. The goal
is that the simulation outcome could be directly comparable to experimental
observables, specifically to the velocity distribution of the photofragments.
The simulation is first-principles in structure: given the Hamiltonian,
initial temperature, charge, and mass, it propagates the dynamics
with no fitted dynamical parameters.
The knowledge of the total
Hamiltonian and initial state is required
to solve the time-dependent Schrödinger equation. The Born–Oppenheimer
decomposition allows us to separate the electronic and the nuclear
terms. Ab initio electronic structure
methods are able to accurately calculate the electronic potential
surfaces and the transition dipole moments associated with the photodissociation.
The present study concentrates on solving the nuclear dynamics, specifically
the dynamics of the vibrational and rotational degrees of freedom
of a diatomic molecule on multiple electronic surfaces. The strong
coupling between the molecular degrees of freedom opens multiple channels
to unbound electronic states where the molecule dissociates.
Within the common experimental framework,
,
various
physical measures of the photofragments are observed far from the
dissociation point. The simulation should be able to calculate the
same observable and specifically the angular distribution of the outgoing
momentum. Such a calculation is set in a large phase-space volume,
which translates to a very large computational Hilbert space. In the
weak-field regime, one can use perturbation theory, and the calculation
is simplified significantly since selection rules limit the accessible
Hilbert space. In strong-field photodissociation processes, many interference
channels become open. As a result, the calculation must include all
of the relevant quantum states that participate in the dynamical process.
Solving the time-dependent Schrödinger equation for a large
Hilbert space is a difficult task. The computation effort scales with
the size of the Hilbert space, the product of the energy range, and
time period. The Hilbert space size is determined by the phase-space
volume containing the problem. The phase-space is restricted by the
maximum energy the molecule can carry and by the asymptotic distance
estimated by the effective dissociation point. The number of angular-momentum
states required for the calculation is determined by the initial temperature
of the molecule and the strength of the radiation field. Each participating
photon changes the rotational angular momentum by one unit of ħ.
Therefore, when the molecule is subjected to a strong radiation field,
the size of the manifold of the accessible states increases significantly.
For a photodissociation study, the Hilbert space is vast, of the order
106 states for the representation of the whole ro-vibronic
manifold (where, the vibration is represented on a grid of size (∼103), rotational states (∼ 102) and spin-orbital
states (∼ 101)). In principle, the initial state
is a thermal mixed state formally requiring a density-operator description.
The dimension of the density operator is the square of the Hilbert
space; this means that such a calculation is prohibitively expensive.
Several studies have addressed the computational complexity of
similar systems. In the study of I
2
–, one of the most studied
diatomic molecules, the complexity is reduced by neglecting the rotational
degree of freedom due to the heavy molecule.
−
For the H
2
– case, the lightest molecule, the study was limited to low initial
temperatures, resulting in only the ground ro-vibrational state.
−
A commonly employed approximation assumes that the initial orientation
is fixed to the perpendicular direction of the polarization of the
laser field. This leads to a significant reduction of the accessible
Hilbert space (practically to M = 0 only) with the price that the simulation scope is
limited.
The random-phase wavefunction (RPW) method is employed
in this
study to overcome the challenge of the large Hilbert space. It replaces
a density-operator calculations by a finite sample of random-phase
thermal wavefunctions.
,
Previously, the method has been
used to calculate the dynamics of molecular systems including rotational
states.
−
The present study aims to develop methods to calculate the full
asymptotic photofragment distribution. This requires studying the
convergence of the calculation with respect to the number of RPWs
employed. Typically in the RPW theory, the convergence of the observables
scales with the number of realizations, K, and the
Hilbert size, N
S, as
[jp6c05555_m001.webp]∼1K·1NS
. We investigate
how many realizations K each observable needs to
reach a target accuracy. We find this number depends strongly on the
observable’s sensitivity to coherence.
The developed
simulation is general, it can be used to calculate
a variety of processes and systems. We use the F
2
– properties
to enable later comparison with the experiments of Shahi, Albeck,
and Strasser.
,
Note, however, that the goal of
the present work is to study the random-phase method and for this
task we simplified the system and employed only two of the four potentials
(see below).
Methods
The dynamics of a gas phase
diatomic molecule takes place on n electronic surfaces
and three ro-vibrational internuclear
degrees of freedom, the internuclear distance, and the two rotational
angles. The system Hamiltonian includes the nuclear Hamiltonian of
all electronic states, and the radiation coupling elements between
the states are mediated by the transition dipole moment. In the Born–Oppenheimer
decomposition, the Hamiltonian can be
written as
1
[jp6c05555_m002.webp]Ĥsys=∑n(Ĥn⊗|n⟩⟨n|+∑k≠nμn,k·ε⊗|n⟩⟨k|)
2
[jp6c05555_m003.webp]Ĥn=P̂r22mr+Vn(r)+12IR̂2(θ,φ)
where Ĥ
n
is the
nuclear Hamiltonian operator of the nth electronic
state. Ĥ
n
contains
the electronic potential of the n electronic state, V
n
(r), at each
internuclear distance r, (presented in ) and the total nuclear kinetic
energy. The kinetic
energy is composed of the linear motion term,
[jp6c05555_m004.webp]P̂r22mr
, and the rotational motion term,
[jp6c05555_m005.webp]12IR̂2(θ,φ)
, where m
r
is the reduced
mass. Here, P
r
is the
linear momentum, I(r) = m
r
r
2 is the moment of inertia, and R̂(θ, φ)
is the nuclear rotational angular-momentum operator.
The laser field, ε, is a transform limited pulse with the parameters
presented in .
The scalar product
between the transition dipole operator, μ
n,k
, and
the radiation field, ε, can generate different
transition schemes. The specific scheme is influenced by the states
involved, symmetries, and the field polarization. The details are
elaborated in our previous work. The
present study contains only two 1Σ singlet electronic
states. Hence, under these conditions, the nuclear rotational angular-momentum
operator merges into the total angular momentum of the molecule, R̂(θ, φ) = Ĵ(θ,
φ). In the absence of other angular momenta, and the Hund’s
case (a), the most common
case and the one used here, the projection of
Ĵ
(θ, φ) on the laboratory axis, Ẑ, is m and on the internuclear axis is Ω.
Similar to our previous work, we expand
the angles’ coordinate by the Wigner rotation matrices, D
m,Ω
j
(φ, θ). With this
basis set, each state is described by r, j, m, and Ω. The density matrix of
the molecular state at time t is
3
[jp6c05555_m006.webp]ρ̂(r,r′,θ,θ′,φ,φ′,n,n′;t)=∑ζ,ζ′|n⟩⟨n′|⊗an,n′,ζ,ζ′(r,r′;t)⊗|ζ⟩⟨ζ′|
where the summation is over all rotational
states, ζ = {j, m, Ω},
thus |ζ⟩ is the wavefunction with the quantum numbers
|j, m, Ω⟩, expanded
by the D
m,Ω
j
(φ, θ) matrices. a
n,n′,ζ,ζ′ (r, r′; t) is the nuclear matrix coefficients at each electronic and rotational
state. This work includes only 1Σ states Ω
= 0, and the matrices are identical to the spherical harmonics Y
j,m
(φ,
θ). For linearly polarized external field, the coupling elements
between the states, due to the interaction with the field, are proportional
to cos(θ) ∝ D
0,0
1(φ, θ).
To mimic the
experimental process, a thermal ensemble of molecules
is subjected to a pulsed laser source. Within the temperatures considered,
the electronic and vibrational degrees of freedom are in their ground
state. The initial state is a Boltzmann distribution of several rotational
components and can be written as
4
[jp6c05555_m007.webp]ρ̂(r,r′,θ,θ′,φ,φ′,n,n′;t=0)=∑ζiPB(ji)a0,0,ζi,ζi(r,r′;t=0)·|0⟩⟨0|⊗P̂0,ji(r)⊗|ζi⟩⟨ζi|
where
[jp6c05555_m008.webp]P̂0,ji(r)=|Φn=0,ji(r)⟩⟨Φn=0,ji(r)|
is the projection onto
the ground vibrational
state that conditioned by the rotational state, j
i
.
[jp6c05555_m009.webp]PB(ji)=exp(−βEji)Z
are the Boltzmann coefficients, where
[jp6c05555_m010.webp]Z=∑l(2l+1)exp(−βEl)
is the partition function. Note, that the
rotational energy eigenvalues that satisfy Ĥ
rot |ψζ
i
⟩ = E
j
i
|ψζ
i
⟩ depend only on the total angular momentum j.
The time evolution of the state is carried out
by a first-principle
propagation of the time-dependent Schrödinger equation: ρ̂(t) = Û(t)ρ̂(t = 0)Û
†(t) where
[jp6c05555_m011.webp]iħddtÛ(t)=Ĥ(t)Û(t)
. Observables
are calculated by the data
provided by the density operator, ρ̂(t), so that for any operator Â
5
[jp6c05555_m012.webp]⟨Â⟩t=tr(Âρ(t))=tr(ÂÛ(t)ρ(t=0)Û†(t))
Under
the common experimental conditions for the molecular beam
considered here, the typical incoherent processes such as collisions
and spontaneous decay take place on the time scales of t ≥ 1 nsec. Dissociation under a 30 fsec laser pulse is achieved
by ∼100 fsec. The dynamics are therefore coherent. Therefore,
the expectation value can be evaluated, employing the basis where
the initial density operator is diagonal |ψ̃0,0,ζ
i
⟩ = |ψ0
e⟩ ⊗ |ψ0
ν⟩
⊗ |ζ
i
⟩, giving by
6
[jp6c05555_m013.webp]⟨Â⟩t=∑n,ν,ζPB(ji)⟨Û†(t)ψ̃0,0,ζi(t=0)|Â|Û(t)ψ̃0,0,ζi(t=0)⟩
Using , we can
decompose the expectation value into the expectation values of the
individual components. Each component requires the solution for time-dependent
Schrödinger equation:
[jp6c05555_m014.webp]iddtψ=Ĥ(t)ψ
. Numerical
methods for the wavefunction
propagation such as the Chebychev approximation scale as N
S
2.
−
Observables that are computed by will be used as a reference to the results that will
be derived by the random-phase method.
The size of Hilbert space
of linear molecules, that is the number
of occupied states N
S, scales linearly
with the temperature. The experiments considered on F
2
– are
carried out at an estimated temperature of T = 100 K, where N
S ∼ 106. We ran two cases at lower temperatures (see ) to keep the exact reference
computable.
The light-induced coupling between these states
further complicates
the problem. At low intensity, the transition can be approximated
perturbatively, which keeps the scale of interacting states to N
S
2, while at high intensities, strong fields, the process requires
full and exact calculation, which increases the scale of the states
to N
S
3.
To overcome the challenge of the large number of states,
the random-phase
wavefunction method is used. An efficient representation of the state
is done by the use of a limited ensemble of pure states, mimicking
the initial thermal ensemble. The initial state of a single random-phase
wavefunction is written as
7
[jp6c05555_m015.webp]|Ψk(r,θ,φ;t=0)⟩=∑ζieiχζikPB(ji)ã0,ζi(t=0)|0⟩⊗|Φ0,ζi(r)⟩⊗|ζi⟩
where the phases χζ
i
k
are chosen randomly, and ã
0,ζ
i
(t = 0) are the coefficients
corresponding to the diagonal elements
of a
0,0,ζ
i
,ζ
i
(r, r; t = 0), and the angular part is introduced
by |ζ
i
⟩. The propagation
of the wavefunction is done by solving the time-dependent Schrödinger
equation, |Ψ
k
(r, θ, φ; t)⟩ = U(t) |Ψ
k
(r, θ, φ; t = 0)⟩ (using
the Chebyshev expansion and absorption
boundary condition in additional auxiliary channels as fully described
at our previous work). For large number
of realizations k, the set of random-phase wavefunctions
is complete basis of states
8
[jp6c05555_m016.webp]limK→∞1K∑k=1K|Ψk(r,θ,φ;t)⟩⟨Ψk(r′,θ′,φ′;t)|=ρ̂(r,r′,θ,θ′,φ,φ′,n,n′;t)
The
efficiency of the random-phase wavefunction method increases
with the size of Hilbert space, i.e., with the number of random phases,
which increases with the temperature.
The time-dependent observables are computed for each random-phase
wavefunction and then averaged over K realization.
The error with respect to the exact solution, the reference, decreases
with the number of realizations, K. The observables
are calculated using the RPW as follows
9
[jp6c05555_m017.webp]⟨Â⟩tK=1K∑k=1K∑n,ν,ζ⟨Ψk(r,θ,φ;t)|Â|Ψk(r,θ,φ;t)⟩
In the following section, we present
the results calculated
by
using the random-phase wavefunction method on the spatially resolved
photodissociation. The quantum observables are then compared to numerically
exact computations, carried out by propagating
of all of the initially occupied states and thermal summation of their
expectation values. For a convergence study, the exact numerical result
is the appropriate reference: it isolates the error of the random-phase
sampling from any error in the potentials. The comparison to experiment
is a separate question, requiring the ab initio potentials discussed
above.
Results
A first-principles simulation has been designed
to simulate the
photodissociation process of a diatomic molecule. The RPW method has
been employed to overcome the large Hilbert-space challenge. Thus,
the initial state was sampled from the Boltzmann thermal state by
random-phase wavefunctions, . The photofragment outcome state is represented by the final
density operator composed from averaging the individual initial RPW
states, .
Two
opposing constraints pin the choices. The RPW method needs
many occupied states to be efficient, so low intensity forces a high
temperature. The exact reference must stay computable, so high-intensity
forces a low temperature. Our two cases therefore differ in both intensity
and temperature and do not isolate intensity by themselves. In each
calculation, we compared the RPW result to the exact computation,
propagating each initial state separately in the full Hilbert space.
First, the dissociation probability and the average kinetic energy
of the fragments were calculated. The dissociation probability,
[jp6c05555_m018.webp]Pk
, and the kinetic energy,
[jp6c05555_m019.webp]Ek
, for each realization k from the final wavefunction
are
[jp6c05555_m020.webp]Pk=⟨Ψn≠0k(r,θ,φ;t=tfinal)|Ψn≠0k(r,θ,φ;t=tfinal)⟩
10
11
[jp6c05555_m021.webp]Ek=⟨Ψn≠0k(r,θ,φ;t=tfinal)|P̂r2|Ψn≠0k(r,θ,φ;t=tfinal)⟩
The exact dissociation probability,
calculated
according to ,
is 9.28 × 10–3 at the high intensity and 6.57
× 10–6 at the low intensity. presents the deviation of
the approximate computation from the exact one, for the dissociation
probability
[jp6c05555_m022.webp]P
, and the
fragments’ average kinetic
energy E
k
.
The error bars correspond to the standard deviation
of the results
from the mean value at each given number of realizations. For any
observable A this is given by:
[jp6c05555_m023.webp]1K∑k=1K(k−K)2
, where
K
is the averaged
value of A after K realizations
and
k
is the
value for a given realization k. The
convergence of the deviation with the number of realizations is visible.
a,b displays
the error of the dissociation probability. Two scales are displayed:
the relative error
[jp6c05555_m024.webp]ΔPrel=|PRPW−P|/P
in the outer axis and absolute error
[jp6c05555_m025.webp]ΔPabs=|PRPW−P|
in the inner axis. As expected,
at the
high intensity the large overall
[jp6c05555_m026.webp]P
leads
to an absolute error much larger
compared to the low intensity. The relative error of this case is,
however, significantly lower than at the low intensity. Note that
a comparison to the experimental output reveals that, at least for
commonly used methods, the resolution for the low intensity regime
is way beyond the experimental. Absolute experimental values for the
dissociation probability for the higher intensity regime are even
harder to be determined.
The kinetic energy of the photofragments
is used experimentally
to track the asymptote of the electronic state from which the dissociation
takes place. Panels (c) and (d) in present the average kinetic energy relative
error. The error at both intensities was already lower than 1% after
200 realizations.
presents
the angular distribution of the momentum fragments calculated by the
direct method, a,b, and the RPW method, c,d. The angular distribution
of the outgoing momentum, P, is calculated as follows
[jp6c05555_m027.webp]Dk(P,θ)=∫dφ∑m,Ω,n>0|∑j∑ζieiχζikPB(ji)ãn,0,ζ,ζi(t=tfinal)·Dm,Ωj(θ,φ)|2
12
where ã
n,0,ζ,ζ
i
(t = t
final) are
the coefficients in momentum representation corresponding to the a
n,0,ζ,ζ
i
(t = t
final) coefficients. Note that the summation over j is
a quantum summation that conserves the phases while the summations
over m and Ω are classical. The difference
is an outcome of the dynamics, which generates coherences only between
different values of j with the same values of m and Ω.
The two distributions are qualitatively similar
with the same central
momentum peaks and parallel orientation. However, changes in small
details could be observed, especially when comparing the distributions
at high-intensity panels, d,b.
A more detailed
analysis is obtained by comparing separately the
directions in the azimuth and the absolute value of the momentum.
The upper panels in presents the distribution as a function of the momentum for P
⊥ = 0. Note, that only a very narrow
range of values from the full momentum grid are plotted. The lower
panel at presents
the distribution as a function of the angle θ, taken at the
most probable momentum. At low intensity, the convergence of both
distributions to the exact distribution is fast, while at high laser
intensity, the distribution has not converged even after 2000 realizations.
In particular, the momentum-magnitude peak at high intensity reaches
only about a third of the exact height and barely changes between
100 and 2000 realizations. Strong interference at high intensity builds
high coherence into this observable, which drives the slow convergence.
Specifically, we see the sensitivity to interference on the momentum
and angular distribution. On top of this, one should beware of the
fact that the grid resolution at the P axis is poor,
due to the very narrow width of the distribution in the momentum space.
The distance between the distributions
as a function of the number
of realizations is calculated by the Wootters metric. The distance is calculated for the momentum distribution, d
W
p, and at the angular distribution, d
W
θ, defined
by the following equations
13
[jp6c05555_m028.webp]DWp(K)=π2cos−1(∫P2dPDrandK(P,θ=0)·Dexact(P,θ=0))
14
[jp6c05555_m029.webp]DWθ(K)=π2cos−1(∫sin(θ)dθDrandK(P=Pmp,θ)·Dexact(P=Pmp,θ))
We
choose to calculate the distribution at the momentum direction, d
W
p, and the angular direction, d
W
θ, separately. For each calculation,
we set one of the direction at the most probable value, θ =
0 and p = p
mp. The distance
was normalized so that its maximal value is 1 for nonoverlapping distributions.
The distances between the distributions as a function of the number
of realizations are present in .
The convergence of the distribution is analyzed
by the distance
between the distributions at different laser intensities. An inspection
of the distances in , reveals that the distribution at low intensity, both in
angular and momentum, is closer to the exact distributions. Furthermore,
in both intensities, the momentum distribution is closer to the exact
solution than the angular distribution. For an extremely narrow distribution
such as that in our case, the distribution distance is sensitive to
small differences and accounts for high levels of coherences. In contrast,
observables that are more classical and contain a lower level of off-diagonal
elements in the energy representation, such as the average kinetic
energy and dissociation probability, are less sensitive to coherences
and are much easier to capture with the random-phase method.
Discussion
and Conclusions
In this work, a computational method for
the first-principle simulation
of strong-field photodissociation and other time domain spectroscopy
was developed. Simulations based on an exact solution of the time-dependent
Schrödinger equation are expensive. Typical computational complexity
scales as N
S
3, where N
S is the
number of states involved in the process. Simulations of diatomic
photodissociation processes at typical experimental conditions require
a vast Hilbert space, of the order 106.
To overcome
this challenge, the RPW method has been employed. The
phases of the initial states are randomly chosen and thermal averaged
observables are obtained. The observables are converged to the directly
calculated outcome after K realizations, where K ≪ N
S. Previous studies
on RPW
,,
showed that
the observable convergence scales as
[jp6c05555_m030.webp]∼1K·1NS
. Here, we show the required K is not universal: it depends strongly on the observable’s
sensitivity to coherence.
As the RPW method relies on randomness
of the relative phase between
the states, it is unlikely that processes which are driven by high-order
coherent interactions, e.g., multiphoton transitions, would be well
approximated by the method. These types of processes are highly coherent
and employ throughout the dynamics nondiagonal elements of the density
matrix, i.e., coherences.
The dissociation probability
[jp6c05555_m031.webp]P
is an example of an observable
in which
the only contribution comes from diagonal elements of the density
matrix in the energy representation. Hence, coherent interferences
have low impact on such observables. Another case is the average kinetic
energy, ΔE
k
rel, an observable that is less
affected by the different interference pathways in the dynamics. At
both intensities, it converges to the exact solution up to the experimental
error of about 0.02 eV after only ∼ 200 realizations, which
means acceleration of the calculation by a factor of ∼20 compared
to the direct method, which propagates every initially occupied state
separately. An explanation can be attributed to the property of the
operators that correspond to these observables that are diagonal in
the energy representation. Thus, they are less dependent on coherences
and interferences.
Another observable that was explored is the
momentum-space distribution
of the dissociated fragments, which provides a visual approach to
study the convergence. Our main result concerning this quantity is
that its dominant features, such as the overall shape and peak position,
are accurately recovered by the method, while fine details such as
the peak height might be difficult to retrieveat high intensity,
the momentum-magnitude peak reaches only about a third of the exact
value, and barely improves between 100 and 2000 realizations. This
is the most coherence-sensitive observable we studied, and its slow
convergence is the clearest illustration of our central finding. Unlike
the dissociation probability and kinetic energy, this distribution
carries substantial coherence between angular-momentum states, so
it converges far more slowly. In addition, the convergence is better
for the angular variable θ than for the magnitude of the momentum,
due to the very different width of the distribution in these two directions.
Within a single calculation, every observable shares the same N
S and the same K, yet at high
intensity, the diagonal observables converge within 200 realizations
while the momentum distribution does not converge after 2000. This
within-run difference cannot be attributed to temperature, and therefore
reflects coherence alone. Consistently, the split is pronounced in
the high-intensity run, where coherence is largest, and negligible
at low intensity, where both observable types converge fast. We remark
that the common use of this tool in the experimental context is usually
more quantitativeits main use is to assign the different fragments
to their electronic channels. Therefore, our finding that the exact
distribution is difficult to recover affects the applicability of
the method only marginally.
The observed error leads us to conclude
that the different observables
converge differently. Each observable is sensitive to different properties
of the photofragments state. The dynamics of the dissociation process
include interferences. Adding random phases changes the interference
pattern, leading to different outcomes. At high intensity, the process
contains multiphoton transitions that induce rare events and interferences
that lead to special states. Thus, adding a random phase leads to
different events and interferences and to different final states.
The observable convergence is subject to the coherence and interference
during the process. We note here that the choice of the wavefunction
basis might be another factor that affects the observable convergence.
However, upon changing the basis from the Wigner D matrices to a direct angular grid representation basis, we did not
observe any significant shift in the convergence rate.