Optimization for a Kerr oscillator¤
In this example we will try to implement DRAG-like pulses on a Kerr oscillator, penalizing population going to leakage states. We will additionally use a slightly more complicated control scheme than in the qubit example. There we performed traditional grape, with piecewise-constant controls. Here we fit a spline to control knot points: it is these knot points that are optimized.
This technique has the benefit of ensuring smooth controls if the control points are spread far enough apart in time, which is not guaranteed in the case of piece-wise constant controls. It is important to emphasize here that the control points are in no way related to the points chosen by the numerical integrator for solving the differential equation! This is a significant difference from typical QOC/GRAPE implementations that perform numerical integration of the Schrödinger equation by calculating step propagators with sufficiently small time steps.
This example is available as a Jupyter notebook here.
import diffrax as dx
import dynamiqs as dq
import jax.numpy as jnp
import jax.tree_util as jtu
import matplotlib.pyplot as plt
import numpy as np
import optax
from jax import Array
import qontrol as ql
time = 20.0
control_dt = 2.0
ramp_nts = 4
ntimes = int(time // control_dt) + 1
tsave = jnp.linspace(0, time, ntimes)
optimizer = optax.adam(learning_rate=0.001, b1=0.999, b2=0.999)
options = ql.OptimizerOptions(
verbose=False,
save_states=True,
progress_meter=None,
epochs=4000,
plot=True,
plot_period=20,
)
We define our Kerr oscillator to have a 100 MHz Kerr. We optimize the drive envelopes on the I and Q quadratures.
dim = 5
Kerr = -2.0 * jnp.pi * 0.100
a = dq.destroy(dim)
H0 = -0.5 * Kerr * dq.dag(a) @ dq.dag(a) @ a @ a
H1s = [a + dq.dag(a), 1j * (a - dq.dag(a))]
H1_labels = ['I', 'Q']
Here we attempt to perform an X
gate, and penalize population leaking to the higher-lying states
initial_states = [dq.basis(dim, 0), dq.basis(dim, 1)]
final_states = [dq.basis(dim, 1), dq.basis(dim, 0)]
_forbidden_states = [dq.basis(dim, idx) for idx in range(2, dim)]
forbidden_states_list = len(initial_states) * [_forbidden_states]
We initialize the guess for the controls and moreover define an envelope of cosine ramps with a flat top to ensure that the control turns on slowly.
init_drive_params = {'dp': -0.001 * jnp.ones((len(H1s), ntimes))}
cos_ramp = (1 - jnp.cos(jnp.linspace(0.0, jnp.pi, ramp_nts))) / 2
envelope = jnp.concatenate(
(cos_ramp, jnp.ones(ntimes - 2 * ramp_nts), jnp.flip(cos_ramp))
)
def _drive_spline(
drive_params: Array, envelope: Array, ts: Array
) -> dx.CubicInterpolation:
drive_w_envelope = jnp.einsum('t,...t->t...', envelope, drive_params)
drive_coeffs = dx.backward_hermite_coefficients(ts, drive_w_envelope)
return dx.CubicInterpolation(ts, drive_coeffs)
def H_func(drive_params_dict: dict) -> dq.TimeArray:
drive_params = drive_params_dict['dp']
H = H0
for H1, drive_param in zip(H1s, drive_params):
drive_spline = _drive_spline(drive_param, envelope, tsave)
H += dq.modulated(drive_spline.evaluate, H1)
return H
exp_ops = [dq.basis(dim, idx) @ dq.dag(dq.basis(dim, idx)) for idx in range(dim)]
se_Kerr_model = ql.sesolve_model(H_func, initial_states, tsave, exp_ops=exp_ops)
costs = ql.coherent_infidelity(target_states=final_states, target_cost=0.005)
costs += 40.0 * ql.forbidden_states(
forbidden_states_list=forbidden_states_list, target_cost=0.4
)
opt_params = ql.optimize(
init_drive_params, costs, model=se_Kerr_model, optimizer=optimizer, options=options
)
Try playing with the cost_multiplier
of the forbidden states cost function to see how it modifies the pulse! Specifically, try setting it to 0 so that the optimizer is happy to populate the 2nd excited state without penalty.
Time-optimal control¤
We can modify the above example to perform time-optimal control, allowing the optimizer to make the pulse longer or shorter. Notice the additional parameter in init_drive_params
that controls the pulse time. We now provide an additional update function that updates tsave
during the optimization.
init_drive_params_topt = {'dp': -0.001 * jnp.ones((len(H1s), ntimes)), 't': tsave[-1]}
def H_func_topt(t: float, drive_params_dict: dict) -> dq.TimeArray:
drive_params = drive_params_dict['dp']
new_tsave = jnp.linspace(0.0, drive_params_dict['t'], len(tsave))
drive_spline = _drive_spline(drive_params, envelope, new_tsave)
drive_amps = drive_spline.evaluate(t)
drive_Hs = jnp.einsum('d,dij->ij', drive_amps, H1s)
return H0 + drive_Hs
def update_H_topt(drive_params_dict: dict) -> dq.TimeArray:
new_H = jtu.Partial(H_func_topt, drive_params_dict=drive_params_dict)
return dq.timecallable(new_H)
def update_tsave_topt(drive_params_dict: dict) -> Array:
return jnp.linspace(0.0, drive_params_dict['t'], len(tsave))
se_t_opt_Kerr_model = ql.sesolve_model(
update_H_topt, initial_states, update_tsave_topt, exp_ops=exp_ops
)
opt_params_topt = ql.optimize(
init_drive_params_topt,
costs,
se_t_opt_Kerr_model,
optimizer=optimizer,
options=options,
)
The pulse and fft don't get plotted in this case because the Hamiltonian is of type dq.timecallable
, which does not technically have a prefactor
attribute. We plot the pulse below.
# helper plotting function for the multiple variations
def plot_pulse(_opt_params: dict, _model: ql.Model) -> None:
result, _H = _model(_opt_params)
_tsave = _model.tsave_function(_opt_params)
drive_spline = _drive_spline(_opt_params['dp'], envelope, _tsave)
finer_times = jnp.linspace(0.0, _tsave[-1], 201)
drive_amps = jnp.asarray([drive_spline.evaluate(t) for t in finer_times]).swapaxes(
0, 1
)
fig, ax = plt.subplots(figsize=(5, 4))
fig.patch.set_alpha(0.1)
ax.set_facecolor('none')
for drive_idx in range(len(H1s)):
ax.plot(
finer_times,
drive_amps[drive_idx] / (2.0 * np.pi),
label=H1_labels[drive_idx],
)
ax.set_xlabel('time [ns]')
ax.set_ylabel('pulse amplitude [GHz]')
ax.legend(framealpha=0.0)
plt.tight_layout()
plt.show()
plot_pulse(opt_params_topt, se_t_opt_Kerr_model)
We can see that the optimizer chooses to make the pulse slightly longer to reduce the amount of leakage.
Master-equation optimization¤
We can also instead perform the optimization directly on the master equation. This example is interesting because photon loss encourages the optimizer to make the pulse shorter, which competes with leakage effects. Continuing the above example, we need now to initialize an mesolve_model
, which calls dq.mesolve
, and to correctly specify the initial and target states as density matrices.
jump_ops = [0.01 * a]
me_initial_states = dq.todm(initial_states)
me_final_states = dq.todm(final_states)
me_forbidden_states_list = len(initial_states) * [dq.todm(_forbidden_states)]
me_costs = ql.coherent_infidelity(target_states=me_final_states, target_cost=0.005)
me_costs += 40.0 * ql.forbidden_states(
forbidden_states_list=me_forbidden_states_list, target_cost=0.4
)
me_Kerr_model = ql.mesolve_model(
update_H_topt, jump_ops, me_initial_states, update_tsave_topt, exp_ops=exp_ops
)
opt_params_me = ql.optimize(
init_drive_params_topt,
me_costs,
me_Kerr_model,
optimizer=optimizer,
options=options,
)
plot_pulse(opt_params_me, me_Kerr_model)