Providers
A variational algorithm is built once, as a product of exponentials of anti-Hermitian generators. Which SDK turns that into a circuit, and which machine executes it, are two keyword arguments — and because the decomposition is exact, the three SDKs return the same state vector to machine precision.
backend_provider="qiskit"
IBM's SDK, the default, and the route to IBM Quantum. Qiskit is little-endian, so Mandacaru qubit k is laid on wire n−1−k; the unitary is unaffected.
backend_provider="braket"
The route to IonQ, IQM and Rigetti. Locally it returns the exact state vector; through AWS the energy is measured from shots over qubit-wise-commuting Pauli groups.
backend_provider="cirq"
Google's SDK. Big-endian like Mandacaru itself, so the circuit reads exactly as the generators are written.
from ase import Atoms
from mandacaru import Mandacaru
atoms = Atoms("LiH", positions=[[0, 0, 0], [0, 0, 1.595]], cell=[15.0] * 3)
atoms.center()
# Build the qubit Hamiltonian once and cache it. Every provider below
# replays it, so any spread between them is the SDK, not the integrals.
atoms.calc = Mandacaru(method="adapt-vqe",
basis="FAO",
pool="qeb",
h=0.25,
save_hamiltonian="lih.parquet")
atoms.get_potential_energy()
for sdk in ("qiskit", "braket", "cirq"):
calc = Mandacaru(method="adapt-vqe",
pool="qeb",
load_hamiltonian="lih.parquet",
backend_provider=sdk, # which SDK builds the circuits
execute_circuits=True, # ... and runs them
trace=False)
result = calc.run() # no geometry needed
print(f"{sdk:7s} {result.optimal_energy:.6f} eV")
| Provider | Energy (eV) | Error vs. FCI | CNOTs | Depth | Time |
|---|---|---|---|---|---|
| internal (matrix) | −187.438634 | 3.65×10⁻⁶ | 208 | 273 | 0.3 s |
| Qiskit | −187.438636 | 1.71×10⁻⁶ | 208 | 273 | 4.1 s |
| Amazon Braket | −187.438635 | 2.97×10⁻⁶ | 208 | 359 | 43.9 s |
| Cirq | −187.438636 | 2.22×10⁻⁶ | 208 | 358 | 8.9 s |
All four grow the same eight-operator ansatz and recover the exact ground state of the cached Hamiltonian; the depth differs only because each SDK schedules one gate stream its own way. Braket's local simulator is the slow one — that is the SDK, not the algorithm.
Ten devices are registered by name, and any Braket ARN, ibm_* or fake_* backend is accepted verbatim. Every QPU needs shots — with shots > 0 the energy is measured from bit strings over qubit-wise-commuting Pauli groups instead of read off amplitudes.
| Kind | device= | Notes |
|---|---|---|
| Simulators | AER_simulator · braket-local · braket-sv1 · braket-dm1 · braket-tn1 | Exact state vector, or shot-based with shots > 0 |
| IBM Quantum | ibm-quantum · ibm_<name> · fake_<name> | Qiskit Runtime; ibm-quantum picks the least busy operational QPU, fake_* rehearses the path locally |
| Braket QPUs | braket-ionq-aria · braket-ionq-forte · braket-iqm-garnet · braket-rigetti-ankaa | IonQ, IQM and Rigetti through AWS; any device ARN also works |
from ase import Atoms
from mandacaru import Mandacaru
from mandacaru.backends.providers import QiskitProvider
atoms = Atoms("H2", positions=[[0, 0, 0], [0, 0, 0.74]], cell=[8.0] * 3)
# physical_qubits pins the layout: rank pairs by readout and CZ error
# first -- a qubit whose readout drifted since calibration ruins the
# energy while its recorded error still looks fine.
atoms.calc = Mandacaru(method="adapt-vqe",
basis="FAO",
pool="fermionic",
mapping="parity_reduced", # H2: 4 qubits -> 2
measurement_provider=QiskitProvider(
device="ibm_fez",
shots=4096,
physical_qubits=[22, 23],
estimator_options={"resilience_level": 2}))
energy = atoms.get_potential_energy() # measured on the QPU, in eV
Amazon Braket is the same shape: device="braket-ionq-aria", shots=4096. A dry run may name a QPU without shots, so a register can be costed against a device before an account exists.
There is no mitigation module: the Qiskit Runtime Estimator's resilience levels — readout mitigation, twirling, zero-noise extrapolation — are used as they come. Writing one is a roadmap item, not a shipped feature.
Every variational energy can be measured natively, but ADAPT-VQE's operator-gradient screening still runs on the classical state vector behind it. Fixed-ansatz VQE is the fully hardware-native method today.
Translating a molecule rigidly across a frozen grid changes its energy — the egg-box effect, tens of meV at a coarse spacing and falling with it. The analytic force differentiates that faithfully rather than hiding it, so converge the spacing before reading a small force as chemistry.
Twelve families of orbitals generated at run time, the frozen core, and the integral engine underneath.
ADAPT-VQE, VQE, excited states, phase estimation and periodic systems — one driver under all of them.