Hardware
Every figure on this page is the project's own. The caveats below are kept as visible as the headline numbers on purpose: a result that hides its own limits is not more trustworthy for it.
ADAPT-VQE on LiH — 8 fermionic excitation operators, 208 CNOTs — built from one shared gate stream (X, H, S, S†, CNOT, Rz) and translated per SDK. No exceptions, no cherry-picking: every backend against the same FCI reference.
| Backend | Ground-state energy | Error |
|---|---|---|
| NumPy (state vector) | −187.438634 eV | 3.65×10⁻⁶ eV vs. FCI |
| Qiskit | −187.438636 eV | 1.71×10⁻⁶ eV vs. FCI |
| Amazon Braket | −187.438635 eV | 2.97×10⁻⁶ eV vs. FCI |
| Cirq | −187.438636 eV | 2.22×10⁻⁶ eV vs. FCI |
"NumPy (state vector)" is the internal reference implementation, not an SDK — the row this table checks the other three against.
Read these before the table above, not after.
On IBM hardware the Qiskit Runtime Estimator's resilience levels (readout mitigation, twirling, zero-noise extrapolation) are used as-is; Mandacaru has no mitigation module of its own — it is a roadmap target, not yet written. Pin good physical qubits with physical_qubits=[...]: a qubit whose readout has drifted since calibration ruins the energy while its recorded error still looks fine.
A molecule translated rigidly across a frozen real-space grid changes energy — the egg-box effect, 18 meV on H₂ at h = 0.25 Å and falling with the spacing. The analytic force differentiates that discretized energy faithfully rather than hiding it, so it is real and it is in the numbers. Converge h before reading a small force as chemistry.
H₂ and LiH relax cleanly. Water does not — not because the gradient is wrong, but because its energy surface in this model runs monotonically downhill as the O–H bonds stretch: −461.20 eV at 0.97 Å against −472.37 eV at 2.40 Å. The optimizer is walking correctly down a surface with no bound minimum. Check the scan before trusting a relaxation on a new system.
On a real QPU every variational energy is measured natively, but ADAPT-VQE's operator-gradient screening step still executes on the classical state vector behind it. Fixed-ansatz VQE is the fully hardware-native method today.
The caveats above are about what is missing. What exists is checked against results that have nothing to do with a single lucky run:
Real QPU access goes two ways. IBM Quantum through Qiskit Runtime: save your account once, optimize on the local state vector, then measure the optimized states in one Estimator job with QiskitProvider(device="ibm_kingston,ibm_fez,ibm_marrakesh", shots=4096) — device="fake_kingston" rehearses the same path on your laptop. Amazon Braket: an AWS account, IAM permissions for Braket, and device=<Braket device or ARN>, shots=<n>. The install guide below covers the Python side; the how-to guide under docs/guide/aws_braket covers both services.
A real error-mitigation backend, hardware-native gradient screening for ADAPT-VQE — today the variational energies reach the QPU but the operator screening stays classical — and a GPU-accelerated integral backend alongside the existing C/OpenMP one.
Seven families of localized orbitals generated from scratch — FAO to def2-QZVPPD — the frozen core, and the C-accelerated integral engine underneath.
ADAPT-VQE, VQE, excited states and periodic systems — every solver sharing one driver, and a dry run before any of them.
The same algorithm compiled and executed on Qiskit, Amazon Braket and Cirq — one keyword, one unitary, machine-precision agreement.