Algorithms
Ground state, excited states, periodic systems — every algorithm here inherits from the same VariationalDriver and reaches the user through the same Mandacaru. The basis and the integrals that feed them are on the Basis Sets page.
VQE optimizes a fixed UCCSD ansatz; ADAPT-VQE grows the ansatz operator by operator, screening the gradient of a pool — fermionic, qubit, qubit-excitation-based (QEB) or compact-excitation (CEO) — and adding only the one that lowers the energy most.
In the program
Mandacaru(method="adapt-vqe", pool="fermionic") · method="vqe" for the fixed UCCSD ansatz
DeflationMixin runs variational deflation (VQD): each new state is optimized orthogonal to the ones already found. SubspaceMixin runs SSVQE instead, optimizing a weighted set of orthogonal states together. Both plug into any driver without duplicating its setup.
In the program
atoms.calc.energy_levels(num_states=3) · Mandacaru(method="subspace-adapt-vqe", num_states=3)
A converged variational state can be handed to QPE instead of being re-optimized: QuantumPhaseEstimation reads a wavefunction checkpoint, evolves it under exp(2πi(H−Elo)/W) and inverse-QFTs the evaluation register, returning the eigenvalue, the peak probabilities and the collapsed eigenstate. The 2ⁿ⁺ᵗ state vector is costed against the machine's free memory before a byte is allocated.
In the program
QuantumPhaseEstimation(n_evaluation_qubits=10).run("state.json") · memory_policy="abort" | "prompt" | "ignore"
The same drivers run on a periodic basis through BlochCalculator, producing Bloch band structure and Born–von Kármán total energies for a crystal, not just an isolated molecule.
In the program
BlochCalculator(atoms, method="adapt-vqe").band_structure(path) · .total_energy(kmesh=(4, 1, 1))
Cutting the qubit count, tracking how much of Hilbert space an ansatz actually reaches, and getting a result that survives a restart.
Drop the chemically inert core orbitals from the qubit count before mapping, without touching the basis or the integral engine.
An even electron count starts from the closed-shell RHF orbitals; an odd count — any radical or ion — is built in the natural-orbital basis of the unrestricted (UHF) solution, one spatial basis shared by both spins, so the same solvers, pools and mappings run unchanged. The spin state is read from the geometry's magnetic moments: a doublet by default, a triplet O₂ or a quartet with magmoms set. Works for every basis family, plane waves included.
From ten qubits up, ADAPT-VQE keeps the pool as sparse matrices rather than materializing every 2ⁿ×2ⁿ generator, screening with the exact analytic gradient and densifying only the operators it selects. From sixteen, the Hamiltonian and the pool are restricted to the (n_α, n_β) particle-number sector as well — LiH in PAW-DZP is 20 qubits but only 100 states.
KL-divergence against the Haar-random distribution measures how much of Hilbert space an ansatz can actually reach, tracked as ADAPT-VQE grows it.
Dynamic parametrization: by default every parameter is re-optimized each ADAPT iteration; quenching=False freezes the earlier ones and optimizes only the newest, a cheaper one-dimensional line search per growth step.
atoms.get_forces() differentiates the energy the solver actually computed, contracting the one- and two-body RDMs against displaced integrals on the same frozen grid — Hellmann–Feynman where the operators move, Pulay where the basis functions do. It is the default on every basis, all-electron or pseudopotential; the neglected orbital-response term is measured rather than assumed, and reported when it matters.
checkpoint="state.json" writes the reference, the generators, the angles and the Hamiltonian after every accepted operator; resume="state.json" continues an interrupted or unconverged run where it stopped, matching stored generators back to the pool by content. The same file is what QPE reads, and what warm-starts the next geometry of a relaxation.
mandacaru geometry.xyz --dry-run reports the qubit register a calculation needs — after the frozen core or plane-wave cutoff — and whether it fits the target device, without computing an integral or executing a circuit.
A built qubit Hamiltonian serializes to Parquet or JSON with automatic format detection — replay a solver against it without recomputing integrals or geometry.
One Mandacaru surface — the choice of backend, basis and mapping composes freely with the choice of algorithm above.
Every solver reaches the user through one entry point — Mandacaru(method=...) — an ASE calculator, so a run looks the same whether it optimizes a UCCSD ansatz on a state vector or screens ADAPT-VQE gradients against a real QPU.
Seven families of localized orbitals generated from scratch — FAO to def2-QZVPPD — the frozen core, and the C-accelerated integral engine underneath.
The same algorithm compiled and executed on Qiskit, Amazon Braket and Cirq — one keyword, one unitary, machine-precision agreement.
Cross-backend validation to 3.5×10⁻⁶ eV, real QPU execution through Amazon Braket, and what still doesn't work.