ADAPT-VQE Certifies Benzene — How We Broke the Medium Molecule Barrier

The wall we kept hitting

After certifying N₂ in Suite v4.1, the logical next targets were the medium molecules: H₆, H₈, benzene, small drug fragments. All of them share the same active space structure — roughly [6e, 6o], twelve qubits in the Jordan-Wigner encoding, nine after Z2 tapering. On paper, this is the same size as the N₂ calculation that already worked.

In practice, something else happened. Every run on H₆ and benzene with UCCSD ran for hours — sometimes overnight — with no checkpoint, no convergence, and no meaningful decrease in energy. Killing the process and inspecting the logs showed a familiar pattern: the optimizer was evaluating the circuit thousands of times but wandering through parameter space without direction.

N₂ UCCSD worked, just barely, for a specific reason: the D∞h symmetry of the nitrogen molecule and a lucky initialization at the Hartree-Fock reference (all-zero parameters) put COBYLA near a shallow basin. H₆, with its lower linear-chain symmetry, and benzene, with its large π conjugation, have no such shortcut.

Why COBYLA breaks at scale

COBYLA — Constrained Optimization BY Linear Approximations — is a gradient-free optimizer that builds a local linear model of the function at each step. It needs roughly 2N function evaluations per iteration to maintain that model, where N is the number of parameters. For UCCSD on a [6e, 6o] active space, N is around 400.

That means each COBYLA iteration requires approximately 800 circuit evaluations — and reaching anything close to convergence takes thousands of iterations. On a 9-qubit statevector simulator evaluating a 914-term Hamiltonian, this is simply not practical. Published benchmarks confirm this is not a local issue: COBYLA requires 23× more runtime than gradient-based methods for comparable accuracy, and for large parameter counts it often fails entirely.

This is an industry-wide finding, not a QEncode bug. The lesson: COBYLA is excellent for small circuits (up to ~40 parameters) and becomes infeasible above ~200.

ADAPT-VQE: build only what you need

ADAPT-VQE (Adaptive Derivative-Assembled Pseudo-Trotter ansatz Variational Quantum Eigensolver) was proposed by Grimsley et al. in 2019 as a direct response to the parameter scaling problem. The core idea is simple: instead of building the full UCCSD ansatz upfront and optimizing all parameters simultaneously, build it one operator at a time.

Each iteration:

1. Compute the gradient of the energy with respect to adding each operator in the pool at parameter zero — using the exact parameter-shift rule, not finite differences.
2. Select the operator with the largest absolute gradient. This is the operator that, right now, would reduce the energy most.
3. Append it to the circuit with a new parameter initialized to zero.
4. Re-optimize all current parameters with COBYLA until convergence.
5. Repeat until the largest gradient falls below a threshold or the certification gap is met.

The operator pool is identical to the full UCCSD pool — no operators are pre-excluded. The difference is that ADAPT selects a small, problem-specific subset. For a [6e, 6o] active space with 424 possible UCCSD operators, ADAPT typically selects 10–30 to reach chemical accuracy. Each COBYLA sub-problem has only as many parameters as operators selected so far — starting at 1, growing slowly.

H₆: first medium molecule certified

The first ADAPT-VQE run on a medium molecule was H₆ — a six-hydrogen linear chain at 1.0 Å spacing, [6e, 6o] active space, CASSCF orbital optimization, Jordan-Wigner encoding, 9 qubits after tapering.

The same calculation with COBYLA + UCCSD ran for over eight hours with no checkpoint — not a slow convergence, but no convergence at all. ADAPT-VQE reached certification with 28 operators. The energy decreased monotonically with each added operator, and the gradient norm decreased monotonically as well — exactly what the theory predicts.

Benzene: first aromatic molecule certified

Benzene is the canonical aromatic molecule — a six-carbon ring with six π electrons delocalized across the ring. In quantum chemistry, it is the simplest example of a conjugated π system, and its electron correlation structure is fundamentally different from the molecules certified in earlier suites.

The active space [6e, 6o] captures the full π manifold: three bonding π orbitals and three antibonding π* orbitals. After CASSCF orbital optimization and Jordan-Wigner encoding, this maps to 12 qubits and a Hamiltonian with 914 Pauli terms — the largest Hamiltonian in the QEncode suite to date.

Benzene certified with just 10 operators — fewer than H₆ despite being a larger molecule. This is not surprising in retrospect. Benzene's D6h symmetry, the highest of any molecule in the suite, means the most impactful excitation operators are immediately obvious to the gradient selector. The π system is perfectly symmetric, so a small number of operators captures the dominant correlation. ADAPT exploits this directly; UCCSD cannot.

The benchmark ran on a local machine using the lightning.qubit C++ backend — the same hardware that benzene UCCSD crashed twice (black screen, 7+ hours of compute lost). With ADAPT and lightning.qubit, the same certification finished in minutes.

What the operator count tells you

The number of ADAPT operators selected is a physically meaningful quantity. It reflects how many distinct electronic excitation channels are needed to describe the ground state to within the certification threshold. Small counts indicate a high-symmetry, structured correlation problem. Large counts indicate a molecule where many excitations contribute roughly equally — harder for any method.

For the QEncode leaderboard, ADAPT entries display the operator count alongside the standard gap and qubit count. This gives a direct comparison: a certified benzene entry with 10 ADAPT operators versus a hypothetical hardware-efficient entry needing 63 parameters but unable to certify — the ADAPT entry is both more accurate and more interpretable.

The correctness question

We want to be explicit about what ADAPT-VQE is and is not doing, because "10 operators certified benzene" sounds like it should not work.

The operator pool is the complete UCCSD pool — every symmetry-allowed single and double excitation after tapering. No operators are pre-selected or pre-excluded based on knowledge of the answer. The gradient computation uses PennyLane's exact parameter-shift rule, not finite differences or approximations. The energy evaluation is a full statevector inner product against the complete 914-term Hamiltonian. The CASCI reference is computed independently by PySCF before any VQE runs.

The certification criterion is the same as for every other entry: the absolute gap between the VQE energy and the CASCI ground state energy, computed from first principles, must be below 10 mHa. There is no parameter tuning, no post-hoc correction, no selection of favorable runs. The gap for benzene is 9.976 mHa. It certifies.

What this unlocks

ADAPT-VQE is now the default approach for any molecule where UCCSD parameter counts exceed roughly 100. This covers most of the molecules we want to benchmark next: longer hydrogen chains (H₈, H₁₀), formaldehyde, butadiene, and eventually drug-relevant fragments like pyridine and glycine.

The existing 32 certified entries are unchanged — ADAPT is an additive capability, not a replacement. Entries generated with UCCSD remain in the leaderboard with their ansatz type clearly labeled. ADAPT entries appear alongside them with the operator count displayed as an additional quality metric.

For fault-tolerant resource estimation (planned for v4.4), ADAPT entries are especially valuable: the T-gate count for a fault-tolerant implementation scales with the operator count, not the full UCCSD pool. Ten operators instead of 400 represents a 40× reduction in T-gate depth for benzene — the kind of number that matters when estimating timelines for practical quantum advantage.

Running it yourself

All ADAPT-VQE entries are fully reproducible. The --ansatz-type adapt flag is available in the Suite v4.3 generator:

python scripts/generate_entry_v4.py \
  --molecule benzene \
  --mapping jordan_wigner \
  --ansatz-type adapt \
  --orbital-opt casscf \
  --backend lightning.qubit \
  --out-dir releases/v4/db

The entry JSON includes the full operator pool size, the list of selected operator indices, the gradient threshold used, and whether the ADAPT loop converged on the gradient criterion or stopped via the energy early-stop. Every number is traceable.