Nadarasa · v0.4.1 · External hook
ADAPT-GQE: generative circuit synthesis for molecular ground states
Koziell-Pipe, Brewer, Guhit, Farag, Panchagnula, Laude, Finger, Gaggioli, Szulakowska, Backhouse, Papalitsas, Mustakis, Soini, Muñoz Ramo, Clark, Kyoseva & Rinaldi (Quantinuum, NVIDIA, Pfizer; July 2026) train transformers on ADAPT-VQE reference circuits and refine the result with reinforcement learning. The generated circuits prepare ground states of drug-scale molecules and were executed on Quantinuum Helios-1.
arXiv:2607.22468 · Learning to Prepare Molecular Ground States with Transformer Models
Keywords: Quantum Computing, Generative Quantum AI, Quantum Chemistry. Demonstrated on imipramine and executed on Helios-1.
Five touch-points
1.ADAPT-VQE as curriculum
Koziell-Pipe et al. §2ADAPT-VQE is expensive because it grows the circuit one operator at a time, re-optimising all parameters at every step. ADAPT-GQE amortises that cost by using the final ADAPT-VQE circuits as training targets for a transformer. The model learns to mimic the structured, compact circuits that ADAPT-VQE produces, without running the adaptive loop at inference time.
2.RL past the training data
Koziell-Pipe et al. §3A transformer alone is limited to the fidelity of its training data. Reinforcement learning then refines the generated circuits, driving state-preparation accuracy beyond the ADAPT-VQE reference. This is a generative-model pattern we do not yet have in the PQP rewriter: data-driven search can exceed the symbolic starting point.
3.Order-of-magnitude speedup
Koziell-Pipe et al. §4Circuit-generation time drops by an order of magnitude relative to ADAPT-VQE, while accuracy is comparable or better. The bottleneck moves from variational optimisation to model inference, which is a far cheaper classical operation.
4.Helios-1 execution
Koziell-Pipe et al. §5The generated circuits were executed on Quantinuum Helios-1, making this one of the first demonstrations of AI-generated quantum chemistry circuits on state-of-the-art trapped-ion hardware. The same hardware envelope that hosts our QPDE and tomography proofs is now being used for generative chemistry.
5.Imipramine target
Koziell-Pipe et al. §1The demonstration molecule is imipramine, a tricyclic antidepressant with multiple reactive sites and conformational flexibility. This is drug-scale, not a toy active space — the same regime where the QPDE ethylene benchmark is a small-molecule warm-up.
Candidate composition model · generative + symbolic
ADAPT-GQE is data-driven; our PQP rewriter is syntax-driven. The two are not competitors. A generated circuit from ADAPT-GQE can be handed to the same rule-(N) / rule-(M) / rule-(P) canonicalisation pipeline that proves identities on Selene — shrinking, verifying, and physically executing the result.
In the other direction, the rewriter can supply the generative model with a compact gate alphabet. If the transformer emits circuits over {H, S, CZ}, the 2q oracle and rule-(M) residue already know how to canonicalise every sequence up to length 5. That gives the RL step a verifiable equivalence class to reward, rather than raw shot fidelity alone.
Deferred to v0.4.3: a concrete proof-of-concept would require a small generated chemistry subcircuit (or a hand-crafted surrogate) and a Selene tomography run that confirms the rewriter's canonical residue matches the generated circuit. The QPDE ethylene track is the closest existing chemistry kernel in the repo; once it runs, it can serve as the canonicalisation test case.