Scaling a CUDA-Q GQE + QSCI pipeline to 40 qubits for EUV photoresist chemistry

AI summary

Overview

Research area: Quantum computing for electronic structure — specifically near-term hybrid quantum–classical quantum chemistry algorithms applied to semiconductor materials chemistry (EUV photoresist precursors).

Technical level: Advanced. The abstract assumes familiarity with active spaces, CASCI/FCI references, UCCSD operator pools, variational bounds, and superconducting-gate hardware metrics.

Scope (one sentence): The paper reports a CUDA-Q-native GQE + QSCI pipeline benchmarked against exact diagonalisation from 14 to 40 qubits on monoalkyltin oxo-hydroxide photoresist chemistry, with selected demonstrations on a 54-qubit IQM superconducting processor.

What This Paper Is About

The authors take a generative quantum eigensolver — where a language model (GPT-2) proposes sequences of unitary excitation operators — and couples it to quantum-selected configuration interaction, a method that samples bitstrings from those circuits, treats them as electron configurations, and diagonalises the resulting subspace on a classical computer. The goal is to test how far this hybrid loop can be pushed toward chemically useful accuracy for a chemically awkward but industrially important family of molecules: the tin-oxo-hydroxo compounds used in extreme-ultraviolet photoresists. The paper charts where the method remains accurate against exact references, where hardware noise and subspace size start to dominate, and how the chemistry of these resists behaves under ionisation.

Key Contributions

1. A scaled, CUDA-Q-native GQE + QSCI pipeline spanning active spaces from 14 to 44 qubits, with a cross-circuit generalised-eigenvalue refinement that guarantees every reported energy is a variational upper bound rather than an uncontrolled estimate.

2. A rung-by-rung validation ladder against exact theory: every point from 14 to 40 qubits is compared to a CASCI or FCI reference, including a 166-million-determinant FCI calculation for SnO at 32 qubits — an unusually complete accuracy audit for a generative ansatz.

3. Hardware execution on a 54-qubit IQM Emerald processor, using depth-truncated trained circuits, a CCSD-amplitude-ordered operator-pool prefix, and per-circuit readout self-calibration, plus classical configuration recovery on the measured counts.

4. A chemically motivated analysis of the resist mechanism, showing how ionisation of the methyl resist sharply lowers the classical UCCSD(T) Sn–C bond dissociation energy, which the authors identify as the solubility-switching event on exposure.

Main Findings

• Chemical accuracy is reached through mid-sized active spaces. Errors stay below 1.6 mHa through 30 qubits for methyltin trihydroxide and through 32 qubits for SnO, on the best seed at the top rungs.

• Circuit depth, not training length, controls scaling. The authors identify depth as the lever that governs how the method behaves as qubit count grows, rather than the amount of optimisation applied to the ansatz.

• The refined subspace stays small. It grows near-linearly with operator count while remaining a vanishing fraction of the full determinant space — 0.017% at the 32-qubit SnO rung.

• Hardware results are accurate but error-bounded. On IQM Emerald, SnO at 14 qubits reached +0.330 mHa using a pool prefix ordered by CCSD amplitudes; the industrial n-butyltin ligand at 22 qubits reached +3.92 mHa from depth-truncated trained circuits, capturing 81% of the active-space correlation energy.

• Classical post-processing recovers much of the hardware error. Configuration recovery applied to the measured bitstring counts tightened the 22-qubit hardware result to +0.18 to 0.21 mHa.

• Ionisation drives the resist's solubility switch. Classical UCCSD(T) shows the Sn–C bond dissociation energy of the methyl resist collapsing from 72.6 to 21.2 kcal/mol upon ionisation.

• The 40-qubit frontier is mapped, not surpassed. The 40-qubit result is support-limited at 22.8 mHa, the full trained ansatz has not yet run on hardware at acceptable fidelity, and classical subspace expansion can already reach the 32–40-qubit spaces without any quantum sampler.

Methodology in Plain English

The strategy is to let a small language model (GPT-2) act as a proposer: it writes out sequences of quantum operations drawn from a standard chemically motivated pool (UCCSD excitations, or a pool ordered by classical coupled-cluster amplitudes). Those circuits are run — in simulation, or on real hardware — and the measured bitstrings are read as electron configurations. A classical computer then builds a matrix in that small set of configurations and diagonalises it to get an energy. To keep results trustworthy, the authors combine circuits from several different points in the optimisation into a single generalised eigenvalue problem, which guarantees the answer sits above the true energy. This is checked by running the same systems with exact methods that enumerate the full configuration space, so each qubit count has a ground truth to compare against. On hardware, circuits are deliberately shortened (depth-truncated) to survive the device's two-qubit gate routing, and readout errors are recalibrated circuit by circuit.

Why This Matters

Research impact: The paper supplies a rare thing in near-term quantum chemistry — a continuous ladder of system sizes where every quantum result has an exact classical reference, including a 166-million-determinant FCI comparison. That makes it possible to see precisely where the generative ansatz stops being useful, and to separate algorithmic limits from hardware limits. The variational-bound guarantee also gives the method a property many hybrid pipelines lack: reported energies are upper bounds by construction, not just in expectation. The explicit statement that classical subspace expansion already covers the 32–40-qubit range is a useful, honest calibration of the quantum advantage boundary.

Real-world applications:

• EUV lithography photoresist design — the tin-oxo-hydroxo systems studied are candidate precursors for next-generation resist chemistries, where the exposure mechanism (ionisation, bond scission, solubility flip) is the property that determines patterning performance.
• Semiconductor process chemistry — understanding bond dissociation and radiation-induced chemistry supports resist screening and process-window prediction at advanced nodes.
• Organometallic materials discovery — the pipeline generalises to other metal–ligand systems where strong correlation makes mean-field methods unreliable.
• Benchmarking quantum hardware for chemistry — the hardware protocol (depth truncation, amplitude-ordered pools, per-circuit readout calibration) is transferable to other device evaluations.

Industry relevance: The target chemistry is directly tied to semiconductor manufacturing, and the paper explicitly reports on an industrial ligand (n-butyltin) alongside the simpler model systems. The inclusion of a 54-qubit commercial superconducting device in the evaluation reflects an interest in practical deployment rather than simulation-only results — though the authors are candid that the largest trained ansatz has not yet run on hardware at usable fidelity.

Future Directions

• Improving hardware fidelity to run the full trained ansatz. The paper states that the complete trained circuit at the largest scales awaits better device fidelities; closing that gap is the obvious next experimental step.

• Beating the support-limited 40-qubit regime. With the 40-qubit result at 22.8 mHa and classical subspace expansion already reaching 32–40 qubits without a quantum sampler, the open question is what algorithmic or hardware change would push the quantum path past that boundary.

• Extending beyond 44 qubits. The abstract names a 44-qubit ceiling for the study; scaling the ladder further requires addressing both circuit depth growth and the cost of the classical diagonalisation step.

• Broadening the chemistry. Applying the validated pipeline to other resist precursors and ligands — and testing whether the ionisation-driven dissociation picture holds across the monoalkyltin family — would establish whether the chemical insight generalises.

Target Audience

Quantum algorithm researchers working on near-term electronic structure methods, particularly those interested in generative or machine-learning-proposed ansätze and in rigorous benchmarking against exact references. Quantum computing hardware teams evaluating superconducting processors on chemistry workloads will find the device-level protocol and the honest fidelity assessment useful. Computational chemists and materials scientists working on EUV photoresist or organotin chemistry will benefit from the mechanistic finding on ionisation-driven bond weakening, though the algorithmic content is aimed at the quantum-computing side.

Authors’ abstract

We scale a CUDA-Q-native pipeline coupling a generative quantum eigensolver (GQE) to quantum-selected configuration interaction (QSCI) across active spaces of 14 to 44 qubits, applied to the extreme-ultraviolet (EUV) photoresist chemistry of monoalkyltin oxo-hydroxides. A GPT-2 policy emits UCCSD operator sequences; sampled bitstrings become determinants, diagonalised classically, and a cross-circuit generalised-eigenvalue refinement makes every reported GQE+QSCI energy a variational upper bound. Every rung from 14 to 40 qubits carries an exact CASCI or FCI reference, up to 166 million determinants for SnO at 32 qubits. The pipeline is chemically accurate, below 1.6 mHa, through 30 qubits on methyltin trihydroxide and through 32 on SnO, on the best seed at the top rungs. Circuit depth rather than training length is the scaling lever; the refined subspace grows near-linearly with the operator count while staying a vanishing fraction of the determinant space, 0.017% at the 32-qubit SnO rung. It also runs on the 54-qubit IQM Emerald processor, at the shallow depths its routed two-qubit gates allow, reaching +0.330 mHa for SnO at 14 qubits from a CCSD-amplitude-ordered pool prefix and +3.92 mHa for the industrial n-butyltin ligand at 22 qubits from depth-truncated trained circuits under per-circuit readout self-calibration, 81% of the active-space correlation; classical configuration recovery on those counts tightens the 22-qubit result to +0.18 to 0.21 mHa. For the methyl resist, ionisation collapses the classical UCCSD(T) Sn-C bond dissociation energy from 72.6 to 21.2 kcal/mol, the switch that flips solubility on exposure. Against that, the 40-qubit result is support-limited at 22.8 mHa, the full trained ansatz on hardware awaits better fidelities, and classical subspace expansion reaches the 32 to 40-qubit spaces with no quantum sampler, so that boundary is mapped, not beaten.