Graph Neural Networks for Fast Operator Selection in Adaptive VQE

Abstract: Adaptive variational quantum algorithms like ADAPT-VQE construct tailored ansätze by iteratively selecting operators from a pool using gradient-based criteria. While this avoids oversized parameter spaces, repeatedly scanning the full pool incurs a classical cost that scales linearly with pool size-a major bottleneck for systems with long-range interactions or large operator sets. Here, we reformulate adaptive operator selection as a graph-based decision problem and introduce a graph neural network (GNN) policy that predicts the next entangling operator directly from the interaction graph and state-dependent observables. Training data are generated from exact simulations of disordered long-range spin chains, using gradient magnitudes as supervision signals. The learned policy accurately reproduces the dominant structure of the greedy gradient-based selection rule, significantly outperforming heuristics based solely on interaction strength. Integrated into a variational quantum eigensolver (VQE) workflow, this GNN-VQE approach achieves energy errors close to standard ADAPT-VQE while drastically reducing full-pool gradient evaluations. To test transferability beyond spin models, we evaluate the policy on small active-space molecular benchmarks (LiH and BeH_$2$). We find the GNN is highly effective as a shortlist generator: exact rescoring over just a few GNN-proposed candidates recovers near-oracle rollout behavior while searching only a small fraction of the pool. These results demonstrate that adaptive circuit construction contains learnable structure that can be exploited to accelerate variational quantum algorithms.

• The paper proposes a GNN-based policy that replaces costly gradient evaluations with a learned, graph-structured operator selection process.
• It achieves ~40–45% top-1 accuracy on operator selection, outperforming random and heuristic methods on disordered spin systems.
• The framework effectively shortlists candidate operators in molecular applications, reducing computational overhead while maintaining near-oracle energy performance.

Overview of GNN-Based Operator Selection for Adaptive VQE

The paper "Graph Neural Networks for Fast Operator Selection in Adaptive VQE" (2606.08794) presents a novel framework for mitigating the computational bottleneck in adaptive variational quantum eigensolver (VQE) algorithms by replacing explicit, iteration-intensive operator scans with a graph neural network (GNN) policy for entangling gate selection. The approach shifts the adaptive ansatz construction problem from an expensive repeated gradient evaluation over large operator pools to a learned decision-making process leveraging the interaction graph and quantum state observables.

Reformulation of Adaptive VQE Operator Selection

Adaptive VQE algorithms, such as ADAPT-VQE, build variational circuits stepwise by selecting, at each iteration, the operator maximizing the gradient with respect to energy reduction. While this delivers ansatzes tailored to the underlying physics and prevents superfluous parameter growth, this operator search induces a linear-in-pool-size overhead per iteration—a significant limitation for systems with disordered, long-range couplings or when handling large candidate operator sets.

This work proposes reframing operator selection as a sequential graph decision process. The quantum system is represented as a weighted interaction graph, with nodes as sites (e.g., spins, orbitals) and edges as candidate two-body operators. Both static features (geometry, coupling strengths) and state-dependent observables (local magnetizations, correlators) are embedded within graph nodes and edges. Operator ranking is then predicted directly from this joint representation, bypassing explicit gradient computation.

GNN Architecture and Training Pipeline

The GNN operator-selection policy is a message-passing neural architecture with the following salient features:

• Input Features: Nodes include geometric and interaction descriptors, as well as quantum-state-dependent quantities; edges encode logarithmic coupling strengths, geometric separations, dominance measures, correlators, and operator variances.
• Sparsification: For scalability, only the $K$ strongest couplings per node are retained, leading to a reduced but informative operator pool.
• Message Passing and Scoring: Multiple edge-conditioned message-passing layers propagate information, combining latent node and edge embeddings. At each construction step, all candidate operators (edges) are scored via a learned multi-layer perceptron, utilizing the embeddings of both incident nodes and the edge feature.
• Supervision: Labels are generated by the commutator-based energy gradient, i.e., $$g_{ij} = \left| \langle \psi | [H, P_{ij}] | \psi \rangle \right|$$, computed for all candidates via statevector simulations. Temperature-controlled soft labels are used for robust target distributions.

The system is trained on data from disordered long-range spin chains of varying interaction exponents, using sampled random product states as initial conditions. The GNN is optimized to mimic the one-step greedy gradient-based selection rule, the de facto standard in adaptive VQE methods.

Numerical Results and Critical Evaluation

Operator Ranking and Rollout Performance

The GNN achieves a significant improvement in top-1 operator selection accuracy over both random and strongest-coupling heuristics, reaching ~40–45% accuracy on validation data for operator pools of typical size 20–30. More importantly, the oracle-selected operators generally appear among the highest-ranked candidates in the GNN's predictions. The model generalizes well across a broad range of disorder and interaction profiles, consistently outperforming static baselines.

When the GNN policy is used to sequentially build circuits in a greedy rollout regime (fixed-angle updates), energy reduction per step tracks well with the true gradient oracle and outperforms all heuristic approaches, though a residual performance gap to the oracle persists, attributable to incomplete recovery of the full state-dependent gradient landscape.

Ansatz Quality After Full Optimization

After compiling a circuit by the GNN policy and globally optimizing all gate parameters via VQE, the final energy error remains close to that of ADAPT-VQE, particularly for moderate circuit depths. The GNN-VQE approach consistently yields lower energy errors than the strongest-coupling heuristic. Notably, random operator selection—while suboptimal in greedy rollout—can occasionally produce competitively low energy errors after full parameter optimization, highlighting the complex relationship between operator ordering and ansatz expressivity for small systems.

Computational Complexity and Scaling

The main advantage of GNN-VQE is the elimination of per-iteration, per-operator quantum measurements required by ADAPT-VQE. The classical cost per layer in GNN-VQE is $O(L |E| d^2)$ (with $L$ the number of GNN layers and $d$ the latent dimension), compared to $O(|E| C_{\text{exp}})$ for ADAPT-VQE (where $C_{\text{exp}}$ is the cost for evaluating each operator gradient expectation on hardware). This cost-shifting is especially favorable for near-term devices and classically unsimulable regimes, where state preparation and measurement dominate runtime.

In current small-scale, classical emulations, overheads related to graph construction and tensor operations in the GNN may exceed the cost of brute-force gradient-based selection, but this is expected to invert at relevant quantum device scales.

Ablation Analysis

The study reveals that state-dependent correlator features and message-passing modules are essential to the GNN's predictive power; removing them halves the validation accuracy and undermines mean rank fidelity. Static structural features alone are insufficient. Operator-variance input provides a modest gain, while explicit inclusion of the interaction exponent or operator-rank features yields negligible impact for the considered system sizes.

Transfer to Molecular Systems and Shortlisting Strategy

The GNN policy, trained on spin system data, is evaluated on LiH and BeH₂ active-space molecular problems, following Jordan-Wigner mapping. As a standalone selector, pure GNN rollout does not fully match logistic-regression baselines; however, as a candidate shortlisting tool, the GNN is highly effective. Exact rescoring over 2–4 GNN-nominated candidates (out of 14–52 total) is sufficient to recover near-oracle selection and energy errors. This "learned shortlist" paradigm demonstrates substantial reduction of the classical screening burden while preserving high-quality ansatz growth in molecular settings.

Theoretical and Practical Implications

This work demonstrates that adaptive ansatz construction exhibits learnable, graph-structured patterns. The GNN-policy not only captures nontrivial subsets of the gradient-based selection rule but, more importantly, reduces the dependence on repeated full-pool evaluations that are a primary bottleneck in adaptive VQE deployments. Especially for platforms with programmable interactions (trapped ions, Rydberg arrays, superconducting qubits), this approach enables scalable, circuit-efficient Hamiltonian engineering.

The most robust role of the GNN is as a shortlist generator—screening large operator sets and relegating fine-grained selection to resource-intensive quantum or classical oracles. Extending training and validation to larger molecular and materials model systems, integrating with hardware-efficient ansätze, and cross-Hamiltonian transfer learning are promising directions for further reducing quantum-classical overheads.

Conclusion

By recasting adaptive variational circuit construction as a graph-based learning task, this paper provides an operator selection framework that leverages both physical interaction structure and quantum state information, efficiently encoded by GNNs. GNN-VQE achieves operator selection accuracy and final variational energies approaching those of state-of-the-art gradient-based methods, but at a drastically reduced classical or quantum evaluation cost. Notably, use as a shortlist generator in molecular active spaces enables near-oracle performance with substantial screening cost reduction, opening new avenues for scaling adaptive VQE and related quantum-classical algorithms in the NISQ era and beyond.