Hybrid Quantum Graph AI Research Summary

Research Summary • Comparative Findings

Learning QAOA Parameters with Graph Neural Networks for Scalable Quantum Optimization

The Quantum Approximate Optimization Algorithm remains one of the central variational paradigms for near-term quantum optimization because it provides a hardware-compatible ansatz for combinatorial objectives while retaining a clear Hamiltonian interpretation. Its practical performance, however, depends not only on circuit depth but also on the ability to identify high-quality variational parameters for each problem instance. For fixed-depth QAOA, parameter optimization is often the dominant classical bottleneck, especially when one repeatedly solves related weighted-graph instances drawn from a common structured family. In this work, we address that gap by learning QAOA parameters from graph structure, modeling fixed-depth parameter selection as a graph-to-angle map and using a graph neural network to predict depth-2 QAOA angles for weighted MaxCut instances derived from transcriptomic co-expression graphs.

Held-out approximation ratio
0.8682

Versus 0.8686 for direct multi-start search on the same held-out transcriptomic graph family.

Median online latency
0.3652 ms

Versus 936.1255 ms for repeated classical search in the executed fixed-depth evaluation regime.

Noisy held-out ratio at eta = 0.05
0.7850

Remaining close to 0.7902 from classical-search angles under the executed depolarizing-noise study.

Research Summary

This work shows that structure-aware learning can predict instance-dependent parameters for fixed-depth QAOA on a structured family of weighted MaxCut graphs and that this mechanism can be analyzed quantitatively. In the transcriptomic setting studied here, the archived benchmark shows near-parity with direct classical search and a large runtime advantage, while the rerun studies show that this retained-quality claim is robust across moderate seed variation but limited by protocol sensitivity and family complexity.

Taken together, the results identify structure-aware parameter prediction as a viable route to transferring fixed-depth quantum optimization across related weighted instances, but only within a bounded operating regime that the present research makes explicit. Extending this framework to broader weighted graph families, stronger graph-learning architectures, more stable training procedures, and hardware-side evaluation is the natural next step.

Comparative Results

The comparison is framed against direct depth-2 search and simpler learned initializers. The strongest claim is retention of the downstream MaxCut objective on held-out transcriptomic graphs with sharply lower inference-time cost.

Evaluation Context Leading Method Winning Evidence Comparison To Existing Methods
Held-out transcriptomic benchmark Graph-conditioned GNN Mean approximation ratio 0.8682, runtime 0.3652 ms Retains 99.95% of the direct-search objective while replacing repeated 936.1255 ms search with fast graph-conditioned inference.
Executed depolarizing-noise evaluation Graph-conditioned GNN Held-out ratio 0.7850 at eta = 0.05 Remains close to 0.7902 from classical-search angles across the tested noise grid on the same held-out family.
Supporting biomedical operating point Adaptive BioGCN configuration 98.8% accuracy with one false positive Secondary result showing the same graph-to-decision framing extends beyond the primary optimization branch.

The page is limited to the strongest fixed-depth findings in the executed repository outputs. It does not present these results as evidence for asymptotic quantum speedup.

Primary Benchmark

Near-direct-search quality on held-out graphs

On the real transcriptomic graph family, the graph-conditioned predictor nearly matches direct multi-start optimization while avoiding per-instance outer-loop search at inference time.

Operational reading: repeated fixed-depth QAOA evaluation becomes materially cheaper without sacrificing the downstream objective in the tested regime.
Noise Evaluation

Stable degradation under executed noise

The learned initializer degrades smoothly rather than abruptly across the tested depolarizing-noise grid, staying close to the classical-search reference even at eta = 0.05.

Operational reading: the fixed-depth learned initialization remains usable when idealized statevector assumptions are relaxed in the executed noise study.
Secondary Branch

Graph-to-decision framing extends to CTG

The biomedical branch supports methodological breadth, but it remains secondary to the transcriptomic QAOA result and should be read as supporting evidence rather than the headline physics claim.

Operational reading: the repository demonstrates a reusable graph-learning interface across optimization and biomedical decision settings.

From Benchmarks to Applications

The benchmark is organized around repeated fixed-depth optimization on structured weighted graphs. In that setting, a better model is not merely numerically stronger. It changes how often full parameter search must be rerun, how cheaply related instances can be evaluated, and how confidently objective retention can be audited.

  • In transcriptomic optimization, the graph-conditioned result means new related biological graphs can be evaluated without rerunning expensive multi-start search from scratch, while staying close to the direct-search objective.
  • In noise-aware evaluation, the supporting result means the learned initialization remains close to the classical-search baseline as local depolarizing noise increases across the tested grid.
  • In the secondary CTG setting, the same structural learning approach can be reused for cohort-level decision tasks, even though that branch is not the primary contribution of the page.

Supporting figures

The figures below are presented at larger on-page size and can be opened at full resolution directly from the root page. Each panel is tied to the corresponding notebook or manuscript evidence.

QAOA landscape geometry

This figure shows that the visible depth-2 high-value basin is narrow rather than diffuse, which is why good learned warm starts materially reduce the local search burden.

Open notebook context
Use the full-resolution view to inspect the basin concentration and near-optimum fraction more clearly.

Held-out benchmark overview

The main comparison is anchored to direct search rather than to weak initializers, which is the relevant operational ceiling for learned fixed-depth QAOA initialization.

Open notebook context
This benchmark panel is sized to keep comparison text and summary statistics readable on the landing page.

QAOA state concentration

Parameter concentration across graph families shows that the learned initializer remains stable and transfers across held-out transcriptomic graphs.

Open paper context
Generated from the executed QAOA demo notebook.

Held-out CTG evaluation

The biomedical branch is shown with its strongest evaluation figure rather than a topology overview so the landing page foregrounds quantitative evidence.

Open notebook context
This replaces the smaller CTG graph overview so the root page shows the clearest evaluation panel available.

Live root-page QAOA demonstration

The root page now includes the project demo directly. Generate a graph, predict QAOA angles, and inspect the returned parameters here without leaving the landing page. When the API is unavailable, the demo falls back to an in-browser exact depth-1 simulation and local refinement.

Interactive Controls

Generate a weighted graph and predict QAOA parameters

This is a compact live view of the repository demo. It is intended for quick inspection of the graph-conditioned parameterization workflow rather than full benchmark reproduction.

6
The range is kept small so exact browser-side fallback remains responsive and the visual graph remains readable.
6 nodes · no edges yet — click Random Graph
Main fixed-depth result: graph-conditioned GNN ratio 0.8682, direct-search ratio 0.8686, learned-runtime 0.3652 ms, direct-search runtime 936.1255 ms.

Primary results

These materials contain the manuscript, rendered notebooks, and repository sources supporting the benchmark presented on this page.

Figure preview
Expanded figure preview