AI for Parameter Selection in Sample-based Quantum Diagonalization

Doriela Grabocka, Dimitrios Christofidellis, Federico Zipoli, Anna van Elst, Patrick Ruch, Antonio Foncubierta-Rodríguez, Siddhartha Mishra, Matteo ManicaarXiv 2026 · arXiv

TL;DR

Researchers have developed a machine-learning "oracle" that predicts the runtime and accuracy of quantum chemistry simulations before they even run on a quantum processor. By encoding the spectral properties of molecular Hamiltonians into a latent space, the model can automatically select optimal configuration parameters, achieving significant speedups and accurately identifying which molecules can reach the "chemical accuracy" threshold on current hardware.

What Problem Does This Paper Address?

Finding the ground-state energy of a molecule is a fundamental task in quantum chemistry, but doing so for large systems is classically intractable. Sample-based Quantum Diagonalization (SQD) methods—like SKQD and SqDRIFT—offer a promising hybrid quantum-classical alternative. These methods sample quantum circuits to build a low-dimensional subspace where the problem can be solved classically.

However, SQD methods are notoriously sensitive to configuration choices: the subspace dimension ($m$), the propagation timestep ($\Delta t$), and the sampling budget. Currently, there is no formal way to pick these parameters; researchers rely on trial-and-error, which wastes precious time and resources on expensive Quantum Processing Units (QPUs). This paper addresses the need for an a priori selection method to make quantum simulations more efficient and predictable.

Methods

The authors propose a two-stage pipeline to bridge the gap between Hamiltonian structure and simulation performance.

  1. Hamiltonian Encoding: The team developed a custom transformer-based encoder called a Multiscale Byte Language Model (MBLM). It takes a Hamiltonian (decomposed into Pauli terms) and maps it to a 768-dimensional embedding. Crucially, the encoder is trained using a contrastive loss function. This ensures that "isospectral" Hamiltonians—those that share the same energy levels but have different structures—are mapped to similar points in the latent space.
  2. Predictive Modeling: These embeddings are combined with configuration parameters and fed into downstream models (Random Forests and Ordinary Least Squares). These models act as "oracles," predicting two key targets: the approximation error (how close the result is to the true ground state) and the QPU runtime.
Figure 1: The main pipeline. The dataset consists of Hamiltonians that are decomposed into Pauli terms. The Pauli terms are used as input for the encoder, that creates 768-dimensional embeddings for each Hamiltonian. The embeddings are paired with all the configurations. The Hamiltonians and the configurations are executed on the quantum computer, yielding runtime and ground state energy approximation error. The embeddings of the Hamiltonians, combined with the configurations and the results from the quantum runs, create the data- set for training the error and runtime models to predict the ground state energy error and the quantum pro- cessing unit runtime.

Datasets

The study utilized two distinct datasets:

  • Synthetic Hamiltonians: A massive dataset of matrices ranging from 3 to 72 qubits. This was used to train the encoder and validate the regressors against classical Krylov-Arnoldi simulations.
  • Molecular Hamiltonians: 50 molecules (20–50 qubits) across 13 chemical classes. This dataset included 600 real-world runs on the IBM Heron r2 quantum processor to test the model's ability to predict hardware-specific performance.

Key Insights and Ideas

  • Spectral Coherence is Key: The most significant insight is that the simulation's success depends more on the Hamiltonian's eigenvalue spectrum than its specific matrix structure. The contrastive learning approach successfully captured these spectral properties, as shown by the clustering of similar matrices in t-SNE plots.
  • Generalization via Log-Linearity: While Random Forests performed well for in-distribution data, the authors found that linear models (OLS) trained on log-transformed targets generalized better to larger, unseen molecules. This suggests a predictable, power-law-like relationship between system size and simulation cost.
  • Feasibility Estimation: Beyond just picking parameters, the model serves as a feasibility checker. It can tell a researcher, "Given this molecule and this hardware, you will (or won't) hit chemical accuracy within $X$ minutes of QPU time."
t-SNE visualization showing how the encoder clusters Hamiltonians by their spectral properties.

Key Results

  • Strong Correlations: On synthetic data, the models achieved Spearman rank correlations of $\rho = 0.969$ for error and $\rho = 0.905$ for runtime, meaning they are excellent at ranking which configurations are better than others.
  • Efficiency Gains: In classical simulations, the model-selected configurations achieved a 2.53x speedup compared to standard theoretical baselines while maintaining lower errors.
  • Hardware Success: When tested on real IBM hardware for molecular systems, the model achieved 100% balanced accuracy in classifying whether five large, held-out molecules could reach chemical accuracy ($1.6 \times 10^{-3}$ Hartree).
  • Out-of-Grid Generalization: The model remained robust even when tested on configuration parameters (like num_groups) that were entirely outside the range seen during training.

Limitations and Future Work

The authors identify several hurdles for future research:

  • Hardware Coupling: The runtime predictions are currently "tightly coupled" to the specific IBM Heron r2 processor used. Transferring these models to different quantum architectures remains an open challenge.
  • Ground-Truth Bottleneck: Validating the model on even larger molecules is difficult because we lack classical "exact" ground-state energies to compare against for systems beyond 50 qubits.
  • Configuration Diversity: While the model generalizes well, the authors suggest that more diverse training data (wider ranges of $m$ and $\Delta t$) is needed to build a truly universal parameter selector.

Summary updated

Abstract

Sample-based Quantum Diagonalization (SQD) methods provide a way to compute molecular ground-state energies, but their accuracy and runtime depend sensitively on configuration choices such as the subspace dimension, propagation timestep, and the allocation of a fixed sampling budget across circuits, for which no formal selection method exists. In this work, we explore whether oracle models can predict the quantum runtime and approximation error from the Hamiltonian and the candidate configuration. Using synthetic Hamiltonians with controllable spectral properties, we show that machine-learned models can map a Hamiltonian directly to the best-ranked configuration within a predefined grid for the SKQD algorithm, validated against classical Krylov propagation, achieving strong rank correlations between true and predicted error and runtime, Spearman $ρ= 0.969$ and $ρ= 0.905$, respectively. Extending the approach to 50 molecular Hamiltonians (20-50 qubits), with reference ground-state energies from classical calculations, we train regressors that predict the simulation error and runtime for a given molecule--configuration pair ahead of execution, and classifiers to distinguish, from 600 SqDRIFT quantum-hardware runs, whether a configuration achieves a ground-state energy within chemical accuracy ($1.6\times10^{-3}$ Ha $\approx 1$ kcal/mol). For the molecular Hamiltonians, the models additionally provide an a priori feasibility and resource estimate, predicting whether a molecule can reach chemical accuracy on current hardware and at what runtime, thus correctly classifying held-out molecules at the molecule level, achieving balanced accuracy 61.8% in-distribution and 100% for five larger held-out molecules.
BibTeX
@misc{grabocka2026ai,
  author = {Doriela Grabocka and Dimitrios Christofidellis and Federico Zipoli and Anna van Elst and Patrick Ruch and Antonio Foncubierta-Rodríguez and Siddhartha Mishra and Matteo Manica},
  title = {{AI for Parameter Selection in Sample-based Quantum Diagonalization}},
  year = {2026},
  eprint = {2611.18307},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18307},
}
Figure