QBM: A Quantum Bayesian Machine for Relaxed Energy Prediction on Metal-Oxide Catalysts

Technical Summary: QBM: A Quantum Bayesian Machine for Relaxed Energy Prediction on Metal-Oxide Electrocatalysts

Problem Statement
Predicting relaxed total energies for oxide catalysts using Density Functional Theory (DFT) is computationally expensive, necessitating machine learning interatomic potentials (MLIPs) as surrogates. While classical equivariant graph neural networks (GNNs) have established strong baselines, the utility of quantum circuits in this domain remains unclear. Specifically, it is uncertain whether quantum-defined components—such as feature maps, subset selectors, or variational circuits—offer genuine inductive bias or search advantages over matched classical controls, or if observed differences merely reflect candidate pruning, generic nonlinearity, or validation reuse. This study addresses this gap by evaluating a Quantum Bayesian Machine (QBM) on the Open Catalyst 2022 (OC22) IS2RE-Total task, isolating quantum components within a fixed-descriptor framework to determine if they provide benefits beyond classical simulation.

Methodology
The QBM architecture is a modular, sparse surrogate that separates structural encoding, subset selection, regression, and residual correction. The study maintains fixed structural descriptors (atom-centered radial, angular, and torsional distributions projected via Chebyshev polynomials) and data partitions, varying only the downstream computational components. The workflow proceeds through five stages:

1. Descriptor Construction: Atom-centered features are pooled into a global vector and a locality-weighted adsorbate vector, then projected via Partial Least Squares (PLS) to a fixed dimension ($p=16$).
2. Composition Baseline: A ridge regression model removes dominant composition dependence.
3. Subset Selection: The study compares two selectors:
   • Baseline: Exhaustive enumeration of all $2^{16}-1$ non-empty subsets.
   • Quantum (QAOA): A Quantum Approximate Optimization Algorithm (QAOA) proposes binary subsets by optimizing a QUBO surrogate derived from singleton and pair fits.
4. Regression: Selected subsets are fitted using either:
   • Classical Radial Basis Function (RBF) kernels.
   • Quantum Fidelity Kernels (Z and ZZ maps) derived from exact state-vector overlaps.
5. Residual Learning: An ensemble of top-$k$ subset models is combined, and a final residual correction is learned. This learner is either a linear Gaussian Process (ridge equivalent) or a variational data-re-uploading circuit.

All experiments were conducted using exact local state-vector simulations on CPUs. No quantum hardware was used. The study employs a "matched control" discipline, where components are compared in pairs (e.g., QAOA vs. Enumeration, Fidelity Kernel vs. RBF) while holding other variables constant.

Key Contributions

• Component-Level Attribution: The paper provides a documented framework for testing specific quantum components (selectors, kernels, residual learners) within a controlled classical pipeline, explicitly separating quantum search/representation from end-to-end atomistic learning.
• QAOA as a Search Heuristic: It demonstrates that QAOA can reduce the number of fitted candidate subsets (from 65,535 to 333) and selection time (323s to 22s) in a CPU-simulated workflow, though this reduction does not translate to a consistent predictive advantage.
• Spectral Diagnostics: The study introduces diagnostics linking fidelity-kernel effective rank to validation error, observing that increasing kernel scale drives the Gram matrix toward a high-rank regime where validation error increases, consistent with known concentration phenomena in quantum kernels.
• Reproducibility Boundaries: The paper explicitly documents the limits of reproducibility, noting that while aggregate metrics are available, original run configurations, candidate identities, and raw logs for specific fits are not fully archived, preventing independent reconstruction of all reported results.

Results

• Subset Selection: The QAOA workflow fitted significantly fewer candidates (333 vs. 65,535) and was faster. However, the selected QAOA subset was only one Hamming step from the enumerated optimum. The resulting Mean Absolute Error (MAE) showed mixed changes: a slight increase in In-Domain (ID) error (+0.00172 eV/atom) and a decrease in Out-of-Domain (OOD) error (-0.00253 eV/atom). No consistent predictive advantage was established.
• Fidelity Kernels: A selected ZZ fidelity kernel reduced ID MAE by a median of 0.00433 eV/atom compared to a paired linear ridge control across six fits. However, this gain did not transfer to OOD validation, where ridge performed better in five of six pairs. The results suggest the improvement is validation-specific rather than a general quantum inductive bias.
• Variational Re-uploading: A trainable re-uploading circuit showed a median ID MAE reduction of only 0.00128 eV/atom over a linear ridge model across five seeds, with overlapping error ranges. The study notes that the residual targets for these ablations were constructed under a "reduced profile" that omitted canonical cross-fitting, preventing a definitive conclusion on residual-learning advantages.
• Overall Performance: None of the quantum variants consistently outperformed the classical baseline across both ID and OOD splits. The observed variations were often within the range of computational seed noise.

Significance and Claims
The paper modestly claims that QBM serves as a component-level evaluation framework rather than a demonstration of quantum advantage. The primary significance lies in:

1. Delimiting Component Behavior: It establishes that under ideal local simulation, circuit-defined search and representation do not currently provide a general accuracy or computational advantage over matched classical controls for this specific task.
2. Defining the Role of Circuits: It argues that quantum circuits in this context should be viewed as testing alternative subset proposals and similarity geometries, not as replacements for equivariant GNN potentials or end-to-end atomistic learning.
3. Methodological Rigor: By explicitly separating simulation from hardware execution and identifying the need for locked partitions, budget-matched controls, and full artifact archiving, the paper sets a standard for future tests of quantum utility in materials science.

The author concludes that the IS2RE-Total results do not establish quantum-specific predictive benefits. Future work requires matched classical controls, hardware execution studies, and the preservation of per-structure predictions and candidate identities to rigorously test whether circuit-defined inductive biases can survive distribution shifts and noise.