ClimateChem-QX: Quantum-Accurate AI for Climate Catalyst Discovery via Active-Learning-Guided SQD+Krylov Simulations

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Researchers in computational chemistry and materials science should note the demonstration of near-exact accuracy for CO2 reduction pathways and the active learning framework.

🏢実務担当者:R&D teams in chemical engineering and catalysis can use these methods to improve screening of MOFs and homogeneous catalysts for CO2 reduction.

Virtual screening of metal–organic frameworks (MOFs) and homogeneous catalysts for CO₂ reduction and carbon capture depends critically on the accuracy of the potential energy surfaces used to rank candidate materials. Density functional theory (DFT), the universal label-generator for machine-learning force fields, incurs systematic relative-energy errors of 22.9 meV for organic intermediates — comparable to thermal energy at room temperature — and 861 meV for transition metal complexes, tens of times larger. We expect these errors to propagate into ML models trained on DFT data. We introduce ClimateChem-QX, a quantum-classical pipeline coupling Sample-based Quantum Diagonalisation with Krylov augmentation (SQD+Krylov) to a deep-ensemble active-learning surrogate. SQD+Krylov generates near-full-CI training labels within chemically relevant active spaces; active learning identifies geometries of maximum uncertainty for prioritised quantum treatment. We specify, but do not execute, a pathway for transferring these corrections to the MACE universal neural network potential. On the formate (HCOO⁻) CO₂ electroreduction pathway (45 geometries), SQD+Krylov recovers the active-space FCI solution to numerical precision (relative-energy MAE = 0.006 meV) where B3LYP incurs 22.9 meV. For the Cu-mediated CO₂ activation catalyst [CuCO₂]⁻, evaluated on a 99-point geometry grid, DFT relative-energy MAE reaches 861 meV (peak error 1458 meV; 57× thermal energy) with Kendall τ = 0.555, corresponding to a discordant-pair fraction of (1−τ)/2 = 22.3%, and DFT compresses the true energy spread by 38%. SQD reaches 3.5 meV MAE at τ = 0.999 on the same grid, at a sampling operating point shown to lie in the converged regime; Krylov augmentation, characterised separately at a single geometry, lowers the residual by a further order of magnitude. Active learning reduces quantum oracle calls by ~35% relative to random sampling on the HCOO⁻ surface. These results are the necessary precondition for quantum-accurate ML catalyst screening: they establish that the oracle reproduces the exact answer, under controlled convergence, wherever the exact answer can be checked. They motivate the thesis this programme sets out to test — that the accuracy bottleneck in AI for climate catalysis lies in the training data rather than the ML model.

• Zenodo https://zenodo.org/records/21298591first seen 2026-07-12 04:14:16 · last seen 2026-07-12 04:17:06