arXiv:2610.26279
Physics-anchored residual quantum neural networks for cloud-cover parameterization
Zhuo Wang, Ce Yu, Bin Yang
Abstract
Compact quantum neural networks have been explored for cloud-cover parameterization, but their predictive value depends on how physical knowledge is incorporated and which classical models provide the reference. We investigate an offline residual formulation that retains a fitted \XR\ cloud-cover estimate and learns a correction to its logit in transformed target space. Using released ICON/DYAMOND-derived data, we compare ten model variants in six- and eight-input settings, with three matched replicates and 10,000 training samples per replicate. All quantum circuits use local Z readout and noiseless exact expectation values. Under a common 30-epoch protocol, residual prediction reduced physical-space test mean squared error relative to direct QNN prediction by 54.3\% and 38.8\% for six-input XYZ and ZZXY circuits, and by 49.2\% and 38.9\% for their eight-input versions. Residual QNNs also outperformed the selected compact residual multilayer perceptron in every paired replicate. A larger, longer-trained Grundner-Q1 adaptation nevertheless achieved lower error: residual Q1 reached mean MSEs of 0.00778 and 0.00791, compared with 0.01124 and 0.01082 for residual XYZ. The results support residual prediction for compact simulated QNNs within the tested budget and establish a clear boundary set by stronger classical references. They do not establish computational quantum advantage or online climate-model skill.
AI Summary
The paper asks whether compact quantum neural networks (QNNs) predict cloud cover better by correcting a calibrated physical diagnostic rather than learning the full mapping. It adds a learned correction to the logit of a transformed Xu–Randall estimate, then applies a sigmoid and inverse target transform. Direct and residual XYZ and ZZXY circuits have identical parameter counts and local $Z$ readout. Experiments use released ICON/DYAMOND-derived data, six- and eight-input configurations, and three matched replicates, each with 10,000 training and 20,000 test cells. Quantum outputs are noiseless exact expectations. Under a shared 30-epoch protocol, residual prediction reduces physical-space test MSE in every QNN comparison: mean paired reductions are 54.3% and 38.8% for six-input XYZ and ZZXY, and 49.2% and 38.9% for eight inputs. Residual QNNs outperform the selected compact residual MLP in every paired replicate. However, larger, longer-trained classical Q1 networks achieve lower errors: residual Q1 reaches mean MSEs of 0.00778 and 0.00791, versus 0.01124 and 0.01082 for residual XYZ. Distributional distances do not improve uniformly. The evidence concerns offline prediction on condensate-filtered samples. Three replicates support descriptive comparisons, not significance claims; temporal or geographical generalization is untested. Residualization simultaneously changes the anchor, output map, and initialization, while input extensions also change circuit size and depth. Stronger compact classical tuning and matched optimization budgets remain untested. The study establishes neither computational quantum advantage nor hardware or online climate-model performance.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.