Partitioned adaptive ensemble of surrogate models toward wavefront aberration prediction of large-aperture mirror...

The analysis and prediction of wavefront aberrations (WFAs) in large-aperture mirror assemblies (LAMAs) is computationally expensive, and selecting a suitable single surrogate model (SSM) is challenging. Existing ensembles of surrogate models (ESMs) often fail to exploit the local-fitting advantages of SSMs across subregions, thereby limiting prediction accuracy. To address these issues, this paper proposes a partitioned adaptive ensemble of surrogate models (PAESM) toward predicting multiple-order WFAs of LAMAs. PAESM employs a three-layer mechanism-local model screening, intra-partition adaptive weighting, and inter-partition robust fusion-based on a partitioned integration architecture. A local-global combined evaluation strategy is developed for model screening using design space partitioning and compensated cross-validation error. An error-modulated distance-weighting method integrates model error with spatial information, thereby reducing reliance on the reliability of the error metric. Numerical experiments on the 36th-order WFA prediction demonstrate that PAESM outperforms conventional SSMs and ESMs across different sample sizes and thermal conditions, achieving mean R² values of 0.919 and 0.906 under 20-50 °C and 0-70 °C conditions, respectively. Furthermore, PAESM is integrated with a genetic algorithm to optimize the support structure of LAMAs, reducing the RMS of WFA by over 70% while cutting single-evaluation time from 8 minutes to 0.031 seconds. These results demonstrate the effectiveness, robustness, and engineering applicability of the proposed method.