MAESTRO: a multi‑fidelity modeling assisted by an ensemble of surrogates for robust prediction

초록

We introduce MAESTRO, a multi-fidelity framework designed to maintain accuracy under small data budgets . The method trains five surrogate families in parallel for both the low-fidelity (LF) response and the high-fidelity (HF)–LF discrepancy, prescreens unstable candidates via a novel gradient-distribution detector and a median-based trimmed Z-score, and then performs leave-one-out cross-validation for model selection. Across seven numerical problems (2-D/5-D/10-D) and two engineering applications (nonlinear wing-rib stress; vehicle bounce dynamics), MAESTRO consistently outperforms single-fidelity surrogates and multi-fidelity baselines—including co-kriging, a multi-fidelity neural network and an ensemble cross-validation multi-fidelity surrogate—in r2 and normalized root‑mean‑square error, while maintaining stable performance across tasks. The framework readily integrates into computer‑aided engineering workflows for design exploration and optimization. © 2026 Elsevier Ltd.

키워드