Intelligent Bayesian Optimization for Semantic-Aware Energy Conversion System Design

The study presents an intelligent Bayesian optimization framework with semantic-aware decision-making for designing coupled material–process energy conversion systems. A Gaussian process surrogate with heteroscedastic noise modeling encodes experimental uncertainty, while a composite acquisition function fusing expected improvement, probability of improvement, and upper confidence bound, together with an adaptive diversity penalty, dynamically balances exploration and exploitation via real-time model diagnostics. Applied to optimizing five fabrication variables of a near-infrared-to-visible upconversion layer in inverted perovskite solar cells, the closed-loop campaign identifies optimal conditions within six rounds starting from 16 initial experiments, yielding a champion device with 22.41% efficiency and 24.52 mA cm-2 short-circuit current density, gains of 13.18% and 8.4% over the control. The adaptive strategy outperforms a fixed-policy baseline and reduces trials by three orders of magnitude versus grid search, establishing a generalizable artificial intelligence–driven methodology for engineering optimization.