Multifidelity Transfer Learning for Data-Efficient Surrogate Modeling of Micro-OLEDs

Microscale organic light-emitting diodes (micro-OLEDs) operating at pixel densities near 10,000 pixels per inch (ppi) require full 3D wave-optics simulations to accurately predict spectral-angular emission characteristics, yet the prohibitive computational cost of such simulations severely limits the training data available for surrogate modeling. In this work, a multifidelity transfer learning framework is used to progressively transfer knowledge across three training stages to construct a high-fidelity surrogate model under limited-data constraints. The proposed surrogate maps device structural parameters to the spectral-angular intensity I(λ, θ) using an encoder–decoder architecture, where wavelength and viewing angle are incorporated as conditioning variables through feature-wise linear modulation. Using only 150 high-fidelity training samples, the transfer-learning pipeline achieves an RMSE of 0.086 in log-intensity space with substantially less high-fidelity data relative to direct learning. A normalized-intensity surrogate is further developed for chromaticity prediction, and rapid gradient-based optimization of micro-OLED emission characteristics is demonstrated with the differentiable surrogates. To the best of our knowledge, this work represents the first surrogate modeling framework for full-wave spectral-angular simulations of micro-OLEDs and enables efficient parametric studies, optimization, and inverse design of high-resolution micro-OLEDs.