Bidirectional deep learning for chromatic-optical predicting and inverse design of LED systems with multi-objective...

A unified experimental-computational framework was developed to model and optimize the chromatic-optical performance of multi-primary light emitting diodes (LEDs). A dataset of 21,296 measurements, obtained by systematically varying heatsink temperatures and per-channel driving currents, was used to train three deep-learning architectures independently: an autoencoder (AE), a long short-term memory network (LSTM), and a gated recurrent unit (GRU). Thereby, forward mappings from electro-thermal setpoints to spectral power distributions (SPDs), optical and chromatic parameters were learned, and conversely inverse mappings from these optical observables to operating temperatures and currents were inferred. These mappings are explained by coupled electro-thermal-optical mechanisms; specifically, temperature-induced band-gap narrowing and current-dependent carrier recombination drive the observed nonlinear spectral shifts and flux variation. Based on the learned forward and inverse models, a normalized multi-objective score was constructed and optimized by a genetic algorithm (GA) across two application CCT bands and three priority scenarios. The AE achieved the highest forward fidelity (R² = 0.999; MSE = 7.0 × 10^-4 on the hold-out test sets), whereas the GRU provided the most accurate inverse estimates, with mean relative errors of 0.24% (heatsink temperatures) and 0.41% (per-channel currents). GA results quantified practicable circadian action factor (CAF) - luminous efficacy (E_f) - color gamut coverage (G_c) trade-offs and delivered implementable heatsink temperatures and currents setpoints. The proposed approach captures nonlinear electrothermal-optical interactions that provide a physics-consistent, application-oriented framework that translates learned spectral-performance mappings into implementable drive and thermal setpoints, thereby offering lighting and display engineers what we believe to be a novel quantitative tool for balancing circadian, photometric, and colorimetric objectives.