Inverse design of high-quality multi-channel LED spectra enabled by feedforward neural network

Abstract High-quality spectra of LED lighting have profound implications for visual comfort, human health, efficiency, and other specialized applications. The ability to precisely tune and control the spectral output of LEDs is becoming increasingly crucial. In this study, we propose an inverse design framework utilizing a Feedforward Neural Network (FNN) to design high-quality spectra of LED lighting. Firstly, a dataset of spectral power distributions (SPDs) of multi-channel LEDs is experimentally collected by applying 20% increment pulse width modulation (PWM) combinations. These experimentally acquired data facilitate an accurate learning process for the nonlinear mapping from PWM combination inputs to SPDS outputs, which is essential for training the FNN. The trained FNN subsequently predicts SPDs at finer 1% PWM increments, from which chromaticity parameters are derived. Optimal PWM inputs are then selected based on target specifications, achieving inverse mapping from desired chromaticity parameters back to the corresponding PWM combinations. Validation of predicted spectra demonstrates that SPDs generated by the selected PWM combinations comply with ANSI standards, with deviations in chromaticity parameters remaining within acceptable limits. This approach overcomes the computational challenges associated with traditional enumeration methods and limitations of conventional optimization algorithms, providing a computationally efficient and dynamically stable solution for real-time spectral tuning in intelligent lighting systems.