Prediction and inverse design of effective optical properties of polymer nanocomposites empowered by machine learning

The complex refractive index of polymer nanocomposites (PNCs) is critical for advanced optical devices, but the agile discovery of PNCs with high refractive index remains challenging. We developed a machine-learning model that instantly predicts wavelength-dependent optical properties across varying volume fractions and particle sizes and inversely designs PNCs with targeted functionalities. A high-quality dataset was constructed using the Finite Element Parameter Retrieval (FEPR) method, integrating finite-element analysis and particle swarm optimization. A Physics-driven Neural Network (PNN) achieved forward prediction with relative errors in the refractive index n below 0.05% for 97.9% of test samples. For inverse design, a Bidirectional PNN (Bi-PNN) with physical constraints achieved rapid single-solution design with 94.2% of relative errors in n below 0.05%, and a hybrid PNN-genetic algorithm (PNN-GA) framework explored multiple viable solutions. This work offers an efficient, readily transferable, and practically valuable computational paradigm for evaluating and optimizing complex PNC systems.