Accurate aerosol composition retrievals support radiative forcing assessment, source attribution, air quality analysis, and improved modeling of aerosol–cloud–radiation interactions. Aerosol retrievals based solely on visible-wavelength aerosol optical depth (AOD) observations provide limited spectral sensitivity, which may be insufficient to reliably distinguish among aerosol types with similar optical properties. In this study, we present a new retrieval framework that combines multi-wavelength AOD observations from both the visible and infrared spectrum, enhancing aerosol type discrimination. A neural network forward model trained on simulations from the Model for Optical Properties of Aerosols and Clouds (MOPSMAP), which relates aerosol optical properties to spectral AOD, is embedded in an optimal estimation method (OEM) to retrieve aerosol composition. This machine learning-based forward model achieves computational efficiency without making compromises in accuracy. The neural network forward model achieves a mean R² of 0.99 with root-mean-square error below 0.01. The retrieval resolves up to four independent aerosol components, with degrees of freedom for signal about 3.75. In the total retrieval uncertainty, the forward model contributes less than 10 %, confirming its robustness. We apply this hybrid method to ground-based observations, including data from the Aerosol Robotic Network (AERONET) and Fourier Transform Infrared spectrometer (FTIR) measurements. The retrieved aerosol compositions are consistent with physical expectations and validated through backward trajectory analysis. Furthermore, we successfully apply this method to satellite AOD observations and demonstrate its potential for global aerosol composition retrievals. The full development of a global dataset will be further addressed in future work.