Solvent composition analysis is key to determining the performance of the carbon capture process. Current efforts focus on the optimization of sample preparation, analysis time, and method reliability. This work introduces an innovative approach that combines a machine learning algorithm with Fourier transform infrared (FTIR) spectroscopy for the precise and accurate quantification of carbon dioxide (CO2) and monoethanolamine (MEA) in carbon capture solvent with a limited number of samples in the training phase. The framework consists of feature selection that combines statistical and domain-knowledge analysis to choose the appropriate wavenumber that represents both MEA and CO2 accurately. Eleven machine learning models are compared in terms of accuracy to predict the MEA and CO2 concentration in solution given the absorbance from the previously selected wavenumber. The result shows that extreme gradient boosting and adaptive boosting are the best models for predicting MEA and CO2 concentration, respectively. This study obtained a similar accuracy compared to the literature with a data set reduced by a factor of 2.63 on average. The framework showcases the synergistic potential of ML associated and FTIR spectroscopy, offering a robust methodology for routine analysis of MEA and CO2 concentration in the solvent.