Population composition is crucial in exploring genetic associations and investigating conditions and diseases. Single nucleotide polymorphisms (SNPs) are the subject of extensive studies, as they represent the most prevalent genetic variability in the human population. This work proposes a hierarchical framework for nonlinear probabilistic clustering of individuals with mixed ancestry population components. Through methods such as Kernel PCA, latent variables that can capture complex patterns of genetic variation are found. Gaussian mixture clustering allows for the inference of the population structure of the data obtained through the proposed feature extraction model. The proposed method is trained with pure populations from Africa, Europe, East Asia, and America, achieving an adjusted rand index of 0.981 for the evaluation set. Validation is achieved by evaluating the method on real data sets and compared with results from previous studies, achieving a Mean Squared Error of 2.77%. The model was tested on Colombian individuals diagnosed with microtia, revealing a robust association between the prevalence of the Native American population component and their conditions.