The potential of machine learning in classifying latin American water isotope values and its implication for forensic provenancing.

The United States-Mexico border has become one of the world's deadliest migration corridors, with hundreds of undocumented migrants perishing each year. This has resulted in humanitarian efforts to identify deceased migrants; however, the environmental context results in the rapid decomposition and scattering of remains, complicating traditional identification methods. Isotopic analysis can provide investigative leads about an individual's region-of-origin and travel history if accurate classifications or predictions can be made. Robust isotopic datasets and associated classification toolkits for Latin America are currently not readily available, limiting the utility of isotopic analysis in the border context. This study introduces IsoPredict, a prototype application developed as a proof of concept from interpolated data to evaluate and compare the performance of six artificial intelligence (AI) algorithms for isotopic classifications within Latin America. Among the tested algorithms, artificial neural networks demonstrated the highest performance, achieving an 83 % accuracy when tested on novel data - substantially outperforming other models. Random forest (77 %) and support vector machine (76 %) models also demonstrated strong performance and are favored over decision tree (69 %) and k-nearest neighbors (66 %) models for this application. The Gaussian Naïve Bayes model yielded the lowest accuracy (49 %), likely due to the specific implementation that is not representative of broader Bayesian approaches. These results demonstrate the potential of AI to enhance isotopic classification in forensic contexts. However, they also highlight the need for more robust and forensically relevant isotopic datasets across Latin America. The further development of these datasets could help refine classification models using advanced AI algorithms.