Crop variety selection is one of the most important factors influencing on-farm yields. Identifying suitable varieties for farms can be difficult as the relative performance of varieties often varies across environments due to genotype by environment interactions. This research seeks to address this issue by improving the accuracy of variety recommendations using machine learning and single nucleotide polymorphism (SNP) data to better capture the genotype by environment interaction effect. We implement Bayesian additive regression trees (BART) to analyze 13 years of variety trials from across Ontario. We find that the BART model is able to consistently provide significantly better variety recommendations relative to the mixed effects models commonly used in variety trials. This improvement in accuracy was in part due to the BART model's ability to capture SNP-SNP interactions and nonlinear SNP effects better than the mixed effects model.