Background <p>Breeding programs prioritize the average performance of a genotype across environments and may overlook promising candidates for specific environments. To address this challenge, we propose a genomic prediction framework to select high-yielding genotypes tailored to individual environments.</p> Results <p>We compiled winter wheat grain yield data from 13,285 genotypes—6,766 lines and 6,519 hybrids—evaluated in yield plots at 31 central european sites from 2010 to 2022. With integrated genomic data, we show that only as the size of the training dataset increase, convolutional neural networks benchmark competitive to superior compared with traditional genomic best linear unbiased predictions (GBLUP) in predicting average genotype performance of lines. We then extend our prediction models to account for genotype times environment (G × E) interactions by incorporating information about the growth environment. We observe a 23% improvement in predicting environment-specific performance of new hybrids within a network of test environments with GBLUP based models. To better understand the environmental variables driving G × E interactions, we conduct analyses on a core set of 500 genetically diverse lines. Using machine learning, we successfully identify pivotal environment variables driving the clustering of study environments in central europe and highlight the benefit of modelling G × E interactions in selection of enviromically adapted varieties.</p> Conclusions <p>Our results suggest that big data in combination with machine learning and deep learning methods offers new ways to widen the genetic bottleneck often encountered when advancing candidates from early limited-environment to late stage multi-environment evaluations. This promises faster delivery of breeding progress to farmers’ fields.</p>