Isla, Malena Méndez; Kozlowski, Diego; Larivière, Vincent
Artificial intelligence (AI) is reshaping scientific practice. While recent advances in large language models (LLMs) are substantial, machine learning (ML) has a long history of developments across science. This study analyses the evolution of machine learning in global science from 1990 to 2025, by examining 4.9 million articles from OpenAlex. Using a hierarchical taxonomy of 252 ML techniques and architectures, we found that while deep learning exploded post-2015 and LLMs after 2022, established methods, such as linear and logistic regressions, declined in their share of the total scientific output. Utilizing state-of-the-art embeddings (mE5) and dimensionality reduction (UMAP), we constructed the ML Semantic Space that reveals a core-periphery structure dominated by the Physical Sciences. Computer Science acts as the primary methodological anchor, while Engineering drives applied developments. Health Sciences, led by Medicine, represents the second-largest expansion zone of ML techniques.