Approximate unlearning removes irrelevant information from vision-language models (VLMs) while preserving performance. However, current approaches are based on class unlearning, which excludes domain-specific recognition and is insufficient for practical applications. Researchers from Tokyo University of Science and National Institute of Advanced Industrial Science and Technology have proposed approximate domain unlearning as a novel approach to differentiate between domains. This innovation outperforms state-of-the-art alternative tuning techniques for VLMs, paving the way for practical and fine-grained unlearning.

The gerotor tooth profile is crucial for determining hydraulic system performance in automotive engineering. In a new development, researchers from Pusan National University have leveraged conditional generative adversarial networks for machine learning-driven gerotor profile synthesis and optimization. The novel approach has remarkably produced designs that outperform human efforts and lead to 32% more efficient hydraulic pumps, potentially revolutionizing the automotive industry.

Using sound to get objects to float works well if a single particle is levitated but it causes multiple particles to collapse into a clump in mid-air. Physicists at the Institute of Science and Technology Austria (ISTA) have now found a way to keep them apart using charge. Their findings, published in PNAS, could benefit materials science, robotics, and microengineering.

This review systematically examines the integration of machine learning (ML) and artificial intelligence (AI) in nanomedicine for cancer drug delivery. It demonstrates how ML algorithms—including support vector machines, neural networks, and deep learning models—are revolutionizing nanoparticle design, drug release prediction, and personalized therapy planning. The article outlines the complete ML workflow from data acquisition to model interpretation, compares key algorithms, and presents real-world case studies spanning multidrug carrier optimization and cancer diagnostics. While highlighting substantial preclinical advances, the authors identify critical barriers to clinical translation such as data heterogeneity, model opacity, and regulatory challenges. The review concludes with a forward-looking roadmap emphasizing data standardization, explainable AI, and clinical validation to bridge the gap between computational innovation and patient-ready nanomedicine.