All articles tagged with #machine learning
Physicists at Emory used a physics-tailored neural network trained on 3D dusty-plasma data to reveal precise, non-reciprocal inter-particle forces, correct key theoretical assumptions, and propose a universal AI-based framework for inferring physical laws in complex many-body systems.
Clouds remain the biggest uncertainty in climate projections, pushing researchers to blend physics with AI. Projects like CLIMA and ACE2 train neural networks on real atmospheric data and high‑resolution cloud simulations to emulate cloud effects more accurately and, in some cases, to forecast with far less computational cost than traditional Navier‑Stokes–based models. CLIMA refines cloud parameters through large-eddy simulations to double model accuracy, while ACE2 uses data-driven forecasts to capture cloud-influenced dynamics and speed up predictions, sparking debate over long-term reliability and the best balance between data and physics.
Two research teams claim to have located the long-lost Soviet Luna 9 lunar lander—one via crowdsourced landscape matching to NASA’s LRO data, the other with a YOLO-ETA machine-learning search trained on Apollo sites. With Luna 9’s precise resting place still unknown after its 1966 landing, the dispute will hinge on higher-resolution images from India’s Chandrayaan-2 over the sites in March, potentially confirming—or debunking—each claim.
Nature reports that a surge of AI-generated, low-quality submissions—dubbed 'AI slop'—is flooding computer science journals and conferences, with ICML 2026 receiving over 24,000 papers and arXiv submissions up more than 50% since ChatGPT; some papers are AI-generated or contain fabrications, prompting arXiv and conference policy changes, expanded reviewer pools, and debates about moving to rolling-journal models to preserve research integrity.
An AI method called CellTransformer analyzed RNA profiles from 3.9 million cells in a mouse brain to delineate 25–1,300 cellular neighborhoods, producing maps that align with known brain structures while revealing novel subregions; researchers aim to apply the approach to humans and other organs to sharpen understanding of brain function in health and disease.
A European study found that blood biomarkers linked to early DNA repair and cellular stress, identified with machine learning, may reveal Parkinson’s long before motor symptoms, paving the way for inexpensive blood tests and earlier treatment—with clinical usage potentially within five years—though brain changes may not fully match blood signals and no cure exists.
Researchers released DinoTracker, a free app that uses unsupervised AI to compare dinosaur footprint silhouettes without relying on pre-labeled data. By analyzing eight footprint features across 2,000 unlabeled casts, the system groups prints in line with expert classifications about 90% of the time, helping palaeontologists test ideas about which animals made which prints. Notably, some Triassic–Jurassic tracks appear birdlike, suggesting birds may have deeper roots than previously thought, though the team cautions that substrate, foot motion, and other context must still be considered by experts.
UIUC's Stomata In-Sight integrates confocal microscopy, leaf gas exchange, and machine-learning analysis to observe stomatal openings and real-time gas transfer in maize, linking pore-scale dynamics to leaf-level water-use efficiency and photosynthesis.
Researchers built the most detailed continental US water table depth map to date (≈30 m) using a random-forest model trained on over a million observations, estimating total groundwater storage at about 306,500 km³ (uncertainty 291,850–316,720 km³). The high-resolution product captures local variability and shallow groundwater near streams and croplands, and shows that coarse-resolution datasets systematically underestimate accessible groundwater. The study emphasizes the importance of depth-aware estimates for drought planning, agriculture, and water security, and highlights the need for expanded groundwater observations and hybrid modeling. All data and code are publicly available.
Scientists at UCL used machine learning to analyze blood tests and brain scans from 634 MS patients, discovering two distinct subtypes of the disease with different progression patterns, which could lead to more personalized treatment approaches.
Computer scientist Yann LeCun emphasizes that true intelligence fundamentally involves learning, highlighting the importance of learning processes in understanding and developing artificial intelligence.
Researchers at Duke University have developed an AI framework that simplifies complex dynamic systems into understandable rules, aiding scientific discovery across various fields by identifying key variables and stable states, and extending the capabilities of traditional physics-based models.
Scientists used lab-grown brain organoids and machine learning to identify distinct electrical signatures associated with schizophrenia and bipolar disorder, potentially paving the way for more accurate diagnoses and personalized treatments.
The article argues for the development of AI models that truly understand the world, rather than just predict outcomes, emphasizing the importance of semantic comprehension in language models to be genuinely useful.
Yoshua Bengio, a pioneer in AI and machine learning, has become the first researcher to be cited over one million times on Google Scholar, highlighting his influential work on neural networks and the rapid growth of AI technology.