UCL researchers built a hybrid where a 20-qubit quantum computer first identifies invariant statistical patterns in chaotic data, then feeds those patterns as structure into classical AI training. Tested on fluid dynamics and chaotic physical systems, the method delivers ~20% greater accuracy than standard AI while requiring hundreds of times less memory and maintaining stable long-horizon predictions. Published in Science Advances, with potential applications in climate modeling, transportation, and medicine.
Berkeley Lab Accelerator Assistant — an LLM multi-agent system — was the first to autonomously prepare and execute multi-stage physics experiments on a live synchrotron light source, cutting setup time by two orders of magnitude compared to expert manual scripting. Engineers issue natural-language goals; the system resolves variables, writes and runs analysis code, and safely controls accelerator hardware within operator-standard safety constraints. Published in Physical Review Research.
Tohoku University built a physics-grounded AI that predicts ionic dielectric tensors by first computing Born effective charges and phonon properties, then combining them via a physical formula — outperforming conventional black-box ML. Screening 8,000+ oxides, it identified 31 previously unknown high-dielectric candidates that could enable smaller, more efficient capacitors in smartphones and computers. Published in Physical Review X, April 2026.
University of New Hampshire scientists used AI to systematically mine scientific literature, building the public NEMAD database of 67,573 magnetic materials, and identified 25 previously unreported compounds that stay ferromagnetic above 180 degrees C — made from abundant iron, manganese, and nitrogen rather than scarce rare earths. The discovery directly targets the strategic supply bottleneck for EV motors and wind turbines, which currently depend on China-controlled rare-earth element exports. Published in Nature Communications, February 2026.
Rice University scientists demonstrated the first use of AI/ML for designing genetic circuits, showing that ML models trained on high-throughput data from the CLASSIC platform are significantly more accurate than traditional physics-based models at predicting how circuits behave in human embryonic kidney cells. The finding upends the assumption that explicit biological knowledge must be encoded — data-scale ML simply learns better predictors. The advance could dramatically accelerate the synthetic biology design-build-test cycle.
University of Cambridge researchers built a nanoelectronic memristor from modified hafnium oxide that combines memory and processing in one location — just like biological neurons — eliminating the energy-wasting data shuttling of conventional AI chips. The device switches at currents roughly one million times lower than standard oxide-based memristors and supports spike-timing dependent plasticity, the biological learning rule used by real brains. The main obstacle to production is a manufacturing temperature of ~700 degrees C, which exceeds standard semiconductor fabrication thresholds.
Oratomic (launched March 2026, spun out of Caltech) used OpenEvolve — an open-source LLM-powered optimizer — to iteratively evolve quantum algorithms the way natural selection works, discovering that a single logical qubit can be encoded with just three atoms instead of the previous 100 to 1,000. The resulting breakthrough means a utility-scale fault-tolerant quantum computer may require only ~10,000 physical qubits, versus earlier estimates in the millions. The AI combined knowledge across niche quantum sub-disciplines in a way no single human expert would have explored.
Google partnered with Georgia Tech and the Wild Dolphin Project to train DolphinGemma, a ~400M-parameter model that learned the hidden structure of Atlantic spotted dolphin vocalizations from four decades of underwater recordings, distinguishing signature whistles, burst-pulse squawks, and courtship clicks. The model — small enough to run on a Pixel phone in the field — can generate novel dolphin-like sound sequences, and researchers are now deploying it toward a shared vocabulary for two-way interaction with free-ranging dolphins. Google plans to open-source DolphinGemma for adaptation to other cetacean species.
Sony AI Ace became the first known autonomous robot system to compete at an elite human level in a widely played competitive sport, defeating 13 out of 29 amateur and professional human players in real matches and winning rallies against world-ranked professionals. Unlike game-playing AIs, Ace must contend with physical unpredictability — ball spin, bounce variance, real-time opponent reads — making this a qualitatively harder class of problem than board games. The project demonstrates that physical-world AI can now close the gap with top human athletes.
Researchers at Emory University used a physics-tailored neural network to analyze dusty plasma (the fourth state of matter) and discovered non-reciprocal particle forces with over 99% accuracy. The AI overturned a decades-old assumption: particle size DOES affect how quickly inter-particle forces weaken, and a leading particle attracts the trailing one while the trailing particle always repels the leading one. The framework runs on a desktop and may generalize to biological many-body systems.
Sandia National Laboratories scientists published a paper in Nature Machine Intelligence showing neuromorphic chips can efficiently solve partial differential equations (PDEs) — math previously reserved for traditional supercomputers. For 12 years after the cortical network model was introduced, no one noticed its non-obvious link to PDEs; the new algorithm exploits that link and opens the path to the world's first neuromorphic supercomputer. This directly challenges the assumption that neuromorphic hardware is only good at pattern recognition.