Today's Tech News: Notable or Not?

https://techxplore.com/news/2026-05-optical-chip-static-power-enabling.html "chips require constant electricity, overcome using silicon optical circuits integrated with phase-change materials where electrical pulse changes state from crystalline to amorphous (staying that way without additional power), altering light movement... near 0 static power consumption... large neural networks can be hard-wired into chip's light paths... combines electronics' programmability (logic/ control) with photonics' energy efficiency... optical switching, edge computing, autonomous vehicles, drones"

https://www.eurekalert.org/news-releases/1125685 "tiny chips made by machines that have to precisely stretch when electricity applied, but sluggish/ clumsy when move back/ forth quickly since remember where they were second ago... overcome using neural network watching machine movement making tiny adjustments instantly... 2 different computer parts working together: field-programmable gate array/ CPU... solves hysteresis in piezoelectric materials/ computational lag in real-time control... molecular level 3D printing/ etching, faster chips"

https://www.eurekalert.org/news-releases/1124871 "more data modes processed require more physical layers/ components adding bulk/ losing light energy passing through each component/ scaling difficult... overcome using high-dimensional unitary transformations with just 3 programmable layers... commercially available liquid-crystal spatial light modulator pixelated arrays: each cell independently controlled via software changing light phase... couple single input to up to 7K outputs...  managing quantum computers' high-dimensional states"

https://www.eurekalert.org/news-releases/1123850 "assumption that connected data points share similar traits (homophily), but relationships like chemical bond/ predator-prey exhibit heterophily where connected nodes differ, often causing graph neural network failure/ accuracy loss... overcome allowing AI to better understand individual data points before being influenced by neighbors, enabling simultaneous homophily/ heterophily processing... better predictions: fraud detection, physics modeling, drug discovery, social network analysis, weather forecasting"

https://techxplore.com/news/2026-04-method-neural-networks-faster-propagation.html "struggles with wave propagation because wave equations mathematically stiff, and standard AI models require many parameters/ data points, so extremely slow heavy, often requiring days of supercomputer time... overcome embedding phase/ amplitude physics into network's architecture... predicts wave variation rather than entire wave itself... coarse wave path model/ lets neural network fill in complex local interference patterns... energy exploration, medical Imaging, telecommunications"

https://www.eurekalert.org/news-releases/1121069 "holographic storage usually based only on light's intensity (amplitude), ignoring its phase/ polarization, making storage/ retrieval incredibly complex... overcome using convolutional neural network solving decoding by learning to identify phase/ polarization signatures within simple diffraction intensity images, allowing all 3 light properties to be reconstructed simultaneously... significantly greater data storage density compared to traditional surface-level storage like hard drives or DVDs"

https://techxplore.com/news/2026-02-atom-thin-ferroelectric-transistor-polarization.html "2 orders magnitude more memory states for more granular synaptic connection strengths in neuromorphic AI... transistor few atoms thick, stable >100K seconds even at room temperature, 93% accurate running image recognition AI algorithm, durable, fast switching... energy efficient, non-volatile... mechanical exfoliation, then dry-transfer stacking without solvents... applying DC voltage pulses/ gate voltages for on-demand modulation within the moiré potential. scale to wafer-size production" Related: Strong correlations and superconductivity observed in a supermoiré lattice https://phys.org/news/2026-02-strong-superconductivity-supermoir-lattice.html

https://www.eurekalert.org/news-releases/1115475 "brightness not negative so forced to either use electronic components to handle negative values (causing delays) or limit AI to simpler, less stable non-negative math... overcome using 2 microring resonators at different wavelengths representing positive/ negative values... cascadable nonlinear activation, Mach–Zehnder interferometer mesh... modeled generator that uses natural optical noise as an input, 98% accurate Iris classification... optical-to-optical image generation without need for digital-to-analog conversion"

https://www.eurekalert.org/news-releases/1110024 "mimics neuron Compute-in-Memory: digital logic more accurate but uses more energy than analog... versatile platforms switching between traditional/ neuro-inspired networks based on application... only communicates (sends spike) when electrical charge (membrane potential) reaches threshold... responds to irregular/ occasional events (sparse signals) enabling AI to operate onboard device (like drone), outside large data centers/ cloud... event-Based cameras send data only when pixel change detected" Related: Single coin-battery runs AI, from multi-sensor fusion to high-fidelity voice/ vision, for years https://mailchi.mp/allaboutcircuits/experience-ai-on-a-coin-cell-at-ces-8871931?e=197a4256c6

https://www.eurekalert.org/news-releases/1108469 "computes while data in transit over optical fibers... speed: tera- to peta-scale operations/ second, energy consumption: femtojoules to attojoules/ operation, latency: pico- to nanoseconds... remote access to optical processor without digital conversion drastically reduces computing latency... performs machine learning computations within optical fiber... proof-of-concept combining computation with data transfer over 20 km fiber access line achieved good classification accuracy for image recognition tasks"

https://techxplore.com/news/2025-12-previously-untrainable-neural-networks-effectively.html "network encouraged to match internal representations of guide network during brief training... transfers structural knowledge, aligning way 2 networks organize information within each layer, even when guide network untrained, as already encode valuable architectural biases steering other networks toward effective learning... even few alignment steps before full training sufficient to confer lasting benefits... stabilizes training, eliminates failure modes, prevents overfitting... success less dependent on task-specific data" Related: For visual AI systems, selecting the right blueprint may accelerate learning https://www.eurekalert.org/news-releases/1108079 Taming chaos in neural networks: A biologically plausible way https://techxplore.com/news/2025-12-chaos-neural-networks-biologically-plausible.html

https://phys.org/news/2025-11-method-based-quantum-processors-neural.html "process information more efficiently than classical Convolutional Neural Networks, but quantum photonic states linear, limiting flexible operations needed for neural networks... overcome injecting adaptive state making photonic circuit more adaptable, so circuit adjusts behavior based on a measurement taken during processing... either injects new photon or sends existing light forward steering computation... proof of concept: encoded 4 × 4 images of horizontal/ vertical bar patterns, >92% classification accuracy, scalable"

https://www.eurekalert.org/news-releases/1104774 "integrate data storage/ signal processing... highly uniform, scalable, using precise method controlling Schottky (contact) barriers achieved exposing selected areas to oxygen, allowing precise/ predictable electricity flow... high resistive switching ratio, switches clearly between on/ off, <6.8% variation between devices, 100% fabrication yield, picture classification accuracy >98%... fabrication strategy also applicable to other materials... AI accelerators when integrated with multi-layer stacking or hybrid CMOS-2D"

https://www.eurekalert.org/news-releases/1104295 "only final intensity or amplitude of target scene is known, not the crucial wavefront phase, so large datasets pre-computed holograms as labels unobtainable... also, trade-off between reconstruction quality/ computational speed... overcome by learning mapping between target scene/ its hologram... lightweight multi-wavelength network creates realistic depth perception by computationally simulating optical field... future: overcome full-color still needing 3 separate network models, which triples computational resources"

https://quantumzeitgeist.com/chinese-researchers-achieve-98-scalability-on-sunway-supercomputer-this-week/ "scaled quantum simulations to real molecular sizes using 37 million processor cores by training neural network to model electron arrangements/ movements within a molecule... SW26010-Pro chips including clustered cores highly effective for parallelizable, computationally demanding tasks... bridges AI/ quantum science: existing classical supercomputing resources accurately modeling complex quantum systems... materials and drug discovery without needing fully realized quantum computers"

https://phys.org/news/2025-10-rechargeable-powers-nanoscale-molecular-machines.html "room temperature circuits spend heat energy stored in kinetic traps, like molecular springs, recharged with heat pulse... reusability through kinetic traps isn't limited to heat, but for any energy source where can break weak bonds between molecules letting them fall back into their traps, such as light, salt, acid gradients... long term neural network and logic circuit tasks like those of living systems... smart medicines, cargo sorting, self-hydrating contact lenses, self-repairing strain cracks, molecular computers" Related: Light-guided 'optovolution' evolves proteins that switch states on schedule https://phys.org/news/2026-03-optovolution-evolves-proteins-states.html DNA molecular-scale computer smaller than 2nm semiconductors; high expectations for bio-computing https://www.eurekalert.org/news-releases/1125806

https://www.eurekalert.org/news-releases/1097680 "math operations performed through light interference mechanisms on mm-scale silicon microchips carried out faster/ using less energy, since no need to digitize information... step forward to make artificial intelligence (which relies on extremely energy-intensive data centers) more sustainable... addresses theme of training, phase in which network learns to perform certain tasks... also enables more sophisticated AI models integrating real-time data processing into portable devices... autonomous cars, intelligent sensors" Related: Advanced design for high‑performance and AI chips https://www.eurekalert.org/news-releases/1101475

https://techxplore.com/news/2025-07-neural-networks-enable-ai-memory.html "not by using more data or computational overload, but with geometry... self-tuning intelligent AI automatically adjusts focus as it recalls, speeding up its response; with fewer mistakes, single tuning parameter lets system balance between memory power/ accuracy... properties not hardcoded, but arise naturally from curved geometry... more adaptive, efficient, easier to understand, instead of black box models that are powerful but hard to explain... brain-inspired computing, neuroscience, robotics"

https://www.eurekalert.org/news-releases/1090114 "biocompatible polydopamine-modified barium titanate nanoparticles enable efficient photo-to-thermal conversion/ ferroelectrics... ferroelectric poly (vinylidene fluoride-co-trifluoroethylene) copolymer generates signals through reversible polarization changes... cellular-scale micropyramid arrays promote neuronal adhesion, neurite outgrowth, interconnection... performs up to 3 months after in vivo implantation... neural interface materials/ devices, adaptive brain-machine interfaces, biomedical/ tissue engineering" Related: Organic neuromorphic devices: Paving the way for next-generation computing and bioelectronics https://www.eurekalert.org/news-releases/1090581 Soft 3D transistors with hosting living cell potential https://www.eurekalert.org/news-releases/1112930

https://techxplore.com/news/2025-06-optical-neural-partial-differential-equations.html "current computational techniques time-consuming/ expensive, overcome by encoding equations in light wave's intensity property, then phase feeding them into diffractive optical neural networks/ optical matrix multipliers... as wave passes through, intensity gradually shifts/ changes until representing solution... takes the spatiotemporal data of an input physical quantity, which is a function of positions and time, to predict the spatiotemporal data of an output physical quantity as a function of positions and time" Related: Control and enhancement of optical nonlinearities in plasmonic semiconductor nanostructures for future reconfigurable optical neural networks https://www.eurekalert.org/news-releases/1087065