Existing electrochemical biosensing systems generally rely on external computing units for data analysis. The necessity of analog-to-digital conversion and data transfer leads to high complexity, low working efficiency and concern of privacy. Zhao et al. realize the in-sensor computing in electrochemical biosensing for the first time, which uses a single neuromorphic organic photoelectrochemical transistor to simultaneously perform multi-target biosensing and artificial neural network classification. The strategy permits in-situ training using real-time data, helping solving the issue of device-to device variations. The proof-of-concept application is demonstrated by direct classification of microRNA fingerprints in real biological samples, which provides a new paradigm for simpler, faster and more secure AI-driven electrochemical biosensing.

Researchers have developed BattFailScholar, a knowledge-augmented large language model framework that dramatically improves battery failure diagnosis. By equipping AI with a structured "case library" of over 11,000 battery failure examples, the system boosts diagnostic accuracy by 19.7% over standard LLMs, effectively handles rare failure types, and provides reliable risk assessments with confidence scores. Beyond diagnosis, it can infer missing links in failure chains with over 85% accuracy, offering a new tool for safer battery development.

With artificial intelligence (AI) workloads driving rising demands for computing power, bandwidth, and energy efficiency, integrated photonic neural networks (PNNs) offer a promising route beyond electronic bottlenecks. This review surveys emerging PNN technologies from devices and computing architectures to integration, packaging, and systems, highlighting optoelectronic co-design, key challenges, and their potential as next-generation AI accelerators.

Researchers developed DexAnyTwist, a learning framework that enables dexterous robotic hands to twist many everyday objects, from bottles and valves to screwdrivers and light bulbs. By identifying different modes and assigning expert policies to them, the system improved simulation performance and transferred zero-shot to real-world objects.

Researchers have updated the classic nonsolvent induced phase separation (NIPS) theory. By optimizing a fluid observation cell, the team decoupled the formation of macrovoids and cellular pores and revealed their distinct formation mechanisms. A new thermodynamic parameter was proposed for accurate pore prediction. The developed ultrathin membrane delivered outstanding performance in vanadium flow batteries, outperforming commercial counterparts.