A growing surplus of refinery sulfur could find a new use in asphalt production, reducing reliance on petroleum-based bitumen and lowering costs. But a new academic review highlights unanswered questions about emissions, human exposure and long-term health implications, calling for safer formulations and life-cycle assessment of sulfur-modified asphalt.

Traditional air quality models struggle with changing emissions, while pure AI systems operate as opaque "black boxes" that can break the laws of physics. To solve this, researchers developed ProNet, a framework that embeds AI directly into atmospheric equations. Tested across 338 cities, ProNet outperformed experienced human forecasters and accurately predicted extreme events—like severe dust storms—without producing "AI hallucinations." Crucially, it also acts as a scientific diagnostic tool, using routine hourly data to uncover and correct hidden flaws in existing chemical models.

Researchers have developed a multimode band-engineering framework in one-dimensional grating-based photonic crystal waveguides designed to meet the growing demand for scalable and flexible filtering in mode-division and hybrid multiplexing systems. Published in PhotoniX Synergy, the work combines multimode photonic band engineering, scalable higher-order-mode filtering, and reusable filter building blocks to enable broadband manipulation of spatial modes, polarizations, and wavelengths, providing a unified route toward compact and multifunctional silicon photonic devices.

Deep learning is rapidly changing fluorescence microscopy by enabling researchers to extract more information from fewer, noisier and faster-acquired images. A new review in PhotoniX Life systematically summarizes recent advances across four major directions: imaging quality, imaging speed, imaging throughput, and intelligent analysis. This article systematically reviews the applications of deep learning across four key areas: imaging quality, imaging speed, imaging throughput, and intelligent analysis. Starting from mainstream neural network architectures, it presents their specific implementations in super-resolution reconstruction, image denoising, aberration correction, fast 3D imaging, cell segmentation, and virtual fluorescent labeling. It also highlights current critical challenges, including strong dependence on training data, limited cross-platform generalization, and the need for validation of biological plausibility in reconstructed results. The aim of this paper is to help researchers understand the limitations of algorithmic capabilities and guide them in making informed choices and evaluations when applying these methods in real experimental settings.