- KAIST researchers develop CURE, a technology that allows smartphone AI to retain and reuse knowledge from a server model, reducing server calls by an average of 55.61% compared with a non-cumulative approach while maintaining high accuracy

A data-knowledge dual-driven machine-intelligence framework enables experiment-efficient, full-chain synthesis of 2D dendrites. It achieves rapid process optimization within one week, constructs high-fidelity recipe-morphology mappings, and deciphers multifactorial growth mechanisms, requiring <70 experiments, which covers <1.5% of the total parameter combinations.

Optical neural networks capable of executing large-scale operations in parallel at the speed of light play an ever more prominent role in energy-efficient artificial intelligence. Towards this goal, scientists in China, inspired by the human brain, invented a backpropagation-free framework (B-ONN) for local learning. The technique will open new avenues for deploying physically robust, large-scale optical computing systems in real-world edge devices and green infrastructures.

A research team from the School of Physics at Renmin University of China has developed PhononBench, a large-scale phonon-based benchmark for evaluating the dynamical stability of AI-generated crystal materials. By combining the MatterSim-v1 universal machine-learning interatomic potential with Phonopy, the study enables high-throughput phonon-spectrum calculations for more than 130,000 AI-generated crystal structures. The benchmark systematically evaluates the dynamical stability of crystals generated by different AI models, identifies 32,995 MatterSim-labeled dynamically stable crystal structures under a strict phonon-frequency criterion, and provides a web-based platform for rapid phonon-spectrum calculation and dynamical-stability assessment.

Researchers at the Institute of Computing Technology, Chinese Academy of Sciences, have developed AtomFlow, a flow-matching model that predicts three-dimensional reactant structures from a product structure. It preserves the number and types of atoms throughout generation. Tested on small organic molecules with no more than seven heavy atoms in the Transition1x benchmark, its MatRIS version achieved 96.2% Top-10 accuracy, meaning that a matching reactant was found among the ten highest-ranked candidates in 96.2% of evaluated cases.

Working backwards from a target molecule to possible precursors, known as retrosynthesis, helps chemists plan syntheses for drug discovery and materials research. Many AI approaches represent molecules as SMILES strings and learn to translate products into reactants. These representations do not explicitly encode three-dimensional atomic coordinates, and sequence generation alone does not guarantee conservation of each element’s atom count. AtomFlow addresses this constraint by keeping the atoms fixed in identity while learning how their coordinates change.