A single transistor can mimic the neural and synaptic behaviours of the human brain, bringing biologically inspired computing closer to reality.

With the growing emphasis on sustainable development, the demand for environmentally friendly solvents in green chemical processes and carbon dioxide capture is increasing. Ionic liquids (ILs), as promising green solvents, offer significant potential but face considerable challenges, particularly in solvent selection. To overcome the limitations of traditional screening methods, machine learning (ML) techniques have recently been applied, offering a more efficient and data-driven approach. This review provides an overview of key ML methods used in solvent screening and compares them with traditional experimental and theoretical techniques. It examines the role of descriptor selection in structure—property-based methods, such as quantitative structure-activity relationships (QSAR) and quantitative structure—property relationships (QSPR), which are critical for predicting IL properties. The review also explores the application of these methods to screen IL properties, including toxicity, viscosity, density, and CO₂ solubility. Additionally, it discusses challenges in selecting appropriate models based on data scale and task complexity, integrating physical information for model interpretability, and achieving multi-objective optimization to balance key properties in ionic liquid (IL) design. Finally, it summarizes the achievements, limitations, and prospects of ML applications in ILs research, offering insights into how these methods can advance the development of sustainable ILs.

Human cognition relies on the seamless integration of multiple senses, allowing the brain to associate, infer, and even imagine across modalities. Replicating this capability in artificial systems has long remained a challenge, particularly under strict energy constraints. This study presents a bioinspired multisensory framework that integrates vision, touch, hearing, smell, and taste within a self-powered architecture. By enabling cross-modal association and adaptive reconfiguration, the system allows one sensory input—such as touch or sound—to trigger corresponding representations in other sensory domains. Beyond conventional recognition, the framework demonstrates higher-level cognitive functions, including inference and generative pattern creation. These advances point toward a new generation of intelligent machines capable of human-like perception and cognition.

Researchers have developed a soft measurement method based on a convolutional long short‑term memory (ConvLSTM) network that predicts product yield levels directly from real‑time image sequences of a microchannel reactor during gas‑liquid sulfonation. To overcome limited experimental data, a frame‑sampling spatio‑temporal augmentation strategy expands the training set. On the experimental data set, the augmented ConvLSTM model achieved an average accuracy of 97.44 %, outperforming the model without augmentation by 19.66 % and a conventional convolutional neural network by 9.94 %. This work provides a robust, non‑invasive tool for monitoring and optimizing complex micro‑chemical processes.

Researchers have developed a novel approach that integrates complete threedimensional molecular structures with traditional quantitative structureproperty relationship (QSPR) methods using deep learning. By combining molecular descriptors with chargedensity fields from density functional theory, a convolutional neural networkenhanced artificial neural network model significantly improves the prediction of critical temperature and critical pressure for 1359 organic compounds. The model achieves high accuracy (for T_c: R²=0.888, MAPE = 5.03 %; for p_c: R²=0.919, MAPE = 6.37 %), outperforming both conventional QSPR and the widely used JOBACK group contribution method.

When robots perform complex tasks, the pose accuracy of the end-effector is critical. However, errors from individual joints tend to accumulate along the kinematic chain, making it challenging to guarantee high pose precision at the end-effector. To address this issue, this study proposes a virtual-constraints-based end-effector pose compensator (VEPC). The method treats the actual angles of specific joints as known inputs and automatically adjusts the remaining joint angles in real time, effectively eliminating the pose errors of the end-effector caused by the joints. Experimental results demonstrate that the method can reduce the maximum end-effector position error by over 75%. Moreover, the method requires no additional sensors, offering low cost and high compatibility.