Many systems that are found in nature display complex, even chaotic, dynamics, yet often there is structure lying within the apparent complexity. Daniel Floryan and Michael Graham devised a method, combining mathematical theory of manifolds with neural networks, that can discover a system’s intrinsic structure, a low-dimensional manifold formed by intrinsic state variables, and learn to predict its dynamics from observed time series data. The cover image represents the low-dimensional manifolds hidden in data and the dynamics on them.

SeeFloryan and Graham

Osteoarthritis is the most common joint disorder, and it affects a substantial proportion of the population. Han et al. explore whether advanced machine learning techniques can predict future radiographs of knee joints. A deep learning model generated the most likely future radiograph and helped medical experts to predict the future course of osteoarthritis.

See Tianyu Han et al.

Immersive virtual reality (VR) and augmented reality (AR) require realistic haptic feedback. But tactile sensitivity varies on different parts of the hands of individual people, and current haptic interfaces are bulky. Yao et al. . develop a skin-integrated haptic interface to reproduce tactile sensations for immersive VR and AR applications. The haptic interface decodes tactile information that is associated with sensation threshold mapping on the hand, and enables haptic feedback of touched virtual objects.

See Kuanming Yao et al.

The outcome of organic reactions can be hard to predict without comprehensive knowledge of organic chemistry and known reactions. To speed up the development of new synthesis pathways (cover image), Chen and Jung use graph neural networks to extract a low number of general templates that can describe a large number of known organic reactions.

See Shuan Chen & Yousung Jung

Complex metabolic behaviour in cells can be captured with dynamic kinetic models. Such models are challenging to develop owing to the lack of knowledge about the characteristic kinetic parameter values that govern the cellular physiology of organisms. A new generative deep learning framework called REKINDLE has been developed by Subham Choudhury et al. to efficiently parameterize large-scale kinetic models, which helps to navigate the complex physiologies of various types of cellular organisms. Transfer learning in the low data regime allows REKINDLE to significantly expand the potential applications of kinetic modelling.

See Subham Choudhury et al.

The development of quantum hardware has reached a stage where meaningful quantum computing tasks are within reach, provided that the effects of noise can be mitigated. Most error mitigation methods require specific information about the noise channels that affect a quantum computation, the hardware implementation or the quantum algorithms themselves. Machine learning provides an alternative route to error mitigation, and Bennewitz et al. demonstrate a new technique that uses neural networks to mitigate errors in finding the quantum ground states of molecular Hamiltonians. The method is highlighted by the experimental preparation of the ground states of LiH at different bond lengths using IBM’s five-qubit chip, IBMQ-Rome.

See Elizabeth R. Bennewitz et al.

Tool use is one of the hallmarks of intelligence. Although robots can be programmed or trained to use a specific tool effectively, using a previously unknown tool is challenging. The robot shown on the cover uses a skill transfer approach developed by Keng Peng Tee et al. to manipulate the object on the table using a tool it has not seen or learnt about before. This skill transfer approach relies solely on experience that the robot has gathered from previous manipulation of objects with its own limbs, which is analogous to tool use in humans. Photos and videos of the Olivia III robot using and shaping previously unseen tools to manipulate objects can be found in the Article.

See Keng Peng Tee et al.

A challenge in robotics is to mimic biological tactile systems in which the perception of touch results from the integration of multiple mechanoreceptors. Massari et al. describe a large-area sensing skin for robotic system applications that detects external tactile stimuli in terms of contact location and magnitude. Such bio-inspired sensitive skin is crucial for effective human–robot collaboration in a variety of workspaces.

See Massari et al.

Understanding the relationship between brain function and behaviour is a goal of neuroscience and psychology. Automated behavioural analysis provides tools for investigating this goal, in combination with recordings of neuronal activity. Marks et al. present a deep learning pipeline that combines identification, tracking, pose estimation and behavioural classification for individual and social animal behaviour, using data only from simple mono-vision cameras in home-cage setups.

See Marks et al.

Studying the relation between the network structure of city roads and socioeconomic features can provide useful insights for urban planning. With the help of graph neural networks, Xue et al. propose a new metric, spatial homogeneity, to compare and analyse patterns of road networks for 30 global cities. They find that intra-city spatial homogeneity is highly associated with socioeconomic statuses such as GDP and population growth. Moreover, inter-city spatial homogeneity obtained by transferring the model across different cities reveals the inter-city similarity of urban network structures that originate in Europe, passed on to cities in the United States and Asia.

See Xue et al.

Learning walking gaits in unstructured environments is a challenging task for multi-legged robots such as the hexapods in the cover image. A modular approach for neural control by Thor et al . in this issue combines multiple primitive closed-loop controllers to allow rapid learning and adaptive behaviour, including pipe and wall climbing, as well as gaits to pass through high, low or narrow gaps.

See Thor et al.

Previous work has indicated that the brain may operate by predictive coding, but how such coding is implemented is unclear. In a paper in this issue, Luczak et al. present computational and electrophysiological data that a single neuron could be an elementary predictive unit from which a variety of predictive networks can be built. The cover image shows an artist’s rendering of how we learn, how individual neurons predict expected future activity, and how humans reflect on the functions of the brain.

See Luczak et al.