Geologically storing carbon is a key strategy for abating emissions from fossil fuels and durably removing carbon dioxide (CO2) from the atmosphere1,2. However, the storage potential is not unlimited3,4. Here we establish a prudent planetary limit of around 1,460 (1,290-2,710) Gt of CO2 storage through a risk-based, spatially explicit analysis of carbon storage in sedimentary basins. We show that only stringent near-term gross emissions reductions can lower the risk of breaching this limit before the year 2200. Fully using geologic storage for carbon removal caps the possible global temperature reduction to 0.7 °C (0.35-1.2 °C, including storage estimate and climate response uncertainty). The countries most robust to our risk assessment are current large-scale extractors of fossil resources. Treating carbon storage as a limited intergenerational resource has deep implications for national mitigation strategies and policy and requires making explicit decisions on priorities for storage use.

Physical neural networks (PNNs) are a class of neural-like networks that make use of analogue physical systems to perform computations. Although at present confined to small-scale laboratory demonstrations, PNNs could one day transform how artificial intelligence (AI) calculations are performed. Could we train AI models many orders of magnitude larger than present ones? Could we perform model inference locally and privately on edge devices? Research over the past few years has shown that the answer to these questions is probably "yes, with enough research". Because PNNs can make use of analogue physical computations more directly, flexibly and opportunistically than traditional computing hardware, they could change what is possible and practical for AI systems. To do this, however, will require notable progress, rethinking both how AI models work and how they are trained-primarily by considering the problems through the constraints of the underlying hardware physics. To train PNNs, backpropagation-based and backpropagation-free approaches are now being explored. These methods have various trade-offs and, so far, no method has been shown to scale to large models with the same performance as the backpropagation algorithm widely used in deep learning today. However, this challenge has been rapidly changing and a diverse ecosystem of training techniques provides clues for how PNNs may one day be used to create both more efficient and larger-scale realizations of present-scale AI models.

Steel truss bridges are constructed by connecting many different types of bars (components) to form a load-bearing structural system. Several disastrous collapses of this type of bridge have occurred as a result of initial component failure(s) propagating to the rest of the structure1-3. Despite the prevalence and importance of these structures, it is still unclear why initial component failures propagate disproportionately in some bridges but barely affect functionality in others4-7. Here we uncover and characterize the fundamental secondary resistance mechanisms that allow steel truss bridges to withstand the initial failure of any main component. These mechanisms differ substantially from the primary resistance mechanisms considered during the design of (undamaged) bridges. After testing a scaled-down specimen of a real bridge and using validated numerical models to simulate the failure of all main bridge components, we show how secondary resistance mechanisms interact to redistribute the loads supported by failed components to other parts of the structure. By studying the evolution of these mechanisms under increasing loads up to global failure, we are able to describe the conditions that enable their effective development. These findings can be used to enhance present bridge design and maintenance strategies, ultimately leading to safer transport networks.