Recent decades have witnessed a tremendous growth in the number of Earth observation satellites (EOSs), which presents a huge challenge for mission planning. For the EOSs with optical sensors particularly, the observation mission is significantly influenced by the uncertainty of cloud coverage, which has been identified as the most dominant factor for the invalidation of remote sensing images. To overcome this uncertainty, uncertainty programming methods, namely, chance constraint programming, stochastic expectation model, and robust optimization, are put forth. Despite their success, these approaches are limited in that: 1) they simplified the complex cloud coverage uncertainty, which may be different from the true cloud conditions, and 2) they did not take the true cloud information into consideration. Motivated by these recent trends towards big data of satellite cloud images and machine learning for spatiotemporal prediction, this paper explores a dynamic re-planning scheme for multiple EOSs based on cloud forecasting. Specifically, we propose a new approach mainly in three steps: first, proactive scheduling based on a chance constraint programming is implemented and uploaded via ground control; second, cloud forecasting can be continuously conducted relying on the predictive recurrent neural network and the latest satellite cloud image; third, mission re-planning can be conducted according to the initial schedule and relatively accurate cloud information. Simulation results show that the cloud forecasting method is effective, and the re-planning approach presents highly efficient and accurate scheduling results.

Understanding the number of sunspots is crucial for comprehending the Sun’s magnetic-activity cycle and its influence on space weather and the Earth. Recent advancements in machine learning have significantly improved the accuracy of time-series predictions, revealing a compelling approach for sunspot forecasts. Our work takes the pioneering work by proposing a hybrid forecasting approach that combines the Seasonal Autoregressive Integrated Moving Average (SARIMA) with machine-learning algorithms like Random Forest and Support Vector Machine, delivering high prediction accuracy. Despite its high accuracy, we highlight the need for caution in deploying machine-learning-based methods for sunspot-number prediction, demonstrated through a detailed case study with only three extra time stamps leading to a dramatic change. More specifically, when making a forecast of monthly averaged sunspot numbers from 2023–2043 based on data from 1749–2023, we found that the observations in June, July, and August 2023 have a significant impact on the forecast, particularly in the long term. Given the multiseasonal and nonstationary nature of the sunspot time series, we conclude that this kind of phenomenon cannot be simply captured by a pure data-driven model, which can be highly sensitive in the forecast in the long term, and requires a more comprehensive approach, possibly with a model that includes physics.

Despite its promising potential for artificial intelligence (AI) applications, current in-memory computing (IMC) technology faces a variety of challenges before mass production. One of the major challenges we face is the absence of efficient toolchains for deploying canonical networks on IMC chips. To address this issue, we propose a co-designed framework that integrates circuit, toolchain, and system elements specifically for IMC. More specifically, our framework consists of several key techniques to improve the key performance, including 1) an 8-bit hardware-friendly quantization-aware training (QAT) approach to quantify the deep learning network from floating-point data to fixed-point data; 2) a novel operator optimization technique to increase the computing precision when running the algorithm models on the IMC chips; and 3) an efficient mapping strategy based on the integer linear programming (ILP) approach to improve the computation resource utilization of the IMC array. We assess our method on our 40-nm embedded Flash-based IMC SoC chip with voice recognition, speech noise reduction, and person detection tasks. Our experimental results show an accuracy over 94.60% in a quiet environment and 87.27% in a white noise environment and a false recognition rate below 1 time per 24 h for voice recognition, a 21.53% improvement for the perceptual evaluation of speech quality (PESQ) for noise reduction, and a 97.80% accuracy in person detection.

For their emergent application in electric vehicles, the development of fast and accurate algorithms to monitor the health status of batteries and aid decision-making in relation to maintenance and replacement is now of paramount importance. Data-driven approaches are preferred due to the difficulties associated with defining valid models for system and parameter identification. In recent years, the use of features to enhance data-driven methods has become commonplace. Unless the data sets are from multiple batteries, however, such approaches cannot be used to predict more than one cycle ahead because the features are unavailable for future cycles, in the absence of different embedding strategies. In this paper, we propose a novel approach in which features are predicted for future cycles, enabling predictions of the state of health for an arbitrary number of cycles ahead, and, therefore, predictions for the end-of-life. This is achieved by using a data-driven approach to predict voltage and temperature curves for future cycles, from which important signatures of degradation can be extracted and even used directly for degradation predictions. The use of features is shown to enhance the state-of-health predictions. The approach we develop is capable of accurate predictions using a data set specific to the battery under consideration. This avoids the need for large multi-battery data sets, which are hampered by natural variations in the performance and degradation of batteries even from the same batch, compromising the prediction accuracy of approaches based on such data.

Automatic transistor sizing in circuit design continues to be a formidable challenge. Despite that Bayesian optimization (BO) has achieved significant success, it is circuit-specific, limiting the accumulation and transfer of design knowledge for broader applications. This paper proposes (1) efficient automatic kernel construction, (2) the first transfer learning across different circuits and technology nodes for BO, and (3) a selective transfer learning scheme to ensure only useful knowledge is utilized. These three novel components are integrated into BO with Multi-objective Acquisition Ensemble (MACE) to form Knowledge Alignment and Transfer Optimization (KATO) to deliver state-of-the-art performance: up to 2x simulation reduction and 1.2x design improvement over the baselines.