Integrating single-cell and spatial multi-omics data is essential for resolving cellular identity, regulatory programs and tissue organization, yet remains challenging under realistic experimental designs. In practice, data are often incomplete, unpaired and partially overlapping across batches and platforms, resulting in mosaic settings where missingness, batch effects and limited correspondence are tightly coupled. Existing methods typically rely on shared cells, explicit anchors or post hoc mapping, and can become unstable as overlap diminishes or spatial structure is not explicitly modeled. Here we present GAMMI (Graph-guided Adversarial Mosaic Multi-omics Integration), a unified graph learning framework for mosaic integration of heterogeneous single-cell and spatial multi-omics data. Rather than performing direct cell or sample alignment, GAMMI learns biologically meaningful relational structure by jointly embedding cells and molecular features in a shared latent space using heterogeneous graphs that encode cell-feature, feature-feature and spatial adjacency relationships. An edge-based contrastive objective with missingness-aware negative sampling avoids false-negative supervision under unobserved modalities, while adversarial domain adaptation suppresses batch-associated variation at the embedding level. Spatial data are incorporated as structurally informative constraints during learning, enabling systematic enrichment of molecular representations across spatial locations. Across diverse mosaic single-cell benchmarks and spatial datasets, GAMMI consistently outperforms state-of-the-art methods in biological conservation, batch correction and spatial reconstruction, including in low-overlap and fully unpaired regimes.

In a mobile edge computing environment where mobile devices are powered by green energy, minimizing power consumption and latency under long-term MEC energy constraints faces several challenges: unpredictable tasks, keeping both the privacy of the mobile device’s battery and efficient task scheduling, stabilization of power consumption and latency accumulation, and non-convex optimization for a continuous-discrete decision space. To tackle these challenges, we investigate the latency-aware resource-constrained scheduling (LARCS) problem in a hybrid-energy edge-cloud MEC system to minimize energy consumption and latency under long-term MEC energy constraints. For such a scenario, for the first time, we address protecting battery privacy and mitigating energy rebound peaks, which indicate sudden power surges from concurrent high-load tasks. We model the LARCS using mixed integer nonlinear programming (MINLP) and prove its NP-completeness. Then, it is reformulated as a Markov game. We adopt the multi-actor-attention-critic learning mechanism to preserve user privacy and employ a reward shape in reward functions for mitigating energy rebound peaks. On the above basis, we propose a latency-aware resource-constrained scheduling algorithm on the basis of multi-agent deep reinforcement learning (LARC-MADRL) to coordinate optimizations for multiple mobile users. This proposed algorithm requires no prior knowledge of uncertain parameters, operates independently of hybrid-energy dynamic models, and maintains user privacy. Numerical simulations validate that our proposed algorithm outperforms benchmark algorithms in minimizing power consumption, latency, and dropout rate while achieving long-term benefits, better convergence, improved stability, and enhanced scalability.

With the advancement of semiconductor manufacturing technology, multifunctional process modules (MPMs) have been widely adopted in cluster tools to enhance production flexibility and efficiency by handling multiple operations concurrently. However, the MPMs lead to challenges in scheduling the robot due to a variety of configurations, complex robot action sequences, and regulating wafer postprocessing residency time. For scheduling such a dual-arm cluster tool (DACT) with MPMs, this article proposes a specific reinforcement learning-based scheduling method. We first develop an adaptive algorithm to generate all feasible MPM configurations. We then employ the masking technique and prioritize a replay experience buffer to improve the dueling double deep$Q$-network (D3QN), enabling it to train and identify scheduling strategies that minimize makespan and reduce wafer residency time under each configuration. We conduct experiments to demonstrate that the proposed method ensures high productivity, providing a robust and flexible scheduling solution for cluster tools in semiconductor manufacturing, and significantly enhances overall production performance.

Cluster tools are vital in semiconductor manufacturing, where multifunctional process modules (MPMs) enhance flexibility and efficiency. However, variable MPMs and processing time in dual-arm cluster tools (DACTs) complicate scheduling, as variable MPM allocation patterns yield distinct productivity. This paper proposes a reinforcement learning-based method for DACTs with MPMs. Firstly, an algorithm enumerates all valid MPM allocation patterns. Then, an adaptive deep Q-Network (DQN) with masking techniques efficiently selects the most efficient pattern and generates robot schedules, minimizing makespan and wafer post-processing residency time across diverse DACT configurations. Experiments validate the proposed approach that offers robust, flexible scheduling solutions to boost semiconductor manufacturing productivity.

Multi-cluster tools are widely adopted in semiconductor manufacturing. When a process module (PM) at a step fails, a multi-cluster tool cannot complete the process recipe of work-in-process wafers and must be forced to enter a closedown process to be empty. To increase the throughput of a wafer fab, it is economically significant to shorten the failure-closedown process. However, due to the wafer residency time constraints, it is highly challenging to respond to a PM failure and find a corresponding optimal schedule. By assuming no parallel PM at a step, this work is the first to study this important issue for scheduling multi-cluster tools. We analyze the task sequences and synchronization conditions for robot activities to avoid deadlock in a shared buffer module. Upon these analyses, for process-dominant multi-cluster tools whose optimal steady-state schedule is known, algorithms are proposed to synthesize the proper sequences for robots in case of PM failures, then, a nonlinear program model is proposed to find an optimal schedule for the corresponding closedown process or decide no feasible solutions. The proposed model can uniformly deal with different scenarios of PM failures. Examples are given to illustrate the application of the proposed method.Note to Practitioners—In a wafer fab, there are hundreds, even thousands, of cluster tools. It is common that a failure of a processing module happens in a cluster tool. How to intelligently respond to such a random failure in a multi-cluster tool is an important issue in real-world production. This paper studies the scheduling problem of a multi-cluster tool in case of a failure at a PM. For a multi-cluster tool in case of a failure module, the proposed method can significantly reduce the loss of work-in-process wafers and shorten the closedown process.

Cluster tools are widely adopted in semiconductor manufacturing. When a process module fails, a cluster tool must undergo a close-down process, transiting to an idle state. Minimizing the makespan of a close-down process is economically significant in increasing the throughput of a wafer fab. However, the optimization of a close-down process is challenging due to wafer residency time constraints. Existing linear programming models must be constructed when specific processing time and robot activity time parameters are given. To address the issue, we propose a scheduling method based on deep reinforcement learning. For this reinforcement learning algorithm, the specific states, actions, and reward functions are designed for the close-down process to follow the Markov property, and a deep Q-network is adopted to find an optimal scheduling policy. Experimental results demonstrate that, in contrast to traditional scheduling methods, the proposed method achieves an optimal policy and also excels in generalization, enabling adaptive real-time production scheduling.

As data volume increases and data parallelism strengthens, cloud service providers must reduce energy consumption by using effective cloud scheduling methods. This paper investigates scheduling multiple servers in cloud data centers to achieve more balanced energy consumption over the long term while ensuring task completion. We propose a method based on long short-term memory (LSTM) and the African vulture optimization algorithm (AVOA), combining prediction and scheduling. The proposed method is divided into two main modules: the prediction phase and the scheduling phase. The purpose of the prediction is to reserve partial resources on servers for large tasks, thereby reducing energy consumption by minimizing the number of active servers. Therefore, we use a prediction method based on LSTM to predict upcoming resource demands and anticipate server requirements. In the scheduling phase, we adjust the priority of tasks to ensure task completion while maximizing the processing of tasks with larger demands. Experiments are performed to validate the application and effectiveness of the proposed method, which outperforms the benchmark algorithms.

The integration of data from multiple modalities generated by single-cell omics technologies is crucial for accurately identifying cell states. One challenge in comprehending multi-omics data resides in mosaic integration, in which different data modalities are profiled in different subsets of cells, as it requires simultaneous batch effect removal and modality alignment. Here, we develop Multi-omics Mosaic Auto-scaling Attention Variational Inference (mmAAVI), a scalable deep generative model for single-cell mosaic integration. Leveraging auto-scaling self-attention mechanisms, mmAAVI can map arbitrary combinations of omics to the common embedding space. If existing well-annotated cell states, the model can perform semisupervised learning to utilize existing these annotations. We validated the performance of mmAAVI and five other commonly used methods on four benchmark datasets, which vary in cell numbers, omics types, and missing patterns. mmAAVI consistently demonstrated its superiority. We also validated mmAAVI's ability for cell state knowledge transfer, achieving balanced accuracies of 0.82 and 0.97 with less 1% labeled cells between batches with completely different omics. The full package is available at https://github.com/luyiyun/mmAAVI.

In semiconductor manufacturing, multicluster tools are widely employed for many wafer fabrication processes. With the demand for high-mix integrated circuit chips and shrinkage of circuit width, a scheme in which multiple wafer types are fabricated inside multicluster tools is adopted by wafer foundries to make more profits. Multiple wafer types, multiple robots, and wafer residency time constraints make the resulting scheduling problems challenging. This work focuses on scheduling a single-arm multicluster tool to process two wafer types concurrently subject to wafer residency time constraints in which a conventional one-wafer cycle and backward strategy are not efficient. With such properties, several necessary and sufficient conditions are presented to check the feasibility of a periodic schedule. Polynomial-time-complex algorithms are proposed to examine a tool's schedulability and coordinate multiple robots to handle wafers for schedulable scenarios. The cycle time of an obtained schedule can reach the lower bound. A practical example is used to show the effectiveness of the proposed algorithm.

In today’s semiconductor manufacturing industry, wafer foundries often face the challenge of producing a variety of integrated circuit chip products using a single manufacturing line. To address this, multicluster tools have become a popular choice for processing multiple wafer types simultaneously. Operating such tools involves coordinating the robots in adjacent individual tools to transport multitype wafers through a shared buffer. This study aims to develop a scheduling method for the concurrent fabrication processes of two wafer types, performed by a multicluster tool with wafer residency time constraints. The proposed approach presents a two-backward sequence, based on a backward strategy of a single wafer type, to convert a one-wafer cyclic schedule into a one-wafer-per-type cyclic schedule while revealing its temporal properties. To ensure a smooth operation of a single-arm multicluster tool system and synchronize multiple robots, several necessary and sufficient conditions are derived for the first time. Two efficient algorithms are then proposed to determine the feasibility of a periodic schedule and obtain a schedule that achieves the lower-bound cycle time under a two-backward strategy, maximizing the productivity of such a multicluster tool. Finally, numerical simulations and two practical examples are presented to demonstrate the applications and performance of the proposed approach.

Mobile edge computing (MEC) is a promising way to improve the application performance on mobile devices in recent years, especially in terms of energy saving. However, users have to pay additional service costs such as communication costs when using services provided by MEC. In this paper, we manage to minimize both the average energy consumption and the average communication cost of mobile devices in a mobile edge computing environment with multiple users. We consider a MEC system that consists of multiple cloudlets and a cloud. Users submit their applications consisting of tasks with dependencies to this MEC system. We combine computation offloading and dynamic voltage frequency scaling (DVFS) technologies to reduce the power consumption of mobile devices. We formulate this problem into a mixed-integer nonlinear programming model and propose an adaptive multiobjective evolutionary algorithm based on Non-dominated Sorting Genetic Algorithm III (NSGA-III) to solve it. Simulation experiments are conducted to compare the proposed algorithm with three baseline algorithms and two multi-objective optimization algorithms, which validates the superiority of the proposed algorithms.

Current offline signature verification methods based on deep learning have achieved promising results, but these methods degrade greatly in cross-domain settings. An efficient offline signature verification model with both high performance and for deployment cross-domain without any adaptation. In this paper, we propose a novel approach to learning generalisable representations for offline signature verification. Firstly, we use the Siamese network combined with Triplet loss and Cross Entropy (CE) loss to learn discriminative features. Secondly, we introduce Instance Normalization (IN) into the network to cope with cross-domain discrepancies and propose an Inference Layer Normalization Neck (ILNNeck) module to further improve model generalization. We evalute the method on our self-collected Multilingual Signature dataset (MLSig) and three public datasets: BHSig-H, BHSig-B, and CEDAR. Results show that while our method achieves comparable results in single-domain setting, it is obviously superior to state-of-the-art methods in cross-domain setting.

As wafer circuit widths shrink down, stringent quality control is required during wafer fabrication processes. Therefore, a cleaning operation that removes chemical residuals inside a processing chamber is recently demanded in wafer fabs. To make a trade-off between quality and productivity, a condition-based chamber cleaning in practice is introduced to execute a cleaning operation with a known state of a chamber. Aiming to schedule a time-constrained single-arm cluster tool with a condition-based chamber cleaning operation, we present the necessary and sufficient conditions to check the schedulability of such a cluster tool. Efficient scheduling algorithms for a feasible schedule are derived. An example is given to show the application of the proposed approach.

Racetrack memory (RM) has high access performance comparable to SRAM. It is a kind of non-volatile memory (NVM), which consists of data block clusters (DBCs) and access ports. However, data accessing on RM is based on shift operations, which will decrease the performance of RM. This paper proposes techniques by using SRAM to reduce the shifts and improve the accessing performance of RM. The key idea is to place randomly accessed data on SRAM ahead of time to relax the data placement on RM. First, a greedy scheduling strategy is proposed to reduce the requirement of SRAM. Second, to further reduce shifts, data with similar association degree are grouped and allocated to each DBC. Experimental results show that the proposed techniques reduce the shifts by 72.3% with only 256-byte SRAM compared to pure RM.

Human age estimation from face image has drawn increasing research attention due to its many meaningful applications such as demographics analysis and surveillance monitoring. However, most existing methods directly extract age-specific features for age estimation and ignore age-related gender information. In this paper, we propose a simplified deep learning network for age estimation by simultaneously learning aging and potential gender features. Specifically, we first learn the potential gender information from face images. Then, we employ a two-stream convolutional neural network to simultaneously learn and concatenate the aging and gender latent appearance features. Third, we feed the multi-type features into a compact convolution network, named AgeNetwork, to further learn the age-specific features. Finally, we use a deep regression function to estimate the detailed ages. Extensive experimental results demonstrate the promising effectiveness and efficiency of our proposed method in comparison with state-of-the-arts.

The semiconductor manufacturing industry adopts multi-cluster tools as wafer fabrication equipment. If one of process modules fails, a multi-cluster tool cannot follow the normal schedule to complete the process recipe of work-in-process wafers and must empty these wafers by a closing down phase. It is economically paramount to shorten the close-down process with a process module failure (FCDP) and make no loss of wafers. However, due to the wafer residency time constraints, it is rather difficult to respond to a module failure and find a corresponding optimal schedule. By assuming no parallel modules at a step, this work studies this important issue for multi-cluster tools. For process-dominant multi-cluster tools whose optimal steady state schedule is known, we analyze the task sequences to avoid deadlock at the shared buffer modules. Algorithms are proposed to synthesize the proper sequences for robots in case of a process module failure, then, a nonlinear program model is derived to find an optimal schedule for the corresponding close-down process or decide no feasible solutions. An example is present to show the application of our proposed method.

As a kind of robotic cells, cluster tools are widely used for semiconductor wafer fabrication processes since they provide a reconfigurable and efficient environment. With recent advances in new semiconductor materials, the circuit line width has continuously being shrunk down, which brings new challenges for manufacturers. They require that a wafer should be moved away from a processing chamber as soon as possible after its processing is finished. To ensure high-quality integrated circuits in a wafer, its post-processing residency time must be minimized. It is also highly desirable to maximize the throughput of robotic cluster tools. This article aims at scheduling such tools with multiple objectives subject to wafer residency time constraints. To do so, new algorithms are proposed to calculate robot waiting time delicately upon the analysis of particular events of robot waiting for dual-blade robotic tools and optimally schedule such tools. The numerical results of industrial examples show that the proposed algorithms can provide an effective method to find schedules for dual-blade cluster tools such that multiple objectives are optimized.

Many to many (M2M) recommendation is one of the fundamental and important problems in commerce. With respect to this idea, profits, clients and products are inseparable. The traditional Top-N method cannot process the M2M recommendation problem. Therefore, this paper deals with the M2M recommendation as a many to many assignment problem via the group multi-role assignment (GMRA). Based on the concise formalization of Role-based collaboration (RBC) and its E-CARGO model, a successful approach using an (Extended Integer Linear Programming) x-ILP planning method and an improved greedy Top-N algorithm is proposed. These methods are verified by simulation experiments. Their results indicate the practicability of both solutions. As a comparison, the greedy Top-N method is faster than the x-ILP planning method via the PuLP linear planning package of Python. On the hand, the latter outperforms the former in recommendation accuracy.

With the interaction of discrete-event and continuous processes, it is challenging to schedule crude oil operations in a refinery. This paper studies the optimization problem of finding a detailed schedule to realize a given refining schedule. This is a multiobjective optimization problem with a combinatorial nature. Since the original problem cannot be directly solved by using heuristics and meta-heuristics, the problem is transformed into an assignment problem of charging tanks and distillers. Based on such a transformation, by analyzing the properties of the problem, this paper develops a chromosome that can describe a feasible schedule such that meta-heuristics can be applied. Then, it innovatively adopts an improved nondominated sorting genetic algorithm to solve the problem for the first time. An industrial case study is used to test the proposed solution method. The results show that the method makes a significant performance improvement and is applicable to real-life refinery scheduling problems.