Embodied artificial intelligence (AI) represents a paradigm shift in the design and construction of intelligent physical systems, where emphasis is placed on the integration of physical interaction and situational awareness of such systems to facilitate adaptive decision-making. This embodiment is realized through an interplay of sensing, control, and actuation, where AI not only interprets data but directly interacts with physical processes. Enabled by advances in deep learning (DL), sensing technologies, and computational infrastructure, embodied AI systems permit new avenues for a broad range of manufacturing applications by building upon a series of transformative capabilities that span semantic data inference, learning-based control, and generative optimization of actuation hardware design. This paper presents an overview of recent advances to enable embodied AI for manufacturing and outlines promising research directions.
Summary: One main conclusion in this paper is that model-based RL is much more data-efficient, but also more robust against neglected high-frequency dynamics. A modularized model-based RL strategy is therefore proposed where a nonlinear state-space model is estimated. In this model some minor physical knowledge can be easily introduced. Combining feedback and feedforward control with temporal optimization based on the estimated model, it is shown that energy and peak power for moving devices can be significantly reduced, utilizing much less data compared to standard model-free RL.
Over the past decades, the rise of a networked society has been driven by rapid advancements in information and communication technology, particularly in computer networking. This has enabled unprecedented hyper-connectivity among organizations, individuals, smart machines, and intelligent systems. As a result, new forms of collaboration have emerged, composed of distributed, autonomous, and heterogeneous entities. This evolution led to the establishment of Collaborative Networks (CNs) as a distinct discipline with a socio-technical character. Nowadays CNs play a key role in the ongoing digital transformation across industries and services. Although still a relatively young field, CNs have evolved through several generations over the past decades. As we move toward Industry 5.0, the complexity of interactions among a diverse range of agents continues to intensify. This article provides a brief overview of these trends, highlighting the role of CNs in the materialization of the goals of Industry 4.0 and Industry 5.0.
Cyber manufacturing services, which aim to connect geographically distributed designers and manufacturing service providers via the internet, are emerging to address the market shift from mass production to mass personalization. Recent advances in the Internet of Things (IoT) and machine learning enable new capabilities that promise improved efficiencies across the cyber manufacturing ecosystem. In this paper, we focus on machine learning methods that facilitate cyber manufacturing services in the areas of manufacturing process planning and design for manufacturing (DFM). To enable automated manufacturing process planning, we review recent advances in manufacturing capability modeling, manufacturing process selection, and feature recognition for process planning. To facilitate DFM, data-driven tools for generative design are reviewed and new methods and results presented. In the context of the literature review, we summarize work from our research group and present some new methods and results in the DFM area. Critical summaries of research challenges are provided to set the stage for recommendations on future research directions toward realizing cyber manufacturing services.
The energy-intensive automotive industry requires sophisticated energy management systems to improve energy efficiency. In automotive workshops, paint drying systems are a significant energy consumer, necessitating real-time monitoring and control to minimize energy waste and potentially prevent system malfunctions. Thus, this study proposed a novel real-time energy consumption anomaly detection and diagnosis methodology (eAnoD) for automotive paint drying systems to enhance their energy efficiency and operational safety. Specifically, an architecture combining a temporal convolutional network and graph attention network (TCN-GAT) was devised to extract spatiotemporal features from multidomain data, including energy consumption, equipment parameters, production states, and environmental conditions. A hybrid neural network combining a backpropagation neural network (BPNN) and variational autoencoder (VAE) was constructed to enable the prompt identification of energy consumption deviations. Furthermore, an anomaly grading method integrating combination weighting and cloud modeling techniques was developed to evaluate anomaly severity, facilitating targeted maintenance and proactive risk prevention. A real-world case study was conducted in a new-energy vehicle factory to validate the effectiveness and practicality of the proposed methodology and demonstrate its potential for energy saving and risk mitigation in automotive manufacturing. This study is expected to serve as a reference for practical implementation and generate new ideas for academic exploration.
The hallucination and black-box nature of large models limit their industrial applications. To address these challenges, a verification mechanism built on confidence intervals of Transformer-based output layers is proposed for trustworthy industrial large models (ILMs). Adopting a Vision Transformer (ViT), customized verification operations are incorporated to monitor the forward propagation process, and samples with probability distributions outside confidence intervals exit the network early and are handed over to technicians. Thus, the ViT is more interpretable because only samples within confidence intervals can propagate forward and be output from the ViT. Subsequently, an over-approximation approach is employed to obtain confidence intervals by linearizing the decision boundary of the ViT. The conservative decision boundary serves as the lower bound of confidence intervals, which can provide provable robustness for confidence intervals because the minimum probability of the ground truth is always higher than that of other samples. Finally, a certified training strategy is employed to enhance the robustness of the ViT. Data disturbances with Gaussian noise are generated using a randomized smoothing strategy to augment the data distribution. A smoothed loss function is used to strengthen the robustness of the ViT against data disturbances, thereby enabling greater confidence intervals. The proposed verification mechanism was validated on two public defect datasets. It achieved 99.98% precision for normal samples and approximately 95% precision for defective samples on a fabric defect dataset. It also achieved 99.21% precision and 99.15% F1 score on a wafer defect dataset. Comparative experiments with other Transformer-based models also demonstrated the generalization ability of the proposed verification mechanism.
Controlling machining deformations resulting from unbalanced stress fields inside structural components is a significant challenge in the manufacturing industry. Prediction of machining deformation fields is fundamental for deformation control and requires numerous iterations to optimize the machining process. Conventional prediction methods such as numerical analysis are tailored to a fixed geometry, making them time-consuming and inefficient for components with various geometries. In this study, a general data-driven model is proposed for predicting machining deformation fields in components with varying geometries and stress fields. This model is based on a geometry-oriented neural operator that incorporates global geometry information into the function space, modeling the relationship between the input function (stress fields) and the output function (deformation fields). Global geometric information is extracted using a graph neural network applied to a geometric graph and embedded into the input and output function space through an encoder-query framework. The proposed model achieved low root-mean-squared errors ranging from 0.001 to 0.016 mm, with maximum prediction errors between 0.003 and 0.047 mm across different types of components, including beams and frames. The main contribution of this research is the significant advancement in the application of neural operators to the development of general models for predicting machining deformation. The underlying principles of the proposed model provide an important reference for wider applications related to the control of machining deformation in the context of digital and intelligent manufacturing.
Additive manufacturing (AM) has been extensively used in various industries to produce complex-shaped parts, enabling the application of advanced design methodologies for lightweight and mechanically optimized structures. Reinforcement learning (RL) has emerged as a powerful tool for optimizing complex structural designs in various mechanical systems. In this study, an RL-based strategy is proposed to optimize the design of three-dimensional lattice structures. RL is employed to optimize the shape variables of each unit cell in a lattice with the aim of maximizing the mechanical stiffness while retaining a lightweight design. The research object is a body-centered cubic (BCC) lattice with strut geometries adjusted through RL optimization to enhance structural performance. The RL environment, which incorporates the state, reward, and action, is integrated into a finite element method simulation, in which the actions derive a set of optimal design variables through the learning process. The feasibility and effectiveness of the RL-generated designs are validated by fabricating optimized structures using a vat photopolymerization AM process and experimentally testing them under three-point bending. The results demonstrate that the RL-optimized lattice structures exhibit superior performance compared to traditional BCC lattice designs. This study highlights the potential of RL for design optimization and its broad applicability in various engineering systems that require complex mechanical components.
Rapid and accurate detection of surface defects has become critical with the increasing demand for highly reliable carbon-fiber composite plates (CFCPs) in advanced manufacturing. This study proposes a dual-stage enhancement framework to detect subtle defects in CFCPs, addressing the limitations of conventional down--sampling and feature--extraction methods, especially when samples are limited and defects are subtle. The framework highlights the critical role of expert knowledge in defect detection and allows effective parameter transfer between a task-specific super-resolution reconstruction module and a residual, multiscale fusion semantic segmentation network. Experiments on a digital--radiography CFCPs data set demonstrate that this method markedly amplifies weak defect signatures and pinpoints their locations with high fidelity. The findings exhibit significant gains in precision, recall, F1--score, and mean intersection—over--Union relative to U-Net, SegNet, and other baselines. In small--sample conditions, the proposed model nearly doubles the performance of the canonical U-Net. This framework offers a broadly applicable solution for automated micro--scale defect inspection across composite-material systems and other advanced--manufacturing contexts.
Recommending maintenance plans presents significant challenges due to the low standardization of maintenance records and unclear pathways for identifying appropriate plans. While knowledge graphs have been extensively researched for integrating and evolving maintenance data, these issues hinder the accurate recommendation of maintenance solutions within large-scale maintenance knowledge systems. This paper proposes a causality and equipment structure enhanced maintenance plan matching and recommendation (CEE-MPMR) method to address these challenges. The method leverages an unsupervised SimCSE model to normalize domain vocabulary in the absence of domain lexicon, and proposes a maintenance plan reasoning method based on RotatE. The proposed method achieves a maintenance plan matching accuracy of 90.80%, effectively improving the precision of maintenance plan recommendations. Finally, we applied and validated the approach on real-world data from a nuclear power enterprise and integrated the algorithm into a maintenance plan recommendation system, supporting intelligent analysis and decision-making for nuclear complex equipment maintenance.
In the digital transformation era of the fourth industrial revolution, prognostics and health management (PHM) is playing increasingly important roles in various engineering fields. As the complexity of industrial systems continues to increase, model-based or data-driven PHM technologies face growing challenges related to interpretability, generalization, and applicability, which limit the widespread deployment of PHM technologies. To address these challenges, this PHM-GPT, a large language model (LLM) specifically designed for PHM, is proposed in this paper. By leveraging LLMs, the PHM-GPT unifies anomaly detection, fault diagnosis, and maintenance decision-making tasks, enabling robust generalization across diverse datasets. In detail, a signal-to-text (S2T) pipeline is presented to develop the InsPHM-456k dataset, focusing on representative components such as bearings and gears. Furthermore, a novel framework for adapting general-purpose LLMs into PHM-specific LLMs is proposed through knowledge injection-based pretraining, PHM-specific instruction tuning, and downstream application fine-tuning during PHM. Additionally, an efficient architecture is introduced to incorporate low-rank adaptation adapters into the group attention module and the feedforward neural network. To validate the effectiveness of the PHM-GPT, extensive simulation studies are conducted, demonstrating its strong generalization across diverse datasets and its broad applicability to machinery components such as bearings and gears. Beyond automatically providing anomalies, diagnoses, and maintenance results, the PHM-GPT exhibits emergent abilities that have never been observed before, such as attribution reasoning, threshold setting, and knowledge discovery, which contribute to enhanced qualitative interpretability and deeper insight into system behaviors. Finally, this paper provides new insights into the PHM field and explores the future of LLMs in terms of advancing PHM technology deployment, accelerating the digital transformation process during the fourth industrial revolution.
Variations in product quality often originate from dynamic changes in the performance of manufacturing process systems. Developing digital twin models for model-based process control and optimization is a key strategy for improving quality; however, this approach is often constrained by the high cost of data acquisition in industrial environments. To address this challenge, this study proposes a quasi-static hypergraph neural network (QS-HGNN) model framework. Grounded in the assumption of microscopic quasi-static behavior and macroscopic evolution in process system performance, the proposed method abstracts process elements as nodes and uses hypergraph topology to represent complex multivariate relationships among process parameters. At the microscopic scale, static associations between nodes are quantified through discrete computation of a weight matrix, enabling the model to capture the system’s short-term performance characteristics. At the macroscopic scale, a long short-term memory network models the temporal evolution of these weights, thereby capturing the long-term evolution of the system. The framework integrates multimodal data fusion and physics-informed neural network constraints to enhance generalization in small-sample scenarios. The method is empirically validated through a case study on the press-fitting process of rubber bushings into track shoes for tracked vehicles. Comparative studies show that QS-HGNN achieves higher modeling accuracy than both static and dynamic hypergraph neural networks. Finally, the model is deployed within an intelligent digital twin system for real-time press-fitting quality prediction and process control. Experimental results demonstrated a substantial improvement in quality performance: the qualification rate increased from 70% to 100%, and process stability was significantly enhanced. This research provides a high-precision, low-data-cost pathway for digital twin modeling of manufacturing process systems with progressive performance evolution, offering a scalable foundation for intelligent process optimization and quality assurance.
Artificial intelligence (AI) plays a crucial role in autonomous driving (AD), advancing its development toward greater intelligence and efficiency. In response to persistent challenges in current AD algorithms, many researchers believe that large language models (LLMs), with their powerful reasoning capabilities and extensive knowledge, may offer promising solutions, enabling AD systems to achieve deeper understanding and more informed decision-making. Both industry and academia have actively explored the application of LLMs in AD tasks, showing early signs of progress in addressing issues such as the long-tail problem. To examine whether and how LLMs can enhance AD, this paper provides a comprehensive analysis of their potential applications, including their optimization strategies in both modular and end-to-end approaches, with a particular focus on how LLMs can address existing problems and challenges in current solutions. Furthermore, we explore an important question: Can LLM-based artificial general intelligence (AGI) serve as a key for achieving high-level AD? We also analyze the potential limitations and challenges LLMs may face in advancing AD technology and extend the discussion to societal considerations, including critical safety and security concerns. This survey aims to provide a foundational reference for cross-disciplinary researchers and help guide future research directions.
The rapid advances of large language models (LLMs) have presented industrial sectors with transformative opportunities for innovative designs beyond the capabilities of human designers. Due to the inherent black-box nature of LLMs, however, existing LLM-based design frameworks lack the explainability and robustness that human designers would otherwise offer. To address this issue, we propose an LLM-driven industrial design automation (LLM-IDA) framework to encompass conceptual design, knowledge-based design, and digital prototyping. The LLM-IDA framework utilizes a multi-tiered artificial intelligence (AI) agent group for four modular processes: ① a specification module, ② a quantification module, ③ a surrogate module, and ④ a prototype module. With every module autonomously executed by the agent group, LLM-IDA thus enhances the generation of novel designs with few-shot prompts and completes the design process with minimal human intervention. To validate this method, the LLM-IDA framework is tested and compared with a state-of-the-art retrieval augmented generation (RAG) method using the pass@10 metric. Ablation tests confirm the positive impact of each module on both design time and quality. Overall, the experimental results show that the LLM-IDA framework delivers performance superior to the RAG method in real-world applications.
Cognitive diagnosis is a fundamental task in online education and serves as the basis for subsequent educational tasks, focusing on measuring students’ mastery levels across various knowledge concepts. Existing methods mainly model the interactions between students and exercises, while largely neglecting the relations among the underlying knowledge concepts, thus failing to fully model the level of student mastery. Although some works propose to characterize the prerequisite relation among knowledge concepts, such modeling is still limited. To fill in the gap, in this work, we present a heterogeneous concept graph augmented cognitive diagnosis model (HCGCDM), a cognitive diagnosis method that strives to characterize and model the relationships among knowledge concepts with large language models (LLMs), hence benefiting the measurement of student mastery levels. First, we propose a heterogeneous concept graph construction module to automatically detect and mine the complex relations among knowledge concepts by using the power of LLMs. Specifically, we extract pairs of knowledge concepts with high similarity and then use retrieval-augmented generation to systematically detect and validate triples to construct a reliable heterogeneous concept graph. Subsequently, we develop a heterogeneous concept graph modeling and aggregation module that adaptively identifies important features and then integrates graph representations into knowledge concepts and exercises for more accurate assessment. We empirically evaluate our proposal on several tasks, and the results demonstrate that HCGCDM and its components are effective, surpassing the state-of-the-art methods.
The Earth observation satellite scheduling problem (EOSSP) is a complex optimization problem involving the selection, sequencing, and timing of observational tasks to maximize task completion rates while adhering to various constraints. However, effectively solving diverse scheduling scenarios with varying characteristics through manual design and modification of algorithm configurations is a laborious task. Concurrently, with the advent of large language models (LLMs), numerous studies have leveraged related technologies to facilitate automatic algorithm design (AAD), thereby alleviating the burdensome task of crafting optimization algorithms. However, these studies primarily concentrated on traditional optimization problems, whereas the EOSSP constitutes a more intricate problem necessitating precise articulation in specialized language and comprehensive elucidation. Moreover, previous research has achieved AAD by depending on the initially defined algorithms and proposing sophisticated, detailed iterative frameworks, thereby still requiring expertise in algorithm design. In this paper, we introduce an LLM-based multi-agent framework AgentAD that automatically generates efficient algorithms for EOSSP—transforming problem descriptions in natural language into executable algorithmic code. Within this framework, the proposed agents collaborate through numerous atomic interactions, completing subtasks in each defined workflow phase. Mirroring human behavior in software development processes, these agents join forces to craft effective algorithms tailored to EOSSP scenarios. The experimental results demonstrate that the algorithms generated by AgentAD outperformed other state-of-the-art human-designed counterparts.
Existing data-driven fire forecast systems often exhibit limitations in real-world emergency response scenarios, particularly with respect to efficient data reuse and vulnerability of sensor networks. This study proposes a smart agent that integrates an artificial intelligence (AI)-driven fire situational awareness engine with a large language model (LLM) to realize the diverse demands of emergency response in complex fire scenarios. First, a fire-resilient deep learning model based on ConvLSTM is developed to reconstruct building temperature fields using limited inputs from a partially failed temperature sensor network. The proposed architecture constructs spatiotemporal correlations between missing and survived sensor data, enabling the transformation of discrete temperature measurements into a continuous two-dimensional (2D) temperature contour. Subsequently, a smart agent powered by a domain-specific LLM is designed to enhance human-AI interaction during fire emergency response. A self-driven framework capable of automatically executing LLM-generated programs is established to deliver real-time, user-specific information to multiple stakeholders. Experimental results demonstrate that, compared with generic LLM-based responses, the proposed agent augmented with fire situational awareness can generate customized operational recommendations through dynamic interactions with the ConvLSTM-based fire model. This hybrid agent improves situational awareness and safety during fire emergencies, improves the resilience of fire services systems, and advances the practical implementation of AI-driven smart firefighting.
To advance sewage sludge (SS) valorization, this study proposed a novel approach that integrated protein recovery from SS with its conversion into a fully bio-based adhesive for plywood production. A thermal-alkaline pretreatment (pH = 12, 90 °C) followed by acidic precipitation enabled efficient recovery of sewage sludge protein (SSP). Compared to the commonly used sulfuric acid, citric acid partially co-precipitated with SSP, potentially promoting amidation reactions between amino and carboxyl groups during the curing process. This interaction contributed to the enhanced adhesive strength of SSP. Inspired by the catechol-mediated adhesion mechanism of mussel foot proteins, tannic acid (Tan) and Zn2+ ions were incorporated into the alkaline-modified SSP. Covalent and hydrogen bonding occurred through cross-linking between the polyphenolic moieties of Tan and the amide/carbonyl groups of SSP, while Zn2+ ions served as coordination centers, further strengthening interfacial cohesion. These procedures reassembled the fragmented peptide chains of SSP, resulting in enhanced hydrophobicity, thermal stability, and mold resistance. Under optimized curing conditions (140 °C, 8 min), the resulting adhesive achieved a wet shear strength of (1.09 ± 0.08) MPa, surpassing the Chinese National Standard (≥ 0.7 MPa) by 55.7%. X-ray micro-computed tomography revealed that 24.1% of the adhesive penetrated the wood micropores during curing, with mechanical interlocking complementing bulk adhesion to improve shear performance. Heat-induced structural reorganization further promoted β-sheet formation and esterification between SSP peptides and Tan, thereby reinforcing the adhesive’s cross-linked architecture. Finally, a preliminary assessment of the process’s economic viability and carbon neutrality potential highlighted its promise for sustainable, engineering-scale implementation.
The transition to sixth-generation (6G) wireless networks is expected to introduce increasingly complex network architectures, disruptive wireless technologies, ultra-high network density, and diverse service requirements, necessitating highly efficient algorithm design for large-scale and non-convex network optimization. However, conventional optimization-based algorithms usually require sophisticated mathematical modeling and exhibit high computational complexity, while classic learning-based algorithms often suffer from poor robustness and generalization, as well as a lack of cross-scenario meta-optimization capabilities. In contrast, given their strong reasoning and contextual understanding abilities, generative and large artificial intelligence (AI) models are emerging as promising technologies to overcome these limitations. In this article, we propose the leveraging of generative and large AI models for scalable and generalizable network optimization, with an emphasis on facilitating information compression, beamforming design, and automated optimization for dynamic wireless networks with limited radio resources. We introduce a diffusion-based generation framework to solve multi-objective optimization problems for efficient information compression and transmission. We also present a large AI model-based framework for solving non-convex continuous optimization problems for beamforming design in both cell-free wireless networks and integrated sensing and communication networks. Finally, we propose an innovative large AI model-based framework that can automatically solve mixed-integer nonlinear programming problems for microservice deployment over satellite networks.
Passive energy savings through the building envelope represent a critical strategy for reducing both energy consumption and carbon emissions. However, traditional technologies are limited by cumbersome control mechanisms and narrow adjustment scopes. To overcome these limitations, we propose a novel phase-change thermal diode with an enhanced unidirectional heat-transfer capacity. This thermal diode utilizes hydrophobic and hydrophilic materials, presenting dual benefits in material fabrication and structural applications. To verify the potential of this approach for building applications, the effect of different hydrophobic materials and vacuum on the performance of the thermal diode were compared experimentally. The optimally formulated material, which is readily manufacturable over large areas, demonstrated thermal rectification ranging from 8.72 to 23.62, thus offering an extensive adjustment range. For structural applications, the thermal diode could be combined with the building envelope to create a dynamic envelope for passive heat dissipation and insulation. Simulation studies confirmed that this novel dynamic adjustment method provides superior adjustment capabilities and achieves greater energy conservation than conventional dynamic methods. Specifically, cooling energy savings between 11.83% and 21.36% were attainable across various climate zones in China. This research fosters cross-innovation in the fields of buildings and materials, serving as a foundational reference for developing dynamic building envelopes.
The trade-off between affinity and specificity in molecular recognition elements (MREs), including antibodies, protein receptors, and aptamers, represents a well-documented challenge. Achieving MREs that combine high affinity with fine specificity is particularly difficult when targeting structurally similar analytes. To address this limitation, we propose a novel ligand- and receptor-based rational hapten design (LRRHD) strategy to regulate the generation of monoclonal antibodies (mAbs) exhibiting both high affinity and fine specificity, using sulfonamides (SAs) as model target analytes. This strategy enabled the identification of several novel haptens containing rigid spacer arms and led to the generation of mAb 10E6, which unexpectedly exhibited half-maximal inhibitory concentration values ranging from 0.23 to 20 μg∙L−1 across 29 tested SAs. The molecular recognition mechanisms underlying 10E6 binding were elucidated through crystal structure determination and molecular dynamics (MD) simulations. The results revealed that the extended complementarity determining region 3 of the heavy chain (CDRH3) of 10E6 forms a broader and more flexible ligand-binding pocket, allowing accommodation of diverse SAs with high affinity. In addition, mAb 10E6 was applied in an immunoassay, demonstrating limits of detection ranging from 0.29 to 4.7 μg∙kg−1 in skimmed milk and chicken samples. This study presents an effective rational hapten design strategy to mitigate the affinity/specificity trade-off and provides new insights into antibody discovery and vaccine development targeting small molecules.
Powdery mildew poses a major threat to global wheat production, highlighting the urgent need to identify resistance genes. In this study, we report the cloning of PmNCA6, a powdery mildew resistance gene originating from Triticum boeoticum. Using bulked segregant exome capture sequencing (BSE-Seq) and genetic mapping, we mapped PmNCA6 to a 17-Mb recombination-suppressed interval (680.1-697.1 Mb) on chromosome 7AL. By applying a mutant exome sequencing (MutExomeSeq) approach, we analyzed six ethyl methanesulfonate (EMS)-induced susceptible mutants and identified non-synonymous mutations in a nucleotide-binding leucine-rich repeat (NLR) gene, NLR1. Functional validation through barley stripe mosaic virus-induced gene silencing (BSMV-VIGS) and transgenic complementation confirmed that two alternatively spliced NLR1 transcripts (NLR1_V1 and NLR1_V2) confer resistance to powdery mildew. Phylogenetic analysis revealed that PmNCA6 is orthologous to the stem rust resistance gene Sr22a. Domain-swapping experiments between PmNCA6 and Sr22a demonstrated that the leucine-rich repeat (LRR) domain of PmNCA6 is critical for powdery mildew specificity. Field trials of near-isogenic lines (NILs) and recombinant inbred lines (RILs) indicated that PmNCA6-mediated resistance does not compromise yield performance. Screening of 553 Chinese wheat cultivars confirmed the absence of PmNCA6, emphasizing its potential for diversifying resistance sources in breeding programs. This study establishes MutExomeSeq as a robust tool for cloning genes in recombination-suppressed intervals and highlights the potential of engineering synthetic NLRs with tailored LRR domains to combat evolving pathogens.
Since 2022, global mpox outbreaks have resulted in 172 510 confirmed cases and 462 deaths as of October 31, 2025. Tecovirimat, a small-molecule therapeutic agent for orthopoxvirus infections (e.g., smallpox and mpox), is clinically limited owing to its poor solubility. A novel tecovirimat formulation was developed and characterized using scanning electron microscopy, X-ray powder diffraction, Fourier-transform infrared spectroscopy, and stability assessments. The antiviral activity of tecovirimat against orthopoxvirus was evaluated using cytopathic effect inhibition assays. Safety evaluations included: ① active systemic anaphylaxis and vascular irritation tests in guinea pigs and rabbits, respectively, evaluated using scoring systems and histopathological examinations; ② visual assessment of hemolytic activity in red blood cells; and ③ repeated-dose toxicity evaluation in cynomolgus monkeys (14-d administration followed by a 28-d recovery period). The novel formulation enhanced the aqueous solubility of tecovirimat to 50 mg∙mL-1. The lyophilized powder formulation 5 (LP5) exhibited exceptional stability under high-temperature, high-humidity, and photolytic conditions and maintained favorable physicochemical properties after 90 days of storage at 40 °C and 75% relative humidity (RH). Furthermore, safety assessments revealed no concerns regarding allergic reaction, irritation, hemolysis, or toxicity in repeated-dose studies. These findings demonstrate that the novel tecovirimat formulation is a stable, safe, and promising candidate for industrial development and clinical applications.
To improve patient stratification and therapeutic response prediction in computational pathology, clinical-grade decision-making has been enhanced by deep learning models, including those used for microsatellite instability (MSI) prediction and molecular subtype classification. However, prior models and training settings have been largely based on the natural image domain, which differs from histopathology data. To mitigate the inductive bias, we introduced surface parameterization, a geometric mapping from the surface to a suitable domain, to transform whole slide images into fixed-sized squares using conformal energy minimization (CEM) and stretch energy minimization (SEM) algorithms. These transformations are tissue-perceptive, enhancing the region of interest (e.g., cancerous areas) to improve model performance. For example, our method improved MSI prediction, achieving an area under the receiver operating characteristic curve (AUROC) of 0.87 for CEM and SEM with a reduced training set, compared with 0.70 for original slides. To validate its clinical applicability, we analyzed consensus molecular subtype (CMS) classification in 17 colorectal cancer (CRC) patients, with concordance rates of 47.1% (CEM) and 41.2% (SEM), outperforming the original slides (29.4%). As a proof-of-concept, we also linked CMS calls to organoid morphology, demonstrating that cystic organoids were more strongly associated with CMS3. This phenotypic feature may be integrated into CMS and used to improve the evaluation of tissue differentiation. Overall, our method provides new insight into the data processing of computational pathology and demonstrates the performance of state-of-the-art (SOTA) in multiomics prediction.