Abstract Global supply chains are increasingly vulnerable to disruptions caused by pandemics, natural disasters, and geopolitical instability. Traditional machine learning approaches often underperform when data are limited or when disruption patterns differ from historical cases. To address these challenges, this study develops a physics-informed machine learning (PIML) framework that embeds physical and operational constraints into the predictive process, including inventory balance, production capacity, and scheduling relationships. This ensures both predictive accuracy and physical consistency when estimating upstream production downtime. Based on these predictions, the study further introduces a capacity-sharing outsourcing algorithm that jointly considers the production schedules of multiple firms and evaluates the loss reduction of disruption-affected firms when selecting optimal products for delegated production. Simulation experiments show that the proposed framework reduces the losses of production-halted firms by more than 60% and significantly accelerates recovery through coordinated capacity sharing. This research advances the application of physics-informed artificial intelligence in supply chain disruption management and provides practical guidance for building resilient and sustainable industrial systems in line with Industry 5.0 and Sustainable Development Goal 9 (SDG9).
Engineering future worlds in a speculative design process requires creative ideas. Speculative design has ties to both engineering design and creative writing, and while artificial intelligence (AI) use has been found to impact the creativity of ideas in both domains, these fields have conflicting findings related to the impact of AI on idea novelty. Through a study of three groups of professional engineers and designers, we explore the relationships between dimensions of creativity and perceptions of creativity across different AI usage conditions during speculative design in a professional setting. Experts and designers assessed 120 ideas that were either fully human-generated, generated by humans using AI-as-tool, or fully AI-generated. In this context, AI use improved the well-craftedness and usefulness of ideas with no significant impact on novelty, although cosine similarity indicated that AI-generated ideas explored a smaller design space than the other conditions. Designers rated AI-generated ideas as less well-crafted and useful than experts, which may be indicative of ownership bias or "human favoritism" by the participants in this study. Additionally, designers who did not use AI displayed higher creative self-confidence, assessing their ideas as more well-crafted, novel, and useful compared to designers who used AI-as-tool. Overall findings indicate that AI use impacted various dimensions of creativity within this speculative design sprint, and while designer ratings often align with experts, perceptions of creativity can be biased by awareness of AI use. The inconsistent findings observed in both prior literature and our groups of designer groups regarding AI use and creativity highlight the need for continued investigation into how AI use affects design outcomes across different tasks and contexts.
The assembly of aero-engine accessory gearboxes is a critical process governed by complex nonlinear geometric error propagation. Achieving high-precision assembly is often hindered by the limitations of traditional analytical models in capturing contact deformations and the scarcity of labeled data for data-driven approaches. To address these challenges, this article proposes a physics-informed graph neural network for selective assembly (PI-GNN-SA). Moving beyond standard data-driven paradigms, our framework formally translates the classical small displacement torsor (SDT) kinematic theory into a differentiable topological regularizer. This mechanism enforces rigorous geometric loop closure constraints, enabling the network to learn physically consistent representations within a highly sparse data manifold, effectively preventing the severe overfitting common in purely data-driven models. The trained PI-GNN serves as a high-fidelity surrogate model within a nondominated sorting genetic algorithm II (NSGA-II) to optimize part matching. Experimental results based on real industrial data demonstrate that the proposed method reduces the root mean square error (RMSE) of backlash prediction by 43.5% compared to state-of-the-art graph networks. Furthermore, the optimized selective assembly strategy achieves a 96.5% one-pass qualification rate, significantly outperforming traditional methods. This work provides a robust, interpretable, and efficient solution for intelligent assembly in the aerospace industry.
Fault diagnosis of rolling bearings using imbalance data has always been a particularly challenging problem. The proposed digital twin technique, which complements the unbalanced dataset, provides a completely new approach that will greatly improve the effectiveness of the diagnosis. In this paper, A digital twin method based on feature embedding in a diffusion model to generate fault data with more accurate feature representation was proposed and applied to fault diagnosis. First, A bearing mechanism model is constructed from the dynamics to enable the generation of raw one-dimensional signals with fault characteristics. Second, diffusion models based on variational Autoencoders (VAE) are constructed, aiming at the application of diffusion models to low-dimensional signals to form a stable signal generative approach. Finally, A feature embedding method in a diffusion model is proposed to enable the model to be generated to include more accurate feature representations, thus enhancing the datasets and improving fault diagnosis performance. In addition, comparison experiments with digital twins based on Generative Adversarial Networks (GANs) show that the method is more accurate and robust.
Abstract Optimizing the energy consumption (EC) of industrial robots plays a crucial role in promoting their large-scale application to support the realization of Industry 4.0. Notably, robotic grinding is particularly energy-intensive, attributed to the complex coupling between robot dynamics and the continuous contact forces required for surface finishing. Furthermore, given the intricate nature of the EC process, it is challenging to establish a precise numerical relationship between robotic process parameters and EC relying exclusively on physical mechanisms. Data-driven models require large datasets and often fail to achieve high accuracy under small-sample conditions. To address this, we propose a physics-guided data-driven EC (EC-PGDD) prediction model. The heat conduction equation is employed to characterize the underlying relationship between temperature and EC, and a regularization term within the loss function of the temperature prediction model—by embedding differential operators—imposes a smoothness prior, enhancing extrapolation in sparse data regions. Based on the physical information provided by this equation, EC-PGDD integrates predicted temperature with process parameters, encoding richer information than purely data-driven approaches. To validate the effectiveness of the proposed model and provide targeted strategies for process parameter optimization, validation experiments on a battery end-plate robotic grinding unit, utilizing the Nondominated Sorting Genetic Algorithm II (NSGA-II) for multiobjective optimization, demonstrated a 7.67% reduction in energy consumption and a 0.23% reduction in execution time.
