Abstract The development of scalable Digital Twins in manufacturing hinges on robust predictive models that can adapt to varying operational conditions. This paper investigates a critical component of this challenge: the application of transfer learning for tool wear prediction in milling. We train Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models using multiple datasets comprised of acoustic emission and vibration sensor data. Our objective is to transfer learned knowledge about tool wear across different machining processes and equipment.

The results reveal a significant gap between the promise and reality of transfer learning in this domain. While models achieve high accuracy within their source domain, their predictive performance collapses when transferred to a target domain with different machining parameters. This failure highlights a fundamental domain adaptation problem, where variations in materials, tools, and processes prevent the effective generalization of learned features.

We conclude that this poor transferability is a primary bottleneck for creating practical and cost-effective Digital Twins. The inability to deploy a single, adapted model across multiple assets necessitates resource-intensive, bespoke model development for each machine. This study underscores the urgent need for controlled experiments to understand and overcome these domain adaptation barriers, paving the way for truly scalable and reliable Digital Twin implementations in manufacturing.

Abstract With increasing market demands and growing competition, the importance of efficient facility layout has become more critical than ever for maximizing productivity and minimizing operational costs. This study investigates the application of metaheuristic algorithms to the facility layout problem, a complex and computationally challenging task in industrial and systems engineering. The primary objective is to determine an optimal or near-optimal arrangement of resources that enhances workflow efficiency and reduces material handling costs.

Due to the computational intractability of the problem, conventional optimization approaches often fall short in terms of scalability and solution quality. Metaheuristic algorithms offer a promising alternative by providing flexible and adaptable techniques that can effectively explore large solution spaces.

The study involves the development and implementation of metaheuristic optimization models, specifically using genetic algorithms and simulated annealing, to solve facility layout problems. These models are evaluated based on key performance indicators such as material handling costs, space utilization, and process flow efficiency.

Simulations conducted in Siemens Plant Simulation replicate various layout configurations and production scenarios to validate the optimization results. This integrated approach enables a detailed comparison of algorithm performance and provides practical recommendations for improving layout planning in industrial environments.

Abstract An increasingly dynamic market environment necessitates enhanced adaptability and flexibility in production systems. Effectively managing this complexity within simulation models requires the externalization of expert knowledge to accelerate model development and reduce manual effort. This paper explores the use of agentic Large Language Models in material flow simulation-based environments for autonomous problem-solving and decision-making. Today, Large Language Models agents can understand, plan, and execute complex tasks through interaction with experts, making them well-suited for conducting simulation studies. We describe the necessary architecture and examine their behaviour in complex scenarios, demonstrating how Large Language Models agents function as cognitive interfaces. A key application towards the industrial metaverse is the enhancement of digital twins of production systems with agentic Large Language Models, making them more adaptive and interactable. Our findings show that, with appropriate architecture, prompting and memory mechanisms, multi Large Language Models agent systems enable scalable, intelligent simulations for system investigation and optimization.

Abstract An increasing number of manufacturers are adopting on-site renewable energy systems to reduce energy costs, introducing new challenges for energy-efficient scheduling. In flexible job shop systems, the fluctuating availability of renewable energy makes energy supply decisions closely dependent on production schedules, increasing scheduling complexity. Traditional optimization methods face rapidly growing computational demands as problem size and complexity increase. This study proposes a quantum annealing-based method that integrates process and energy supply scheduling to minimize energy cost, energy consumption, and makespan. Results show that quantum annealing outperforms classical methods in both solution quality and computation time, demonstrating its potential for industrial application.

Abstract Laser-based Powder Bed Fusion (PBF-LB) is widely used for manufacturing complex, high-performance metallic parts through a layer-by-layer process. To ensure part quality and optimize processing parameters, PBF-LB relies on accurate thermal-structural simulations to analyse melt pool behaviour. However, traditional numerical simulation techniques, such as Finite Element Methods, are computationally expensive and time-consuming. Machine learning–based modelling offers a faster alternative but typically requires large, labelled datasets that are difficult to obtain. To address these limitations, a Physics-Informed Generative Adversarial Network (PIGAN) is trained to incorporate governing physical laws while generating realistic thermal data, improving prediction accuracy with limited data. The simulated thermal fields produced by PIGAN are then used to train an anomaly detection pipeline, enabling process parameter selection and early identification of defects in PBF-LB.