Simulation-Optimization (SO) has become a reliable framework to optimize complex systems via stochastic simulation. However, recent SO methods face the significant challenges in balancing optimality and efficiency due to expensive evaluation in handling output noise. This paper proposes the SO with Adaptive Model Exchange (SO-AME) where metaheuristics dynamically switches between rigorous simulation model and fast-to-evaluate Machine Learning (ML) model based on prediction performance. The intention is to perform a vast majority of its exploration and exploitation using cheap ML, reserving expensive simulation runs only for high-performing candidates. Methodologically, ML model learns the complex input-output relation of simulation adaptively within Simheuristics loop. A case study is given to design pharmaceutical production. Compared with other robust SO methods, results validate that SO-AME significantly reduces the computational burden yet improves the optimality, offering a scalable AI-driven solution for optimizing complex systems under uncertainty.

Recent advances in AI have enabled agentic systems to act in operational environments with limited human oversight, raising the risk that they may violate or ignore human instructions and cause damage. The use of a digital twin as an intermediate environment can mitigate this risk by having actions examined before they reach the real system. While digital twin has been examined as environments for AI simulation more broadly, this study focuses on agentic AI and the use of DT as an intermediate environment for examining actions before real world execution. This paper develops a review informed conceptual framework that characterizes digital twin as testing grounds for agentic AI. By examining studies that integrate digital twin with agentic systems, it organizes the role of the digital twin into several stages, connected by pathways. Human role remains vital and takes different shapes around directing, checking, authorizing, and maintaining agentic systems.

Validating automotive software systems produces large multivariate test recordings that are still examined through effort-intensive manual review and rule-based evaluation, which detects faults beyond predefined rules poorly. Machine and deep learning have advanced fault diagnosis, yet most supervised models require large labelled datasets, generalise poorly to unseen conditions, and offer little insight into their decisions. We propose a generalisable and explainable two-phase framework for fault detection and classification during real-time validation. A denoising autoencoder trained only on healthy signals first flags abnormal behaviour through reconstruction-error analysis, removing the need for fault labels. Each abnormal window is then encoded as compact textual statistical evidence relative to a time-aligned healthy reference and classified by a frozen large language model under zero-shot and few-shot prompting, returning the predicted class, ranked alternatives, a confidence value, the fault location, and a short evidence-based explanation. Eight open-source models are evaluated across two powertrains and three driving regimes. The detector attains average F1-scores of 0.97 across powertrains and 0.98 across regimes, with average mean error below 0.03. Zero-shot prompting proves insufficient (best 0.519 F1-score), whereas few-shot prompting reaches perfect discrimination under stable regimes, showing that prompting strategy, rather than parameter count, governs classification quality: a nine-billion-parameter model surpasses every zero-shot medium and large model. Mistral Small 24B is adopted as the main pipeline model for its balance of accuracy, class balanced reliability, calibration, and inference cost, giving engineers interpretable diagnostic reports and more efficient validation.

The manufacturing sector is a key driver of global economic growth, yet it substantially contributes to resource consumption, waste generation, and environmental degradation. Conventional quality control methods in manufacturing primarily focus on statistical and economic outcomes, with limited attention to environmental implications, particularly for attribute data. This research proposes an economic-environmental design of the cumulative sum (CUSUM) chart for attribute-based quality monitoring. The proposed model optimizes charting parameters while maintaining acceptable false-alarm and inspection rates and enhancing the detection of nonconforming products. The model is validated through a case study from the plastic extrusion blow molding industry. The findings reveal that the optimal CUSUM chart provides the most effective monitoring scheme, with an economic-environmental cost reduction of 35% compared to the traditional CUSUM chart and 13% compared to the optimal np chart. The optimized CUSUM chart enhances quality control and reduces costs and emissions.