https://www.mdu.se/

Background: Machine learning components are being deployed across nearly every business sector and their importance is continually growing. However, the engineering practices for building these systems remain poorly understood compared to those for conventional software systems. Objective: This work provides practical guidance to support architects in designing and implementing machine learning-intensive systems, and identifies areas where there are gaps in understanding and achievement. Method: Building on our prior research, we developed a checklist of quality concerns for architects of machine learning-intensive systems. This checklist was iteratively refined through expert interviews and subsequently validated in a workshop with experienced architects. Results: The main result of this work is a comprehensive list of 40 checks, organized into two main categories and 16 subcategories. Also, we present the results of a workshop where the importance and degree of achievement of each check was assessed by 25 practicing architects of ML-intensive systems. Conclusion: The findings of this study contribute to a better understanding of the unique challenges of MLintensive systems and offer initial guidance to practitioners, and researchers on areas where future work should be directed. The findings of this study offer valuable support to architects in addressing the unique challenges of ML-intensive systems and provide guidance to practitioners and researchers in terms of where future work should be focused.

Model transformation is a core mechanism in Model-Driven Engineering (MDE), but its use requires expertise in metamodels, transformation languages, execution engines, runtime configuration, and diagnostics. Although LLMs can generate and revise transformation artifacts, direct prompt-to-output generation does not by itself provide an executable MDE workflow grounded in metamodel conformance and transformation rules execution.

This paper proposes agentic model transformation: a service-mediated workflow in which an LLM agent generates candidate transformation specifications and coordinates MDE services for metamodel access, example retrieval, checking, execution, diagnostics, and repair. Model Context Protocol is used to expose these services to the agent. When execution succeeds, target artifacts are produced through the configured transformation infrastructure, serialized by the target modeling environment, and loadable against the registered target metamodel.

Early results from adopting the proposed framework for the UML-to-Java and SysMLv2-to-AAS cases show that constructing and repairing executable transformations is possible. The results also show that executable success and metamodel conformance are only first checks; structural completeness, mapping coverage, and semantic adequacy require additional MDE services. The framework, therefore, shifts LLM support for model transformation from one-off artifact generation towards service-supported transformation development, where specifications, diagnostics, and execution traces can be inspected, reused, and extended.

Architecture as Code has been promoted as a way to bring code-centricity, automation, and approachability to software architecture, yet empirical evidence on how it is realized in industry remains scarce. This paper reports a multiplecase study conducted with a European consulting company serving the financial and insurance sectors. We analyze three engagements of the company with different customers, which offer three complementary instantiations of Architecture as Code. Data were collected over six months through biweekly online workshops, follow-up emails, slide decks, and detailed researcher notes. Analysis proceeded iteratively, alternating between internal consolidation and researcher-industry validation, and concluded with a cross-case synthesis against the existing Architecture as Code framework. Findings show that the Architecture as Code paradigm is present in all cases, albeit at varying levels of maturity. The study provides empirical evidence of how Architecture as Code is practiced across various industrial settings, sharpening its operational meaning by indicating where its characteristics are sufficient and where they require refinement.

Context: Digital Twins (DTs) have gained significant momentum over the past decade, reshaping industrial practices across various sectors, including manufacturing, construction, and infrastructure management. While this rapid adoption has driven innovation in cyber–physical integration, the scientific foundations and engineering principles of DTs are still being actively developed. Objective: Although foundational contributions to model-based engineering (MBE) for digital twins exist, mainly in the form of conceptual frameworks, critical dimensions, such as traceability and federation, remain underexplored. This study aims to provide a comprehensive overview of the state of the art in model-based digital twin engineering, with a focused investigation of these two pivotal concerns. Methods: We conducted a multivocal literature review including 20 peer-reviewed publications and 56 gray literature sources, complemented by a survey of 51 expert stakeholders. The survey was administered during an open session of the MATISSE project plenary meeting, which included both consortium members and external participants from academia and industry. This mixed-method approach was designed to capture both academic and practitioner perspectives on the enablers and barriers to traceability and federation in DTs. Results: Our study reveals how traceability and federation are currently conceptualized, supported, and challenged within model-based DT engineering. We synthesize a landscape of relevant modeling languages, tools, application domains, and technical enablers, and identify key limitations and research gaps. Furthermore, we evaluate existing MBE solutions in light of industrial needs, distinguishing between areas of maturity and those requiring further innovation. The survey results should be interpreted in light of the sampling context, as the respondent pool mainly reflects researchers and early adopters. Conclusion: This work offers evidence-informed guidance for researchers and practitioners by identifying concrete opportunities for improving traceability and federation in DTs. Our findings point to standardized modeling approaches, cross-domain interoperability, and AI-supported traceability mechanisms as promising directions warranting further investigation. Future research priorities include the development of protocols for federated DT systems and the adoption of a system-of-systems perspective to address complex industrial and societal challenges. © 2026 Published by Elsevier B.V.

Large language models are increasingly embedded in software development, yet most evaluations still treat them as one-shot generators for isolated tasks such as code completion or refactoring. In real workflows, however, artifacts such as application programming interfaces, data models, and database schemas co-evolve, and changes must propagate across representations without breaking consistency. When propagation fails, developers incur extra validation, retries, and manual repair, which increases latency and infrastructure cost and undermines sustainable operation. In this study, we ask whether large language models can preserve cross-artifact consistency under change in a round-trip workflow. We apply a controlled edit to one artifact, translate it to its coupled counterpart, and translate it back, then check whether the intended edit persists without drift (i.e., unintended semantic changes or syntactic invalidity). We instantiate this question by synchronizing class-oriented data models with relational database schemas. Using a curated dataset of paired models and schemas and a suite of controlled edit operations, we evaluate four large language model, GPT-5, Qwen3-Next-80B-A3B, DeepSeek V3, and Gemini 2.5, under a unified, reproducible protocol that measures (i) edit persistence, (ii) structural validity (parsability/loadability), and (iii) run-to-run consistency over repeated executions. Our results show that the models handle small, routine edits reliably, but they struggle when edits require structural reorganization. Gemini 2.5 is the most consistent across runs; DeepSeek V3 often preserves the intended semantics but occasionally produces unparsable outputs; Qwen3-Next-80B-A3B exhibits high variance; and GPT-5 often recognizes the change but fails to propagate it coherently through the coupled representation. We contribute a reproducible benchmark and evaluation framework for assessing LLM reliability under artifact co-evolution, together with empirical evidence of current limitations. Overall, the findings reveal a gap between detecting a change and propagating it coherently, underscoring the need for structural validation and human oversight to achieve dependable and cost-efficient LLM-assisted software evolution.

The rapid growth of the Internet of Things has led industries to quickly realize the value and opportunities offered by smart-manufacturing. Within smart-manufacturing, smart-troubleshooting has been introduced, which encompasses a structured set of activities and tools designed to collect failure information from heterogeneous connected devices, analyze it, and align it with appropriate troubleshooting instructions and fixes. In this paper, the general smart-troubleshooting framework is validated using real computerized numeric control (CNC) machine data, in which heterogeneous sensor and event logs are transformed through large-scale feature engineering into a unified time-indexed dataset. A LightGBM (Light Gradient Boosting Machine) model has been trained to predict upcoming faults, enabling early warnings and reducing unplanned downtime. The model has been validated using timeaware cross-validation through an embargo gap to prevent temporal leakage. The proposed approach has been evaluated on a real-world industrial dataset comprising 64 heterogeneous parameters and 1,583 recorded fault timestamps. Using a LightGBM-based baseline, our method achieves a rowlevel F1 score of ≈ 0.781 under time-aware cross-validation. More importantly, when shifting to event-level evaluation, the model attains a global operating point F1 score of ≈ 0.992, demonstrating near-perfect fault event detection with minimal false alarms. In summary, results of model evaluation on CNC machine data show reduced downtime and false alarms, while establishing a scalable foundation for resilient system management in complex industrial environments.

Digital twins, dynamic digital representations of physical systems, are emerging as transformative tools for enhancing crisis preparedness and resilience in critical societal sectors. By enabling real-time monitoring, simulation, and optimization, these technologies offer actionable insights to support proactive risk mitigation, efficient resource allocation, and continuous improvement of crisis response strategies. This study provides a comprehensive knowledge overview of digital twins, focusing on their applicability and impact in key sectors such as energy, healthcare, and transportation. Specifically, it examines the essential services most suited for digital twin adoption, the role of safety-critical data throughout their life-cycle, and their utility in identifying and mitigating risks within critical infrastructure. We employed a mixed-methods research design, combining systematic and gray literature reviews with expert interviews to integrate academic insights with practical perspectives. The findings reveal significant opportunities for digital twins to enhance operational efficiency, strategic planning, and crisis management. However, practical implementation remains in its infancy, with challenges related to cost, complexity, and limited real-world applications. In addition, this study provides actionable recommendations for stakeholders, emphasizing investment in digital twin technologies, robust data governance, and the development of standardized protocols. Future research directions include exploring applications of DTs in emerging sectors, such as crisis preparedness and societal resilience, advancing artificial intelligence integration, and adopting a system-of-systems perspective to address societal challenges comprehensively.

One of the most transformative developments today is the integration of generative artificial intelligence into the development of critical software-intensive cyber-physical systems. From autonomous vehicles to industrial robotics, these systems are entering a new era shaped by artificial intelligence-driven development and automation. In this paper, we consider software engineering, artificial intelligence, artificial intelligence ethics, and social aspects, to explore how such technologies can be harnessed safely, transparently, and with human values at the center. Our contributions include a vision for software engineering, guiding the engineering of future trustworthy safety-critical cyber-physical systems under the influence of generative artificial intelligence. We critically analyze how three established certification principles can be leveraged to cope with the societal and technical tensions introduced by generative artificial intelligence adoption, and propose a research and practice agenda to ensure that future cyber-physical systems development and operations cycle remain trustworthy, both from a system (hardware and software) and from a societal perspective. 

Software quality assurance is pivotal in safety-critical domains such as railway systems, where failures could have catastrophic consequences. In this context, the train control and management system, which enables communication and control across multiple subsystems (such as doors and information panels) within a modern train, and its software must undergo rigorous validation. Alstom Rail Sweden AB employs a digital twin infrastructure to simulate and validate train control and management system software. While this setup improves system-level testing, root-cause analysis of test failures remains a manual, time-consuming bottleneck. In this study, we explore the potential of large language models to automate root cause analysis by interpreting test execution logs generated during digital twin-based testing. We benchmark nine state-of-the-art large language models: Aion-1.0, DeepSeek R1, DeepSeek V3 0324, Mistral Small 3.1 24B, GPT o3-mini, Gemini 2.5 Pro Experimental, QwQ 32B, Gemini 2.0 Flash Experimental, and Amazon Nova 2 Lite using zero-shot chain-of-thought prompting to assess their ability to reason about fault patterns in real-world industrial test execution logs. The logs were sourced from Alstom's digital twin-based testing environment and captured complex operational behaviour typical of embedded, safety-critical systems. Our results showed that long-context LLMs tended to achieve higher accuracy than smaller models. Gemini 2.5 Pro Experimental achieved the best performance with 66.7% accuracy and produced coherent step-by-step explanations. However, coherent reasoning did not consistently translate into correct root-cause identification, suggesting that LLMs are currently better suited to support human-in-the-loop root cause analysis than to fully automate it, and motivating further work to improve prediction accuracy for log-based RCA.

In resource-constrained Internet of Things environments, reducing data transmissions is essential for minimizing energy consumption, network load, and operational costs. Overly aggressive reduction may compromise accuracy, a critical factor in applications such as industrial control. This paper aims to offer practical guidance for selecting suitable data reduction techniques by experimentally evaluating three promising methods from common reduction categories: Data Filtering, Data Aggregation, and Data Prediction. We perform a parameter sweep for each algorithm across three real-world temperature scenarios: stable, rising, and fluctuating. Each configuration is evaluated in terms of data reduction percentage and accuracy, using Total Accumulated Deviation, Mean Absolute Deviation, and Maximum Deviation. Results show that Data Prediction generally achieves the highest accuracy across all scenarios, while Data Filtering tends to yield the greatest reduction at the expense of accuracy. However, all algorithms can be tuned to meet specific scenario demands or accuracy criteria, underscoring that no one-size-fits-all solution exists. We conclude that context-aware algorithm selection and parameter tuning are critical for effective Internet of Things data management.