https://www.mdu.se/

Artificial intelligence (AI) is increasingly portrayed as a disruptive force and is often integrated into organizational routines as part of broader digitalization efforts across organizations in different industries in order to transform work practices and improve decision-making. While previous research has highlighted the anticipated effects of algorithms on organizational routines as well as how their use affects the performance of routines, less attention has been devoted to understanding how AI influences routines design, that is, the intentional change of routines, particularly in cases where algorithms and AI technologies are simultaneously developed and implemented as part of the change process.

Drawing on Routine Dynamics and adopting a processual perspective, this thesis develops a deeper understanding of AI’s influence on the design of organizational routines, particularly in situations where the development of machine learning algorithms and integration of AI technologies are central to routine change. This is done based on a longitudinal ethnographic study of a routine design case in which an AI software company and an electricity distribution utility company engaged in an AI digitization project aimed at transitioning from manual to virtual inspection of overhead powerlines using drones and AI.

The analysis of this case shows that routine design unfolds through a broadly distributed sociomaterial assemblage spanning multiple organizational and spatiotemporal boundaries, involving a range of human and nonhuman actors. The routine design process is therefore conceptualized as a distributed algorithmic routine design, characterized by iterative design actions for preparing, repairing, and post-pairing the routine under design. Through these actions, actors are specifically engaged in developing algorithms while simultaneously orchestrating the sociomaterial assemblage required to make AI function in practice. The study further identifies design disjunctures as emergent sociomaterial challenges that reveal tensions, misalignments, and coordination difficulties across the design assemblage.

The thesis contributes to Routine Dynamics research by showing that designing routines with AI is not merely a matter of developing and integrating artifacts. Rather, it involves unpacking and continuously orchestrating the distributed sociomaterial assemblage through which AI’s agency is enacted and sustained over time. This specifically requires aligning all heterogeneous elements that constitute the design assemblage and collectively enable its continuity and performance. In doing so, the study highlights the importance of attending to both visible and invisible forms of work, the balancing of continuity and discontinuity, material artifacts, and the emergence of intentionality within ongoing nonlinear design processes. Overall, this thesis advances understanding of how AI reconfigures routine design and highlights the complex ongoing sociomaterial work required for designing and implementing algorithmic routines.

This thesis investigates how data-driven analytics can be systematically implemented in legacy industrial batch processes to support robust, interpretable, and operationally relevant decision-making.Industrial batch environments are characterised by heterogeneous data structures, variable process trajectories, evolving operating conditions, and limited contextualisation, complicating the direct application of conventional analytical methods.

The work develops an integrated analytical framework combining batch data analytics, machine learning, cost-sensitive learning, and model-agnostic post-hoc analysis.Batch data analytics is used to contextualise and consolidate irregular industrial process data into analytics-ready representations suitable for downstream modelling.Machine learning methods are subsequently applied to perform classification, regression, and degradation modelling across multiple industrial case studies, with emphasis placed on parsimonious and interpretable models rather than unnecessary model complexity.

To extend predictive modelling beyond conventional accuracy-oriented evaluation, cost-sensitive learning is introduced to align model behaviour with operational and economic objectives.The proposed framework demonstrates how predictive confidence and model coverage can be balanced against operational risk and cost constraints, enabling selective and value-aware decision-support strategies.

Beyond predictive optimisation, the thesis introduces a model-agnostic post-hoc analysis framework for examining model behaviour across operational regimes.By embedding interpretability metrics into reduced-dimensional representations and constructing continuous behavioural landscapes through surrogate modelling, regions associated with confidence, systematic error, uncertainty, and interaction-driven behaviour can be identified and analysed.The results demonstrate that predictive performance is not uniformly distributed across the input space, but instead governed by distinct operational regimes with varying levels of reliability and interpretability.

The framework is validated through multiple industrial case studies within alloy production, ceramic manufacturing, and degradation modelling of electrical resistance heating wires.The results show that structured batch contextualisation improves the suitability of industrial data for machine learning, that selective modelling strategies can achieve substantially higher predictive performance within identified operational regions, and that region-aware post-hoc analysis enables diagnostically grounded evaluation of model behaviour under changing industrial conditions.

The thesis contributes a coherent methodological framework for trustworthy industrial analytics in legacy batch environments by integrating structured data contextualisation, interpretable machine learning, value-aware evaluation, and region-aware behavioural analysis into a unified decision-support perspective.

District heating systems are central to the Nordic energy transition. Yet as supply temperatures fall, renewable integration increases, and buildings take on a more active role, network performance depends increasingly on what happens at the interface between buildings and the network. This thesis develops methods for the interface under uncertainty by following a progression from model representation to control, fault management, and strategic planning. A nonlinear gray-box modeling framework first combines a resistance and capacitance thermal network with a physically motivated radiator heat emission model. The framework predicts indoor air temperature with high accuracy in both residential and commercial buildings and retains the same structural form across building archetypes through parameter re-identification. Building on this foundation, a risk-aware training framework augments the conventional prediction loss with a conditional value-at-risk penalty on the operational cost so that the surrogate model is shaped not only for nominal accuracy but also for reliable control under stressed conditions. In closed-loop evaluation under weather-stress scenarios, the resulting controller reduced occupied cold degree-hours and peak cold violations relative to a fidelity baseline while making the comfort and energy trade-off explicit. A systematic review of 140 district heating and cooling control studies situates these developments within the wider field and highlights the need for methods that bridge accurate predictions and reliable operations. Robust operation also requires resilience to hardware degradation; therefore, an integrated fault detection and compensation framework was developed by combining unsupervised anomaly detection, signature-based diagnosis, and supervisory supply temperature modulation. This framework establishes a compensability spectrum that distinguishes faults that can be mitigated autonomously from those that require rapid maintenance. Finally, the thesis extends from single-building operations to prosumer and portfolio-level decision support, showing that operational strategy materially affects economic performance and that a data envelopment analysis ranking framework can identify robust rooftop photovoltaic investment candidates under joint climate and market uncertainty. Overall, this thesis shows that uncertainty is not a residual complication but a condition that should be addressed explicitly across model development, operational control, fault management, and long-term investment prioritization.

This doctoral thesis describes research aimed at creating a unified understanding of the fundamentals of systems of systems to support modeling practices and thinking principles. This is approached through the development of an ontology-driven conceptual modeling of systems of systems. This approach combines an abstract mental model of a reality with the formalities of ontology theories to build structured knowledge representations of systems of systems using a core ontology artifact. The development of this core ontology adopts the constructive research methodology employing various research methods, tools, and processes. Narrative and systematic literature reviews and exploratory studies were applied to study the systems of systems problem domain and the foundations for system of systems knowledge representation. Iterative ontology development processes, case studies illustrations, formalization and usability analysis studies facilitated the development of the ontology artifact. Different assessment and feedback mechanisms were employed. These include feedback from the iterative ontology development process, an ontology alignment for philosophical clarity, internal and external consistency checking of the formalization process, and a survey-based evaluation study. The outcome of this thesis is a knowledge representation that facilitates semantic interoperability, machine readability, knowledge reuse, and overall, a proposal of an ontological source of truth for the formulation and development of systems of systems. This can play a role in facilitating industry-related tasks such as data integration, task optimization through semantic harmonization and capability mapping, therefore providing a baseline that minimizes ambiguity and facilitates design space exploration.

Static program analysis aims to analyse program behavior without executing it, and has traditionally been implemented as offline tools running on desktop or server machines. As embedded and edge platforms gain compute capability, however, parts of these analyses are increasingly expected to run in new contexts—for example, inside Integrated Development Environments (IDEs), where developers require continuous, low-latency feedback as code evolves, and on embedded systems and hardware, where integrating selected analysis tasks can improve performance or enable new functionality. These settings impose strict constraints—low latency, limited memory, and energy efficiency—which favor demand-driven and incremental analyses over batch-oriented processing. A major obstacle to deploying static analyses in such settings is their reliance on multi-stage graph construction and fixed-point iteration; for example, computing control dependencies commonly involves transforming the program into a node-based representation, constructing a Control Flow Graph (CFG), and deriving a post-dominator tree (PDT), and materializing these intermediate structures increases both latency and memory footprint. This thesis enables the use of static program analysis techniques in such constrained environments through two distinct and complementary directions: first, it reduces unnecessary computation in unstructured code by eliminating full fixed-point iteration and avoiding complete construction of CFGs and PDTs; second, it introduces a syntax-directed method for computing control dependencies without constructing intermediate graphs or reformulating the program into additional node-based representations. The method is single-pass, requires only a constant number of auxiliary in-memory entities, and eliminates graph overhead; experimental evaluation shows that it preserves accuracy while substantially reducing memory usage and improving performance—typically by 6–12× compared to representative graph-based baselines—for well-structured programming languages. Although these two directions are distinct, they are complementary: the syntax-directed method is designed to be integrated into the demand-driven approach.

The rapid adoption of digital technologies has transformed the organizational structures of multinational corporations (MNCs) and altered how they coordinate geographically dispersed knowledge activities, centralizing some value‑chain activities while decentralizing others. Yet, despite growing scholarly interest, IB research continues to conceptualize digitalization largely as a unidimensional phenomenon, overlooking the heterogeneous and sometimes opposing effects of different forms of digital technologies on intra‑MNC coordination and knowledge transfer. This dissertation addresses this conceptual and empirical gap by examining how distinct forms of digital technology adoption trigger different organizational conditions and, in turn, influence subsidiaries’ capacity to transfer knowledge within the MNC network and their market-facing agility to contribute strategically. It comprises four papers: one conceptual study and three empirical studies based on two datasets.

The first empirical study builds the theoretical and methodological foundation for this dissertation by conceptualizing and operationalizing a firm’s digital technology adoption from a functional perspective. It develops a psychometrically validated 19-item scale manifested in four core dimensions: digital communication technology (DCT), digital in situ technology (DST), digital data analytics technology (DDAT), and digital networking technology (DNT), which reflect the coordination, optimization, data access, and location access functions that digital technologies perform within firms. The resulting measurement instrument, based on survey data from 113 technology-intensive Swedish firms, offers scholars a precise means to examine how digital technologies influence organizational processes. 

The second study advances a theoretical argument suggesting that MNCs’ adoption of distinct forms of digital technologies shapes value-chain activities through opposing centrifugal and centripetal forces associated with DCTs and DSTs, respectively. These opposing dynamics generate differentiated organizational conditions, leading to variations in subsidiaries’ capacity to disseminate knowledge to headquarters.

The third and fourth studies empirically demonstrate how MNCs’ adoption of different forms of digital technologies affects lateral knowledge transfer and subsidiaries’ responsiveness to market changes, drawing on data from 100 foreign subsidiaries of 14 Swedish MNCs. The findings reveal that while DST alone enhances operational efficiency but limits knowledge transfer, a combination of DST and DNT reduces a focal subsidiary’s dependence on local embeddedness and enhances lateral outward knowledge transfer via digital platforms. Moreover, lateral inward knowledge transfer and DDAT adoption enable subsidiaries to interpret market insights and respond rapidly to emerging opportunities.

Overall, this dissertation contributes to the IB literature by demonstrating that MNCs’ adoption of different digital technologies has heterogeneous organizational effects on knowledge flows and subsidiary roles. The findings offer guidance for managers assessing digital technology investments and strategic positioning.