https://www.mdu.se/

Machine learning is increasingly called upon to guide decisionsin critical industrial applications. Its predictive powerpromises gains in efficiency, yet its black-box nature and lackof guarantees pose risks in contexts where behavior must remainanalyzable and safe. This thesis asks how machine learning can bemade trustworthy, explainable, and efficient enough for engineersto deploy in practice. Three gaps hamper broader adoption. Few works provide formal orstatistical guarantees on ML outputs paired with explanationsthat engineers can act on (Gap~A). Data-driven models thatgeneralize across hardware configurations without retrainingremain rare, and existing simulators are prohibitively slow(Gap~B). Many contributions address individual components ofindustrially motivated problems without combining them intovalidated end-to-end pipelines (Gap~C). To address Gap~A, we apply abstraction to neural networks,showing that inputs with negligible effect on the output can beformally identified and removed, producing simpler yet boundedmodels open to verification. We then introduce a conformalprediction framework for CPU load forecasting that providesstatistically guaranteed coverage intervals, combined withShapley value analysis to trace individual task contributions tothe predicted load. To address Gap~B, we develop a data-drivencache memory surrogate using long short-term memory networks,reproducing cache miss distributions across unseen hardwareconfigurations at a fraction of the simulator's computationalcost. To address Gap~C, we present HASCO, a Hybrid AI SimulationCompiler that translates natural language accident reports intoexecutable vehicular simulation scenarios through a structuredcompilation approach with deterministic validation. Together, these contributions establish a path toward machinelearning that is not merely powerful but trustworthy, explainable,and practically deployable in the industrial workflow.