Licentiate Proposal Approved

Machine learning is increasingly called upon to guide decisions in critical industrial applications. Its predictive power promises efficiency and adaptability, yet its black-box nature and lack of guarantees pose risks in contexts where behavior must remain analyzable and safe.

This thesis addresses how machine learning can be strengthened to become not only powerful, but also accountable, explainable, and usable by engineers in practice. Key contributions include:

• Neural Network Abstraction: Formally identifying and removing inputs with little effect on outcomes, producing simpler, bounded, analyzable models.
• Conformal Prediction with Shapley Values: Providing coverage guarantees and tracing load contributions back to individual tasks for safety-relevant insight.
• Data-Driven Cache Simulator: Reproducing hardware behavior at a fraction of the computational cost.
• HASCo (Hybrid AI Simulation Compiler): An end-to-end pipeline combining all principles, generating runnable simulation scenarios from natural-language safety reports.

These contributions establish a path toward machine learning that becomes a transparent partner in the industrial workflow, rather than remaining an opaque black box.