Towards Formalized Training and Validation of AI-based Systems with Digital Twin Scenarios

Artificial intelligence (AI) has emerged as a pivotal technology for autonomous systems across various domains, but quality assurance remains challenging due to limited training data and inadequate validation methods. This paper presents a framework combining scenario engineering and digital twins to address these challenges. It supports AI system development from synthetic data generation to validation within a unified simulation-based approach. Scenario engineering focuses on real-world scenarios, while digital twins provide virtual replicas of system components throughout their lifecycle. The framework ensures flexibility in system design and testing, offering formalized standards for quality assurance workflows. Two application case studies are presented: training a LiDAR-based pose estimation model in aerospace and validating an autonomous driving function. These case studies demonstrate the applicability of the framework and its potential for continuous verification and validation.