Digital twins are increasingly applied for real-time monitoring in engineering systems. Data-driven digital twins trained exclusively on experimental measurements, are valid only within the limited range of operating conditions under which the data were collected. Extending their predictive capabilities beyond these conditions typically requires extensive measurement campaigns, which are often costly or impractical. In this work, we propose a hybrid, physics-informed strategy in which simplified finite-element (FE) simulations are used to generate physically consistent training data for data-driven models. By decomposing complex real-world loads into physically meaningful components, the proposed approach enables systematic enrichment of training datasets and improves the generalization capability of AI-based digital twins beyond the conditions observed in measurements, while keeping computational cost low. We illustrate the approach on a representative engineering use case: detecting the driving state of a vehicle (straight driving, turning at an angle) using strain measurements from a car rim, for which we previously developed a purely data-driven digital twin. The rim response is decomposed into a set of physically interpretable loading scenarios, including vertical, tangential, axial, and inclined loads, representing braking, acceleration, turning, and combined loading conditions. The parameters for these individual loading cases are informed by laboratory measurements on a single wheel, while the resulting digital twin is validated against strain data collected under dynamic test-track conditions.