The rapid advancement of digital technologies has revolutionized industrial maintenance practices, with predictive maintenance emerging as a proactive and cost-effective strategy. The conventional machine learning (ML) model for damage identification is trained by data collected from sensors, where majority of the data is collected when the structure is in undamaged condition. Due to the scarcity of damaged condition data, the extrapolation capacity of conventional ML model is questionable, especially dealing with new unseen damages. Furthermore, it is not practical to create new damages on the physical structure just for data collection purposes. This research explores the feasibility of utilising modal-based features in a hybrid digital twin damage identification scheme by integrating physics-based finite element (FE) model and ML model, in which artificial neural network (ANN) is used. A lab-scale plate-like structure supported at 4 corners is used as the test rig, with damage simulated by loosening support fasteners. Modal data collected from the test rig by Impact-Synchronous Modal Analysis (ISMA) is processed and used for ML training, with high severity damage treated as unseen damage intentionally left out. The ML model failed to identify the unseen damage. A correlated high-fidelity FE model of the test rig is developed, where the natural frequency error is less than 6% and the modal assurance criterion (MAC) is greater than 0.95 for the first 3 modes when undamaged. The FE model created synthetic high severity damage (unseen damage) data generated to train a new FE model-informed ML model. Validation accuracy of the FE model-informed ML model is 100.00%. When tested with real high severity damage (unseen damage) data, the FE model-informed ML model predict accurately the damage location and damage severity for all the damage cases. By utilising modal-based features and integrating FE and ML models, the FE model-informed ML model showcases enhanced accuracy in predicting unseen damages.