<p>Conventional numerical and data-driven methods for structural damage identification often struggle with problems such as poor generalization, sensitivity to noise, and limited accuracy in quantifying and localizing damage, especially under real-world uncertainties. This study introduces a novel parallel physics-informed neural network framework that addresses these limitations by integrating physical constraints directly into the learning process and employing a nondimensionalized formulation of the governing equations to improve training stability and convergence. The method is applied to simultaneously identify the location and severity of damage in numerical and experimental shear frame model using measured acceleration responses from the structure. In the numerical case, different damage scenarios are analyzed under white Gaussian noise and real earthquake excitations, across various noise levels. The model demonstrates high accuracy in identifying damaged locations even under high noise conditions, and outperforms traditional response surface methods by reducing false positive cases. Experimental validation using open-source shear frame data further confirms the model’s effectiveness. Despite the presence of modeling uncertainty and measurement, the model accurately identifies damage locations and estimates their severity in agreement with experimental observations. Overall, the model proves its efficacy in predicting damage locations and assessing the severity levels even under uncertainty and noise levels.</p>