Adaptive physics-aware models are central to reliable engineering decision support, particularly for structural digital twins where model inaccuracies can lead to biased state estimation and unreliable predictions. Nominal finite element (FE) models inevitably suffer from model-form errors (MFEs) due to simplifications in geometry, boundary conditions, damping, and material behavior. These discrepancies becomes especially critical in high-dimensional systems and limit the practical use of digital twins for diagnosis and prognosis. This work proposes an adaptive physics-aware probabilistic framework for rectification of MFEs and state estimation in high-dimensional linear structural systems by integrating Gaussian Process Latent Force Models (GPLFMs), Bayesian filtering, and mesh-invariant neural surrogates. The framework is adaptive in the diagnosis stage by explicitly modeling modal MFEs as latent, non-parametric discrepancy forces within a GPLFM formulation. A Bayesian filtering scheme jointly estimates the structural modal states and modal discrepancies from measurement data, enabling adaptation of the model by eliminating bias induced by modeling errors. This allows the digital twin to provide unbiased state estimates along with quantified uncertainty over the state estimates in the presence of noisy measured data. Adaptivity is further achieved in the prognosis stage through a neural network surrogate trained in the reduced modal space to learn the discrepancy dynamics. By operating in modal coordinates, the learned correction becomes mesh-invariant and can be transferred across different FE discretizations without retraining. This enables adaptive prediction and forward simulation for arbitrary mesh resolutions, making the framework scalable and practical for high-dimensional digital twins. The proposed approach preserves the underlying physics-based structure while augmenting it with data-driven corrections. Validation on a structure with distinct MFE demonstrates substantial improvements in predictive accuracy, with up to 90–99% reduction in normalized mean square error (NMSE).