Metamaterials are widely utilized in advanced engineering applications due to their customizable mechanical properties. However, achieving efficient and accurate full-field mechanical analysis for these materials remains significant challenges: conventional Digital Image Correlation (DIC) techniques are restricted to surface observations, while numerical methods like Finite Element Analysis (FEA) suffer from prohibitive computational costs due to multiscale features[1,2]. Consequently, developing a non-invasive method capable of accurately and efficiently reconstructing full-field responses from accessible surface data is critical. To address this problem, we propose a Fourier-Embedded Physics-Informed Neural Network (PINN) framework that integrates sparse surface measurements. The framework is constructed in three stages. First, the architecture employs Fourier feature mapping to embed input coordinates into a higher-dimensional space and utilizes a decoupled network structure to independently predict displacement components, which collectively enhance the network’s ability to capture high-frequency geometric details[3]. Subsequently, training is driven by a composite loss function designed to enforce the governing equations of linear elasticity while fusing boundary conditions with sparse strain data acquired from the six outer surfaces[4]. Finally, an ensemble learning strategy based on randomized initializations is implemented to rigorously quantify the epistemic uncertainty in the predictions[5]. Upon validation against numerical simulations and experimental DIC data, the framework maintains a global relative L2 error of 5–10% and provides reliable uncertainty bounds. These results demonstrate the method's capability to invert high-fidelity full-field strain distributions from sparse surface measurements, paving the way for the intelligent non-destructive evaluation and digital twinning of metamaterials.