A compact and efficient physics-informed architecture for reconstructing and predicting complex physical fields from...

Complex physical fields are traditionally characterized through high-fidelity simulations or dense sensor networks, which provide detailed information but are often computationally or experimentally costly. Deep learning offers a promising approach for reconstructing and forecasting such fields from limited observations. However, many existing AI-based models rely on large labeled datasets and heavily parameterized architectures, making the models both data-intensive and computationally demanding to train. To overcome these limitations, we introduce compact and computationally efficient physics-informed neural networks (PINNs) designed to reduce both architectural complexity and computational overhead. We demonstrate its performance to reconstruct and forecast complex fields with limited or low accuracy multiscale spatiotemporal dynamics using laminar and turbulent flows as representative benchmarks. We first demonstrate accurate super-resolution of lid-driven cavity flows at Reynolds numbers up to $Re=1000$ , where the PINNs reconstruct high-resolution, higher accuracy solutions (401 × 401) from coarse, low-accuracy inputs (40 × 40), reducing error significantly. Extending to three-dimensional Homogeneous Isotropic Turbulence (HIT), the model recovers fine-scale, highly accurate structures from coarse and inaccurate data, preserving key physical and statistical properties. We further show that the framework enables forecasting from limited temporal snapshots, with controlled error growth and accurate prediction of turbulent kinetic energy evolution. Finally, we investigate spatial domain extension, where the model reconstructs flow fields in previously unseen regions. The results reveal that while large-scale structures are recovered robustly, reconstruction fidelity depends strongly on the placement of sparse supervision, with physically informed sampling improving accuracy and energy consistency. The results also demonstrate that the proposed framework scales effectively to larger domains with increasingly complex flow structures. Across the tasks considered, the architecture contains approximately 1-1.5 million trainable parameters and our implementation requires a 3-5 hours of training stage for each task on a single NVIDIA A100 GPU. Together, these results demonstrate that the proposed compact architecture provides a unified and training-efficient route to physics-informed generalization across space and time, enabling the reconstruction and prediction of complex physical fields with multiscale spatiotemporal dynamics from limited observations.