Accelerating PINN Convergence for Signal Reconstructions

In this paper, we explore the PINNslope method for seismic data interpolation, leveraging Physics Informed Neural Networks (PINNs) to simultaneously reconstruct wavefields and estimate local slopes. Our study used synthetic data derived from the Marmousi model, where traces were randomly removed to simulate gaps. Comparing a non-adaptive network with our proposed adaptive network architecture, we observed that while both approaches achieved effective reconstruction, the adaptive network demonstrated superior efficiency, achieving comparable results in significantly fewer epochs. This efficiency gain underscores the potential of adaptive techniques in accelerating convergence and improving the robustness of PINNslope for challenging data reconstruction tasks.