Data for Physics-Informed Neural Networks for Partial Differential Equations: From Training Objectives to Solution Accuracy

This dataset contains the numerical outputs and provenance materials for the controlled diagnostic study accompanying the manuscript “Physics-Informed Neural Networks for Partial Differential Equations: From Training Objectives to Solution Accuracy.” The baseline factorial experiment comprises 120 physics-informed neural network runs across three partial differential equations (Poisson, forced Helmholtz, and viscous Burgers), two numerical precisions (FP32 and FP64), two optimizer settings (Adam and Adam followed by L-BFGS), two collocation strategies (fixed and residual-adaptive), and five random seeds. The archive includes run configurations, sanitized execution-environment metadata, trained model parameters, training histories, final collocation points, predictions, reference solutions, residual evaluations, quantities of interest, run-level metrics, grouped statistics, paired comparisons, diagnostic figures, integrity reports, and a frozen experiment-code snapshot. File-level SHA-256 checksums and a data dictionary are included. The maintained source code and reproduction instructions are available at:https://github.com/wyq12211/pinn-objectives-to-solution-accuracy