Researchers compare physics-informed neural networks using different thermodynamic frameworks to improve accuracy and physical consistency in solving differential equations.

Study systematically compares multiple thermodynamic structure-informed neural networks (PINNs) incorporating Newtonian, Lagrangian, and Hamiltonian mechanics for conservative systems

Evaluates advanced formulations including Onsager variational principle and extended irreversible thermodynamics for handling dissipative systems

Comprehensive experiments measure impact on accuracy, physical consistency, noise robustness, and interpretability across representative ODEs and PDEs

Findings indicate Newtonian-residual-based PINNs can reconstruct system states but show reliability limitations in certain applications