ABSTRACT
This paper introduces a Multi-Fidelity Physics-Informed Deep Neural Network (mPFDNN) framework to predict non-equilibrium dynamic properties in porous materials, aiming for a 100×–1000× speedup over conventional NEMD simulations while maintaining physical consistency.
PAPER · PDF
Loading PDF...
Key findings
Integrates multi-fidelity data fusion, physics-informed constraints, and Fourier neural operators.
Achieves significant speedup over conventional NEMD while maintaining accuracy.
Validated against NEMD simulations for zeolites, MOFs, and carbon molecular sieves.
Limitations & open questions
Current applications are limited to static equilibrium properties and simple diffusion.
Challenges in capturing multi-scale phenomena in heterogeneous media.