mPFDNN datasets and related models

To rigorously evaluate the universality and accuracy of Material-Property-Field-based Deep Neural Network (mPFDNN), we curated a diverse set of benchmark datasets spanning inorganic crystals, organic molecules, and catalytic systems. As summarized in Table 1, these datasets cover a wide spectrum of critical properties, including standard force field targets (e.g., energies and atomic forces) and key electronic properties derived from first-principles calculations (e.g., polarizability, dielectric constants, and dipole moments). Notably, the included elements span nearly the entire periodic table, providing a stringent testbed for assessing the model’s transferability. Table 1 Summary of benchmark datasets. Category Dataset Sampler Number Property Crystal Jarvis 75,908 Formation Energy, E Total Energy, Total E Voigt bulk, Kv Shear modulus, Gv Dielectric constant, є MPtraj (M3GNet) 187,687 Energy/Force/Stress Molecule QM9 133,885 Isotropic polarizability, α Electronic spatial extent, r2 Dipole moment, μ Internal energy at 0K, U0 Drug 1.4 Million Energy/Force OC2M 2 Million Energy/Force Metal Alloy 72,722 Energy/Force Aqueous solutions Water & Salt-wate 33,819 Energy/Force HEA IrPdPtRhRu 16,772 Energy/Force