10:45 AM - 11:15 AM
[SCG49-01] Molecular Dynamics Simulation with Machine-Learning Interatomic Potential and Its Applications to
Computing Free Energy and Thermal Conductivity
★Invited Papers
Keywords:Molecular Dynamics, Machine-Learning Interatomic Potential, Free Energy, Themal Conductivity
Machine Learning Interatomic Potentials (MLIPs) not only has low computational cost but also can achieve high accuracy by training first-principles (FP) calculation data. These powerful advantages have led to MLIPs being used in a wide range of material science. However, the arbitrariness in what and how to train motivates us to construct better MLIPs.
We have developed a training scheme for MLIPs that can perform highly accurate molecular dynamics (MD) simulations. Since accuracy of energy, atomic force, and pressure are important for MD, we trained MLIPs with FP calculation data of these three physical quantities. When using Na as a test system, we confirmed that the solid-liquid phase transition can be reproduced [1]. Incidentally, to simultaneously train such three physical quantities with different dimensions, adjustment of coefficients in the cost function C (Eq. (1)) that are minimized during training have a marked effect on the accuracy of MLIPs [2]:
C = pEΔE + pFΔF + pPΔP, (1)
where ΔE, ΔF, and ΔP are mean square errors of energy, atomic force, and pressure, respectively, between FP and MLIP output values. pE, pF, and pP are the adjustable coefficients. The coefficient adjustment was essential to reproduce the atomic structure of the complicated materials such as Rb-Na mixtures [2].
The MD simulations with MLIPs provide sufficient statistics so that essential physical quantities such as free energy and thermal conductivity can be calculated. In the free energy calculation by thermodynamic integration method, we were able to estimate the melting point of Na with high accuracy [1]. For the thermal conductivity calculation, a method based on the Green-Kubo formula, which can be applied to disordered systems, was adopted. We found that the accuracy of thermal conductivity is improved by derivation of the rigorous heat flux formula based on the many-body effect of MLIP and by training pressure [3]. The latter is because the heat flux formula is closely related to the one of pressure. It has also been found that heat flux regularization during training of MLIPs was effective for calculating not only total thermal conductivity but also partial thermal conductivities such as elemental contributions [4]. It is simply achieved by adding a heat flux squared term J2 to the original cost function C as follows:
C' = C + pJJ2 , (2)
where pJ is the adjustable coefficient. The regularization term works to reduce nonphysical heat fluxes caused by the many-body effect of MLIPs. Using the above training approaches, we estimated total thermal conductivity and their elemental contributions of silver chalcogenides [3,4].
The development of these training methods has enabled highly accurate MD simulations and estimations of the physical quantities such as free energy and thermal conductivity. In our presentation, specific applications will be presented, including important technical aspects in the training methods.
References
[1] A. Irie., et al., J. Phys. Soc. Jpn., 91, 045002 (2022).
[2] A. Irie., et al., J. Phys. Soc. Jpn., 90, 094603 (2021).
[3] K. Shimamura, et al., Chem. Phys. Letts., 778, 138748 (2021).
[4] K. Shimamura, et al., arXiv:2204.01405 (2022).
We have developed a training scheme for MLIPs that can perform highly accurate molecular dynamics (MD) simulations. Since accuracy of energy, atomic force, and pressure are important for MD, we trained MLIPs with FP calculation data of these three physical quantities. When using Na as a test system, we confirmed that the solid-liquid phase transition can be reproduced [1]. Incidentally, to simultaneously train such three physical quantities with different dimensions, adjustment of coefficients in the cost function C (Eq. (1)) that are minimized during training have a marked effect on the accuracy of MLIPs [2]:
C = pEΔE + pFΔF + pPΔP, (1)
where ΔE, ΔF, and ΔP are mean square errors of energy, atomic force, and pressure, respectively, between FP and MLIP output values. pE, pF, and pP are the adjustable coefficients. The coefficient adjustment was essential to reproduce the atomic structure of the complicated materials such as Rb-Na mixtures [2].
The MD simulations with MLIPs provide sufficient statistics so that essential physical quantities such as free energy and thermal conductivity can be calculated. In the free energy calculation by thermodynamic integration method, we were able to estimate the melting point of Na with high accuracy [1]. For the thermal conductivity calculation, a method based on the Green-Kubo formula, which can be applied to disordered systems, was adopted. We found that the accuracy of thermal conductivity is improved by derivation of the rigorous heat flux formula based on the many-body effect of MLIP and by training pressure [3]. The latter is because the heat flux formula is closely related to the one of pressure. It has also been found that heat flux regularization during training of MLIPs was effective for calculating not only total thermal conductivity but also partial thermal conductivities such as elemental contributions [4]. It is simply achieved by adding a heat flux squared term J2 to the original cost function C as follows:
C' = C + pJJ2 , (2)
where pJ is the adjustable coefficient. The regularization term works to reduce nonphysical heat fluxes caused by the many-body effect of MLIPs. Using the above training approaches, we estimated total thermal conductivity and their elemental contributions of silver chalcogenides [3,4].
The development of these training methods has enabled highly accurate MD simulations and estimations of the physical quantities such as free energy and thermal conductivity. In our presentation, specific applications will be presented, including important technical aspects in the training methods.
References
[1] A. Irie., et al., J. Phys. Soc. Jpn., 91, 045002 (2022).
[2] A. Irie., et al., J. Phys. Soc. Jpn., 90, 094603 (2021).
[3] K. Shimamura, et al., Chem. Phys. Letts., 778, 138748 (2021).
[4] K. Shimamura, et al., arXiv:2204.01405 (2022).