Deep Potential Molecular Dynamics Research Articles

Thermoelectric Figure of Merit of a Superatomic Crystal Re6Se8I2 Monolayer

Superatomic materials are newly emerging candidates for high-performance thermoelectric (TE) devices due to their intrinsic ultralow thermal conductivities. However, the low TE power factor becomes a huge obstacle to reaching the required dimensionless figure of merit (ZT) values for practical applications. Here, motivated by the recently synthesized superatomic ${\mathrm{Re}}_{6}{\mathrm{Se}}_{8}{\mathrm{I}}_{2}$ monolayer [He et al. J. Am. Chem. Soc. 144, 74 (2022)], we study its superior TE properties by using density functional theory combined with phonon Boltzmann transport theory and deep potential molecular dynamics. We show that the large mass and anharmonic $\mathrm{Re}---\mathrm{I}$ bonds introduce strong phonon scattering and result in a low lattice thermal conductivity of 1.20 W m${}^{\ensuremath{-}1}$ K${}^{\ensuremath{-}1}$ at 300 K, while the strong and harmonic $\mathrm{Re}---\mathrm{Se}$ network ensures a high TE power factor of 4344 \textmu{}W m${}^{\ensuremath{-}1}$ K${}^{\ensuremath{-}2}$ in the b direction for n-type doping. This is an order of magnitude higher than those of other cluster-based materials. Owing to the low thermal conductivity and high TE power factor, ${\mathrm{Re}}_{6}{\mathrm{Se}}_{8}{\mathrm{I}}_{2}$ exhibits higher ZT values of 1.20 (500 K) and 1.43 (900 K) with n-type doping along the b direction compared with other cluster-based TE materials reported so far.

Metal Affinity of Support Dictates Sintering of Gold Catalysts.

Sintering during heterogeneous catalytic reactions is one of the most notorious deactivation channels in catalysts of supported metal nanoparticles. It is therefore critical to understand the effect of support on the sintering behavior. Here, by using in situ aberration-corrected transmission electron microscopy and computational modeling, the atomic-scale dynamic interactions are revealed between Au nanoparticles and various supports. It is found that Au nanoparticles on ceria have a smaller contact angle and are apparently less mobile, especially at surface steps when compared with those on the amorphous silica. Analogous to hydrophilicity, we attribute the origin of mobility of small nanoparticles to metal affinity, which determines the interaction between metal and support material. Ab initio molecular dynamics (AIMD) and machine learning-based deep potential molecular dynamics (DPMD) simulations directly capture a coalescence process on the silica surface and the strong pinning of gold on ceria. The joint experimental and theoretical results on the atomic scale demonstrate the metal affinity of active and inert supports as the key descriptor pertinent to sintering and deactivation of heterogeneous catalysts.

Predicted superconductivity and superionic state in the electride Li5N under high pressure

Recently, electrides have received increasing attention due to their multifunctional properties as superconducting, catalytic, insulating, and electrode materials, with potential to offer other performance and possess novel physical states. This work uncovers that Li5N as an electride possess four novel physical states simultaneously: electride state, super-coordinated state, superconducting state, and superionic state. By obtaining high-pressure phase diagrams of the Li–N system at 150–350 GPa using a crystal structure search algorithm, we find that Li5N can remain stable as P6/mmm structure and has a 14-fold super-coordination number, as verified by Bader charge and electron localization function analysis. Its superconducting transition temperature reaches the highest at 150 GPa (T c = 48.97 K). Besides, Li5N exhibits the superionic state at 3000 K, in which N atoms act like solid, while some Li atoms flow like liquid. The above results are further verified at a macroscopic level by using deep learning potential molecular dynamics simulations.

Prediction on local structure and properties of LiCl-KCl-AlCl3 ternary molten salt with deep learning potential

LiCl-KCl-AlCl3 molten salt is the main electrolyte for Al-Li alloys electrolysis. It’s beneficial for the electrolyte optimization to have a deeply understand to the local structure and properties of LiCl-KCl-AlCl3 molten salt. Deep potential molecular dynamics simulations were applied in this study, which can balance efficiency and accuracy of the calculations. A deep potential model for LiCl-KCl-AlCl3 molten salt was trained to predict the local structure and properties of LiCl-KCl-AlCl3 molten salt. The structure and properties with temperature were analyzed. The structure information includes the partial radial distribution functions and coordination number distribution; the properties consist of density, self-diffusion coefficient, viscosity, ionic conductivity, thermal expansion coefficient and heat capacity. The structure analysis shows that due to the strong and stable covalent bond between Al-Cl ion pairs, the increase of temperature has no obvious effect on the coordination structure of Al3+ and Cl–. For K-Cl and Li-Cl ion pairs, as the temperature increases, the connection between the anion and cation clusters weakens, and the overall structure of the molten salt tends to be loose, so the density and shear viscosity of the LiCl-KCl-AlCl3 molten salt decrease with the increase of temperature, the self-diffusion coefficient and ionic conductivity increase with increasing temperature, and the thermal expansion coefficient increases with increasing temperature. The variation of the self-diffusion coefficient, shear viscosity and ionic conductivity with temperature follows the Arrhenius rule. Moreover, this work provides an effective and accurate way to explore other molten salts.

Lattice Thermal Conductivity of MgSiO3 Perovskite and Post-Perovskite under Lower Mantle Conditions Calculated by Deep Potential Molecular Dynamics

Lattice thermal conductivity (κ lat) of MgSiO3 perovskite and post-perovskite is an important parameter for the thermal dynamics in the Earth. Here, we develop a deep potential of density functional theory quality under entire thermodynamic conditions in the lower mantle, and calculate the κ lat by the Green–Kubo relation. Deep potential molecular dynamics captures full-order anharmonicity and considers ill-defined phonons in low-κ lat materials ignored in the phonon gas model. The κ lat shows negative temperature dependence and positive linear pressure dependence. Interestingly, the κ lat undergos an increase at the phase boundary from perovskite to post-perovskite. We demonstrate that, along the geotherm, the κ lat increases by 18.2% at the phase boundary. Our results would be helpful for evaluating Earth’s thermal dynamics and improving the Earth model.

Thermal transport across copper-water interface from deep potential molecular dynamics

Thermal transport across copper-water interface from deep potential molecular dynamics

Thermal transport across copper-water interface from deep potential molecular dynamics

Thermal transport across copper-water interface from deep potential molecular dynamics

Molecular dynamics simulations on AlCl3-LiCl molten salt with deep learning potential

AlCl3-LiCl molten salt is a promising candidate used in high-temperature batteries as cathode material to promote the development of renewable energy. Properties of AlCl3-LiCl molten salt are scarce, however, accurate and effective prediction from experienced molecular dynamics and ab initio dynamics remains a challenge. A fast and accurate simulation method based on ab initio datasets and deep neural networks, using machine learning technique, was adopted in this work. A deep potential model was constructed and trained to reproduce the energy and force of AlCl3-LiCl molten salt. Deep potential molecular dynamics simulations were carried out to investigate the local structure and properties using the deep potential model. Structural analysis including partial radial distribution function, coordination number distribution and angular distribution function suggests that the coordinated structure of Cl− around Al3+ is a stabilized and regular tetrahedron, these tetrahedrons form a sparse network liquid structure in mixtures mainly through corner-sharing. Meanwhile, properties like density, thermal expansion coefficient, specific heat capacity, self-diffusion coefficient and shear viscosity were discussed. Property discussion reveals that density and shear viscosity shows a negative relationship with temperature, the diffusivity of each ion species in AlCl3-LiCl molten salt mixture follows the order Li+ > Al3+≈ Cl− and the diffusivity increases with the rising temperature. This work enriches the fundamental data of property for AlCl3-LiCl molten salt and suggests an effective and accurate approach to other molten salt investigations in the future.

Investigation of the Redox Potential of Lithium and Its Dissolution in the LiCl–KCl Eutectic

This present paper reports two fundamental thermodynamic data, including redox potential of Li+ and dissolution rate of metallic lithium in the LiCl–KCl eutectic for promoting the development of metallic lithium. Firstly, data from the first-principles molecular dynamics (FPMD) simulation was used for deep potential (DP) model training. The model’s reliability was then evaluated by the consistency of the energy, force, and radial distribution functions by comparing with FPMD simulation. After that, machine learning-driven molecular dynamics was performed by deep potential molecular dynamics (DPMD) to predict the redox potential of Li+/Li vs Cl2/Cl− in the LiCl–KCl eutectic. Compared with electromotive force measurements results, the maximum relative error is 6.86%. Finally, we investigated the effects of current density and temperature on lithium metal dissolution rate in the LiCl–KCl eutectic by the strategy of current reversal chronopotentiometry, respectively. In the current density range (0.041–0.245 A cm−2), the effect of current density on the dissolution rate of metallic lithium is negligible. As expected, lithium metal prefers to dissolve at high temperatures, and the activation energy of dissolution of lithium in LiCl–KCl eutectic is Ea = 27.78 kJ·mol−1.

Study on the structural properties of refining slags by molecular dynamics with deep learning potential

The CaO-Al2O3 is the most important basis of the multicomponent slag in the steel making industry. The microstructure of molten CaO-Al2O3 systems affects the physical, chemical, and metallurgic properties of refining slags. It is significant to establish a computational method to accurately describe the microstructure of the slag at the atomic level and speculate the properties of large slag models. In this paper, the effect of compositions on the structure of molten CaO-Al2O3 systems was studied by ab initio molecular dynamics, deep learning theory, and deep potential molecular dynamics. The structures of various molten slags were simulated by ab initio molecular dynamics, which accurately describes the interactions between atoms. The potential functions were obtained by deep learning theory. The properties of the large slag models were simulated by molecular dynamics with deep learning potential. As the molar fraction of CaO increases from 0.5 to 0.7, the combination of O and Al mainly forms [AlO4]-5, [AlO3]-3, and [AlO5]-7. Ca presents mainly as free cations in the molten CaO-Al2O3 system. In addition, the diffusion coefficient of Ca decreases from 9.0 × 10-10 m2/s to 5.4 × 10-10 m2/s. The diffusion coefficients of Al and O slightly decrease and are close to 1.1 × 10-10 m2/s and 3.0 × 10-10 m2/s, respectively. It is deduced that the dissolution of CaO provides free O2– and promotes the formation of [AlOn]-b and increases the polymerization degree, which causes deterioration of the fluidity of molten CaO-Al2O3 slags as XCaO increases from 0.5 to 0.7.

Deep learning molecular dynamics simulation on microwave high-temperature dielectric function of silicon nitride

Silicon nitride (<i>β</i>-Si<sub>3</sub>N<sub>4</sub>) is a most promising thermal wave-transparent material. The accurate measurement of its high-temperature dielectric function is essential to solving the “black barrier” problem of hypersonic vehicles and accelerating the design of silicon nitride-based thermal wave-transparent materials. Direct experimental measurement at high temperature is a difficult job and the accuracy of classical molecular dynamics (CMD) simulations suffers the choice of empirical potential. In this work, we build a <i>β</i>-Si<sub>3</sub>N<sub>4</sub> model on a nanoscale, train the deep learning potential (DLP) by using first-principles data, and apply the deep potential molecular dynamics (DPMD) to simulate the polarization relaxation process. The predicted energy and force by DLP are excellently consistent with first-principles calculations, which proves the high accuracy of DLP. The RMSEs for <i>β</i>-Si<sub>3</sub>N<sub>4</sub> are quite low (0.00550 meV/atom for energy and 7.800 meV/Å for force). According to the Cole-Cole formula, the microwave dielectric function in the temperature range of 300–1000 K is calculated by using the deep learning molecular dynamics method. Compared with the empirical potential, the computational results of the DLP are consistent with the experimental results in the sense of order of magnitude. It is also found that the DPMD performs well in terms of computational speed. In addition, a mathematical model of the temperature dependence of the relaxation time is established to reveal the pattern of relaxation time varying with temperature. The high-temperature microwave dielectric function of silicon nitride is calculated by implementing large-scale and high-precision molecular dynamics simulations. It provides fundamental data for promoting the application of silicon nitride in high-temperature thermal transmission.

Molecular dynamics simulation of molten strontium chloride based on deep potential

In this work, a scheme based on ab initio calculations and machine learning is applied to train deep learning interatomic potentials (DP) for molten SrCl2. With the trained DP, we investigate the local structures and properties of molten SrCl2 in the temperature range of 1200–1500 K. The results show that deep potential molecular dynamics (DPMD) simulations can accurately reproduce the radial distribution functions and angle distribution functions of ab initio molecular dynamics simulations. Meanwhile, the simulated partial structure factors of DPMD simulations are in good agreement with those obtained by the thermal neutron scattering technique. Additionally, the predicted density, self-diffusion coefficient, viscosity, and ionic conductivity are satisfactory compared with the experimental data. This study indicates that DP models can serve as an effective tool for the theoretical study of molten salts.

Machine‐learning‐based interatomic potentials for advanced manufacturing

AbstractThis paper summarizes the progress of machine‐learning‐based interatomic potentials and their applications in advanced manufacturing. Interatomic potential is essential for classical molecular dynamics. The advancements made in machine learning (ML) have enabled the development of fast interatomic potential with ab initio accuracy. The accelerated atomic simulation can greatly transform the design principle of manufacturing technology. The most widely used supervised and unsupervised ML methods are summarized and compared. Then, the emerging interatomic models based on ML are discussed: Gaussian approximation potential, spectral neighbor analysis potential, deep potential molecular dynamics, SCHNET, hierarchically interacting particle neural network, and fast learning of atomistic rare events.

Local structure elucidation and properties prediction on KCl–CaCl2 molten salt: A deep potential molecular dynamics study

KCl–CaCl 2 molten salt has emerged as a potential candidate which can be used for heat transfer and thermal energy storage in the next concentrating solar power generation. In this work, deep potential molecular dynamics (DPMD) simulations, based on a many-body potential and interatomic forces generated by a deep neural network trained with the first-principles simulation results, were carried out to predict the local structure and thermophysical properties of KCl–CaCl 2 molten salt at 1100 K. It was found that the steric hindrance effect became more and more intense as the network structure forming and growing with the addition of CaCl 2 , but the increase of CaCl 2 made the coordination shell of Ca 2+ more dynamic and active, and the two reactions constantly compete. Afterwards, thermophysical properties like shear viscosity, heat capacity and thermal conductivity were calculated from DPMD simulations. DPMD simulations could balance the proper description of complex atomic interactions to overcome the challenge of missing potential parameters in classical molecular simulations and become more efficient than first-principles molecular dynamics simulations. Our work based on deep potentials development will provide a promising alternative way to explore more molten eutectic salts for the next concentrating solar power generation. • Deep potential molecular dynamics simulations were carried out to explore the KCl–CaCl 2 molten salt. • Cl − ions more likely to be associated with Ca 2+ ions than K + ions in the KCl–CaCl 2 molten salt. • For KCl–CaCl 2 molten salt, heat capacity decreases with the addition of the CaCl 2 , shear viscosity and thermal conductivity increase slightly with the addition of the CaCl 2 .

Deep neural network potentials for diffusional lithium isotope fractionation in silicate melts

Diffusional isotope fractionation has been widely used to explain lithium (Li) isotope variations in minerals and rocks. Isotopic mass dependence of Li diffusion can be empirically expressed as D7LiD6Li=67β, where D is the diffusivity of a Li isotope. The knowledge about temperature and compositional dependence of the β factor which is essential for understanding diffusion profiles and mechanisms remains unclear. Based on the potential energy and interatomic forces generated by deep neural networks trained with ab initio data, we performed deep potential molecular dynamics (DPMD) simulations of several Li pseudo-isotopes (with mass = 2, 7, 21, 42 g/mol) in albite, hydrous albite, and model basalt melts to evaluate the β factor. Our calculated diffusivities for 7Li in albite and model basalt melts at 1800 K compare well with experimental results. We found that β in albite melt decreases from 0.267±0.006 at 4000 K to 0.225±0.004 at 1800 K. The presence of water appears to slightly weaken the temperature dependence of β, with β decreasing from 0.250±0.012 to 0.228±0.031 in hydrous albite melt. The calculated β in model basalt melt takes much smaller values, decreasing from 0.215±0.006 at 4000 K to 0.132±0.015 at 1800 K. Our prediction of β in albite and hydrous albite melts is in good agreement with experimental data. More importantly, our results suggest that Li isotope diffusion in silicate melts is strongly dependent on melt composition. The temperature and compositional effects on β can be qualitatively explained in terms of ionic porosity and the coupled relationship between Li diffusion and the mobility of the silicate melt network. Two types of diffusion experiments are suggested to test our predicted temperature and compositional dependence of β. This study shows that DPMD is a promising tool to simulate the diffusion of elements and isotopes in silicate melts.

Unravelling the Fast Alkali-Ion Dynamics in Paramagnetic Battery Materials Combined with NMR and Deep-Potential Molecular Dynamics Simulation.

Solid-state nuclear magnetic resonance (ssNMR) has received extensive attention in characterizing alkali-ion battery materials because it is highly sensitive for probing the local environment and dynamic information of atoms/ions. However, precise spectral assignment cannot be carried out by conventional DFT for high-rate battery materials at room temperature. Herein, combining DFT calculation of paramagnetic shift and deep potential molecular dynamics (DPMD) simulation to achieve the converged Na+ distribution at hundreds of nanoseconds, we obtain the statistically averaged paramagnetic shift, which is in excellent agreement with ssNMR measurements. Two 23 Na shifts induced by different stacking sequences of transition metal layers are revealed in the fast chemically exchanged NMR spectra of P2-type Na2/3 (Mg1/3 Mn2/3 )O2 for the first time. This DPMD simulation auxiliary protocol can be beneficial to a wide range of ssNMR analysis in fast chemically exchanged material systems.

Unravelling the Fast Alkali‐Ion Dynamics in Paramagnetic Battery Materials Combined with NMR and Deep‐Potential Molecular Dynamics Simulation

AbstractSolid‐state nuclear magnetic resonance (ssNMR) has received extensive attention in characterizing alkali‐ion battery materials because it is highly sensitive for probing the local environment and dynamic information of atoms/ions. However, precise spectral assignment cannot be carried out by conventional DFT for high‐rate battery materials at room temperature. Herein, combining DFT calculation of paramagnetic shift and deep potential molecular dynamics (DPMD) simulation to achieve the converged Na+ distribution at hundreds of nanoseconds, we obtain the statistically averaged paramagnetic shift, which is in excellent agreement with ssNMR measurements. Two 23Na shifts induced by different stacking sequences of transition metal layers are revealed in the fast chemically exchanged NMR spectra of P2‐type Na2/3(Mg1/3Mn2/3)O2 for the first time. This DPMD simulation auxiliary protocol can be beneficial to a wide range of ssNMR analysis in fast chemically exchanged material systems.

Theoretical study of Na+ transport in the solid-state electrolyte Na3OBr based on deep potential molecular dynamics

Deep potential molecular dynamics is used to study Na+ transport in Na3OBr.

Isotope effects in molecular structures and electronic properties of liquid water via deep potential molecular dynamics based on the SCAN functional

Feynman path-integral deep potential molecular dynamics (PI-DPMD) calculations have been employed to study both light (${\mathrm{H}}_{2}\mathrm{O}$) and heavy water (${\mathrm{D}}_{2}\mathrm{O}$) within the isothermalisobaric ensemble. In particular, the deep neural network is trained based on ab initio data obtained from the strongly constrained and appropriately normed (SCAN) exchange-correlation functional. Because of the lighter mass of hydrogen than deuteron, the properties of light water are more influenced by nuclear quantum effect than those of heavy water. Clear isotope effects are observed and analyzed in terms of hydrogen-bond structure and electronic properties of water that are closely associated with experimental observables. The molecular structures of both liquid ${\mathrm{H}}_{2}\mathrm{O}$ and ${\mathrm{D}}_{2}\mathrm{O}$ agree well with the data extracted from scattering experiments. The delicate isotope effects on radial distribution functions and angular distribution functions are well reproduced as well. Our approach demonstrates that deep neural network combined with SCAN functional based ab initio molecular dynamics provides an accurate theoretical tool for modeling water and its isotope effects.

Warm dense matter simulation via electron temperature dependent deep potential molecular dynamics

Simulating warm dense matter that undergoes a wide range of temperatures and densities is challenging. Predictive theoretical models, such as quantum-mechanics-based first-principles molecular dynamics (FPMD), require a huge amount of computational resources. Herein, we propose a deep learning based scheme called electron temperature dependent deep potential molecular dynamics (TDDPMD), which can be readily applied to study larger systems with longer trajectories, yielding more accurate properties. We take warm dense beryllium (Be) as an example with the training data from FPMD simulations spanning a wide range of temperatures (0.4–2500 eV) and densities (3.50–8.25 g/cm3). The TDDPMD method well reproduces the principal Hugoniot curve and radial distribution functions from the FPMD method. Furthermore, it depicts the reflection point of the Hugoniot curve more smoothly and provides more converged diffusion coefficients. We also show the new model can yield static structure factors and dynamic structure factors of warm dense Be.