Deep Potential Molecular Dynamics (DeePMD) is a scalable molecular simulation method that uses a deep neural network to generate accurate interatomic potentials and forces based on ab initio data. The neural network preserves natural symmetries and is "first-principle-based," with no ad hoc components aside from the network. DeePMD provides efficient and accurate results for various systems, including bulk materials and molecules, with computational cost scaling linearly with system size. Molecular dynamics (MD) simulations are widely used but depend on the accuracy of atomic interaction models. Ab initio MD (AIMD) offers high accuracy but is computationally expensive, limiting its use to small systems. Empirical force fields (FFs) are used for larger systems but lack accuracy and transferability. Machine learning (ML) methods are improving this state by accurately reproducing atomic configurations and forces when trained on large datasets. DeePMD addresses the limitations of symmetry functions and Coulomb matrices by assigning a local reference frame and environment to each atom. The neural network input is derived from local coordinates, preserving translational, rotational, and permutational symmetries. The network is trained with a flexible loss function, enabling accurate reproduction of AIMD trajectories, both classical and quantum, at a cost much lower than AIMD. DeePMD's potential energy is a sum of "atomic energies," where each energy depends on the local environment. The neural network captures the complex, nonlinear dependence of these energies on atomic coordinates. The additive form of the potential energy naturally preserves its extensive character, and the model is conservative in principle. The method is tested on extended and finite systems, including liquid water, ice, and organic molecules. DeePMD reproduces MD trajectories with high accuracy, showing excellent agreement with AIMD results. The method is scalable, with computational cost linear in system size, and can be easily parallelized. DeePMD is applicable to a wide range of systems, including biological molecules, alloys, and liquid mixtures. It offers a new paradigm for molecular simulations, enabling accurate quantum mechanical data to be parametrized by ML algorithms, allowing simulations of AIMD quality on much larger systems and for longer times than direct AIMD. While more predictive than empirical FFs, DFT is not chemically accurate. DeePMD could be trained with chemically accurate data from high-level quantum chemistry or quantum Monte Carlo, though this is currently limited by computational costs. DeePMD is also useful for coarse-graining atomic degrees of freedom, using a reduced set for training while maintaining a full set for simulation. The method addresses the long-standing dilemma of accuracy versus efficiency in molecular simulations, enhancing the application of AIMD.Deep Potential Molecular Dynamics (DeePMD) is a scalable molecular simulation method that uses a deep neural network to generate accurate interatomic potentials and forces based on ab initio data. The neural network preserves natural symmetries and is "first-principle-based," with no ad hoc components aside from the network. DeePMD provides efficient and accurate results for various systems, including bulk materials and molecules, with computational cost scaling linearly with system size. Molecular dynamics (MD) simulations are widely used but depend on the accuracy of atomic interaction models. Ab initio MD (AIMD) offers high accuracy but is computationally expensive, limiting its use to small systems. Empirical force fields (FFs) are used for larger systems but lack accuracy and transferability. Machine learning (ML) methods are improving this state by accurately reproducing atomic configurations and forces when trained on large datasets. DeePMD addresses the limitations of symmetry functions and Coulomb matrices by assigning a local reference frame and environment to each atom. The neural network input is derived from local coordinates, preserving translational, rotational, and permutational symmetries. The network is trained with a flexible loss function, enabling accurate reproduction of AIMD trajectories, both classical and quantum, at a cost much lower than AIMD. DeePMD's potential energy is a sum of "atomic energies," where each energy depends on the local environment. The neural network captures the complex, nonlinear dependence of these energies on atomic coordinates. The additive form of the potential energy naturally preserves its extensive character, and the model is conservative in principle. The method is tested on extended and finite systems, including liquid water, ice, and organic molecules. DeePMD reproduces MD trajectories with high accuracy, showing excellent agreement with AIMD results. The method is scalable, with computational cost linear in system size, and can be easily parallelized. DeePMD is applicable to a wide range of systems, including biological molecules, alloys, and liquid mixtures. It offers a new paradigm for molecular simulations, enabling accurate quantum mechanical data to be parametrized by ML algorithms, allowing simulations of AIMD quality on much larger systems and for longer times than direct AIMD. While more predictive than empirical FFs, DFT is not chemically accurate. DeePMD could be trained with chemically accurate data from high-level quantum chemistry or quantum Monte Carlo, though this is currently limited by computational costs. DeePMD is also useful for coarse-graining atomic degrees of freedom, using a reduced set for training while maintaining a full set for simulation. The method addresses the long-standing dilemma of accuracy versus efficiency in molecular simulations, enhancing the application of AIMD.