Deep Potential Molecular Dynamics

The core neural network architecture that decomposes the total potential energy into atomic contributions based on local environment descriptors.

• Constructs translationally, rotationally, and permutationally invariant descriptors from relative atomic positions
• Uses a smooth cutoff function to ensure forces and virial decay continuously to zero at the cutoff radius
• Employs a sub-network (embedding net) to map local environments to atomic energies, then sums for total energy
• Supports both isolated systems and periodic boundary conditions

Example: A water simulation with 128 molecules using DeepPot-SE achieves ~1 meV/atom accuracy relative to SCAN DFT while running 5 orders of magnitude faster.

DeePMD-kit learns directly from labeled datasets of atomic configurations with energies, forces, and virial tensors computed by DFT or other electronic structure methods.

• Supports VASP, CP2K, Quantum ESPRESSO, ABACUS, PWmat, SIESTA, Gaussian output formats via dpdata
• Loss function jointly optimizes energy, force, and virial predictions with tunable prefactors
• Implements the Adam stochastic gradient descent optimizer with learning rate decay
• Training data is typically generated via concurrent learning (Deep Potential Generator) to iteratively explore undersampled regions of configuration space

Workflow: Run short AIMD → Train initial DP model → Run DP-MD → Identify extrapolation regions via model deviation → Label new configurations with DFT → Retrain.

DeePMD-kit provides model compression tools to reduce inference latency for large-scale production simulations without sacrificing accuracy.

• Tabulated embedding: Pre-computes the embedding net output on a fine grid and stores it in lookup tables, eliminating neural network evaluation for the descriptor
• Mixed precision (FP16/FP32): Reduces memory footprint and accelerates tensor operations on modern GPUs
• Model pruning: Removes redundant neurons in the fitting net with minimal accuracy loss
• Compressed models achieve 2–5× inference speedup while maintaining meV-level energy fidelity

Benchmark: A compressed DeepPot-SE model for a 10,000-atom copper system runs at 2.3 ns/day on a single NVIDIA A100, compared to 0.8 ns/day for the uncompressed model.

DeePMD-kit has been used to construct a neural network potential for water that accurately reproduces the complex phase diagram, including the anomalous density maximum and various ice polymorphs. The model, trained on Density Functional Theory (DFT) data using the SCAN functional, captures nuclear quantum effects and predicts the melting lines of ice Ih, II, and III with near-experimental accuracy. This demonstrates the ability of deep potentials to model hydrogen-bonded systems where subtle quantum mechanical effects govern macroscopic behavior.

Deep potentials enable the study of warm dense matter—a regime where matter is partially ionized and neither classical nor degenerate quantum theories suffice. Researchers have trained DeePMD models on Kohn-Sham DFT data to simulate carbon, silicon, and water at extreme pressures (up to terapascals) and temperatures (up to 10^5 K). These simulations inform models of planetary interiors, including the structure of ice giants like Neptune and Uranus, where superionic water phases exist.

The search for next-generation solid-state electrolytes for lithium-ion batteries requires accurate modeling of Li-ion diffusion pathways. DeePMD potentials trained on DFT data for systems like LLZO (Li₇La₃Zr₂O₁₂) and LGPS (Li₁₀GeP₂S₁₂) capture the correlated motion of lithium ions and lattice vibrations. These simulations reveal how dopants and defects modulate ionic conductivity, guiding experimental synthesis toward materials with higher performance and stability.

Understanding radiation damage cascades in nuclear fuel materials like UO₂ requires simulating thousands of atoms over picosecond timescales with quantum accuracy. DeePMD models trained on DFT+U data capture the complex charge-transfer and strongly correlated electron physics of actinide oxides. These simulations predict primary knock-on atom displacement thresholds and defect recombination rates, informing fuel performance codes used in reactor safety analysis.

Deep potentials have been applied to model catalytic reactions on transition metal surfaces, such as CO oxidation on Pt and water-gas shift on Cu. By training on DFT data that includes adsorbate-surface interactions, DeePMD captures the energy barriers and reaction pathways with ab initio fidelity. This enables the simulation of temperature-programmed desorption and reaction kinetics at realistic coverages, bridging the gap between ultra-high vacuum surface science and industrial catalytic conditions.

The mechanical and thermal properties of 2D materials like graphene, MoS₂, and hexagonal boron nitride are governed by long-wavelength phonons and rippling modes. DeePMD models trained on van der Waals-corrected DFT capture these subtle interlayer interactions. Simulations of twisted bilayer graphene and Moiré superlattices reveal how stacking angle influences thermal conductivity and phonon localization, informing the design of nanoelectronic and thermoelectric devices.