molecular dynamics simulation

Accurate prediction of energy and forces for 3D molecular systems is one of fundamental challenges at the core of AI for Science applications. Many powerful and data-efficient neural networks predict molecular energies and forces from single atomic configurations. However, one crucial aspect of the data generation process is rarely considered while learning these models i.e. Molecular Dynamics (MD) simulation. MD simulations generate time-ordered trajectories of atomic positions that fluctuate in energy and explore regions of the potential energy surface (e.g., under standard NVE/NVT ensembles), rather than being constructed to steadily lower the potential energy toward a minimum as in geometry relaxations. This work explores a novel way to leverage MD data, when available, to improve the performance of such predictors. We introduce a novel training strategy called FRAMES, that use an auxiliary loss function for exploiting the temporal relationships within MD trajectories. Counter-intuitively, on two atomistic benchmarks and a synthetic system we observe that minimal temporal information, captured by pairs of just two consecutive frames, is often sufficient to obtain the best performance, while adding longer trajectory sequences can introduce redundancy and degrade performance. On the widely used MD17 and ISO17 benchmarks, FRAMES significantly outperforms its Equiformer baseline, achieving highly competitive results in both energy and force accuracy. Our work not only presents a novel training strategy which improves the accuracy of the model, but also provides evidence that for distilling physical priors of atomic systems, more temporal data is not always better.

To address energy drift and violation of equipartition among degrees of freedom caused by large time steps in molecular dynamics (MD) simulations, this work proposes a variational-principles-based machine learning integrator. The method employs neural networks to directly learn the system’s action functional, thereby constructing a data-driven, symplectic and time-reversible mapping that intrinsically satisfies Hamiltonian constraints. Unlike black-box state predictors, our framework formulates numerical integration as a differentiable action minimization process and incorporates an iterative correction mechanism to enhance long-term stability. Extensive validation across multiple molecular systems demonstrates that the integrator enables time steps 5–10× larger than conventional Verlet integration, reduces energy error by one to two orders of magnitude, and perfectly preserves energy equipartition across all degrees of freedom. Consequently, it significantly improves both accuracy and efficiency for nanosecond- to microsecond-scale MD simulations.

Parallel computation of three-body interactions—where three molecules from distinct spatial cells participate simultaneously—poses a significant challenge in molecular dynamics simulations. Method: This paper introduces the first linked-cell framework supporting traversal across three distinct cells. It extends conventional pairwise neighbor searching to triplet-based neighborhood construction and integrates geometric culling with distance-based pruning to drastically reduce redundant computations. The approach is validated using the Lennard-Jones fluid model under both uniform and non-uniform density conditions. Contribution/Results: The algorithm achieves strong node-level scalability, delivering a substantial increase in molecular updates per second over baseline methods. It is the first systematic solution to the efficiency bottleneck inherent in parallel traversal of three-cell coupled neighborhoods for short-range many-body forces. By enabling scalable, high-fidelity simulation of many-body potentials, this work provides a foundational algorithmic infrastructure for next-generation, high-accuracy molecular dynamics.

To address the high computational cost and memory footprint of machine-learned interatomic potentials (MLIPs) in large-scale molecular dynamics (MD) simulations, this work introduces knowledge distillation to MLIP training for the first time, proposing a teacher–student collaborative framework. A high-accuracy teacher model provides implicit supervision via atomic energy predictions, guiding the training of a lightweight student model with a customized, resource-efficient architecture. Remarkably, the student achieves superior accuracy to the teacher under identical training data. Experiments on benchmarks including QM9 demonstrate a 12% reduction in mean absolute error (MAE), a 2.3× speedup in MD simulation throughput, and a fivefold reduction in memory consumption relative to the teacher. The core contribution lies in adapting the knowledge distillation paradigm to the MLIP domain, enabling simultaneous optimization of predictive accuracy, computational efficiency, and memory efficiency.

This work identifies a fundamental flaw in atomic-scale machine learning: directly modeling non-conservative forces inherently violates energy conservation, leading to geometric optimization failure and numerical instability in molecular dynamics (MD). Through systematic experiments—including atomic neural networks, multi-algorithm geometric optimization, long-timescale MD simulations, and energy conservation diagnostics—we empirically demonstrate, for the first time, that non-conservative force models fail to satisfy physical consistency requirements. To address this, we propose a hybrid “conservative energy + direct force” paradigm. This approach preserves predictive accuracy while substantially improving stability: MD trajectories remain bounded without divergence, geometric optimization converges reliably, and backpropagation computational cost decreases by ~40%. Our framework establishes a new force-field modeling paradigm that simultaneously ensures physical interpretability—via explicit energy conservation—and computational efficiency.

This work proposes DINaMo, a novel framework that introduces the first fully unsupervised neural trajectory solver for molecular dynamics. Unlike conventional neural approaches that rely on simulated trajectories, forces, or energies for supervised training and thus remain data-dependent, DINaMo models molecular trajectories as differentiable functions of time and trains exclusively through physical principles—namely Newton’s equations of motion, conservation laws, and analytically defined interaction potentials—without any reference simulation data. The method successfully reproduces key structural and dynamical observables in a Lennard-Jones argon system, including short-time coordinate evolution, energy conservation, and the radial distribution function of the liquid phase, thereby demonstrating the feasibility of learning molecular motion solely from physical constraints.

This study addresses the computational inefficiency bottleneck in long-timescale molecular dynamics simulations by proposing a machine learning force field framework based on Langevin flow mapping. The core innovation lies in incorporating a stochastic Langevin integrator into a machine learning model for the first time, fusing stochastic differential equations with deep learning techniques to enable direct learning of the stochastic integration process, thereby overcoming conventional small-timestep limitations. While faithfully preserving system dynamical properties and maintaining strong transferability, this approach achieves efficient large-timestep simulations, accelerating computation by an order of magnitude compared to existing methods.

This study investigates whether AI agents can autonomously design and execute molecular dynamics (MD) simulation workflows, confronting challenges such as modeling physical intuition, reasoning about boundary conditions, diagnosing numerical stability, and interpreting physical outcomes. To this end, the authors introduce MDGYM, the first structured benchmark for evaluating MD agents, encompassing both LAMMPS and GROMACS—the two dominant MD software packages—and comprising 169 expert-curated tasks categorized into three difficulty levels. The evaluation employs an end-to-end framework integrating Claude Code, Codex, OpenHands, and four large language models. Results reveal that even the strongest agent solves only 21% of the easiest tasks, with success rates dropping below 10% on harder tasks. Common failure modes include generating physically unstable configurations, fabricating outputs, and premature termination, highlighting significant limitations of current large models in embodied physical reasoning.

This work addresses a critical limitation of machine learning interatomic potentials (MLIPs)—their frequent failure to reproduce the physical smoothness of quantum mechanical potential energy surfaces, which can lead to unphysical behaviors in molecular dynamics simulations that conventional energy/force regression metrics fail to detect. To tackle this issue, the authors propose the Bond Smoothness Characterization Test (BSCT), a computationally efficient probe that systematically evaluates potential energy surface smoothness through controlled bond deformations. For the first time, BSCT is integrated as a closed-loop feedback tool to guide MLIP architecture optimization. By incorporating differentiable k-nearest neighbors and temperature-controlled attention mechanisms into a Transformer-based MLIP, and co-optimizing it with BSCT, the resulting model achieves low regression errors while significantly enhancing simulation stability and the reliability of atomic-scale property predictions, outperforming traditional, costly, and limited dynamic evaluation approaches.

This work proposes DMTS-NC, a method to accelerate high-accuracy neural network potential-driven molecular dynamics simulations by integrating multiple timesteps with non-conservative forces, achieving substantial computational speedup without fine-tuning. The approach incorporates physical priors—such as embedded rotational equivariance and atomic force component cancellation—into a distillation architecture, enhancing numerical stability and mitigating potential energy “holes.” A two-level reversible RESPA integration scheme couples a high-fidelity conservative potential with a lightweight distilled model, making the framework compatible with arbitrary neural network potentials. Compared to conventional conservative DMTS, DMTS-NC delivers 15–30% faster simulations while preserving accuracy near the system’s physical resonance limit and demonstrating superior long-term stability.