molecular dynamics simulation

Rare-event conformational transitions—such as ion channel gating—in molecular systems exhibit high energy barriers, rendering them poorly sampled by conventional molecular dynamics (MD) or Markov chain Monte Carlo (MCMC) methods. To address this, we propose an end-to-end, physics-informed neural network (PINN)-based continuous optimization framework for computing minimum energy paths (MEPs), the first to integrate PINNs with the string method philosophy. Our approach eliminates the need for an initial path guess and avoids explicit path sampling and reparameterization. Leveraging automatic differentiation, differentiable force fields, and an implicit neural representation of the reaction path, it scales to large, explicit-solvent biomolecular systems (>8,300 atoms). Applied to proteins including BPTI, the computed MEPs are physically realistic and quantitatively accurate. Computational efficiency exceeds that of conventional string methods by over one order of magnitude, establishing a new, scalable paradigm for discovering transition pathways in complex biomolecules.

Traditional reaction path visualization methods—e.g., energy–displacement plots—project high-dimensional structural evolution onto a one-dimensional trajectory, obscuring configurational details and impeding cross-algorithm comparison of optimization dynamics. To address this, we propose a two-dimensional projection framework based on permutation-corrected RMSD: the configuration plane is defined by RMSD distances from reactant and product structures; radial basis function interpolation and energy-based coloring yield a continuous, interpretable potential energy surface. This enables simultaneous trajectory visualization, convergence diagnostics, and precise saddle-point localization, facilitating intuitive comparative analysis across optimization algorithms. Validation on cycloaddition reactions shows that machine-learned interatomic potentials predict saddle points with minor geometric deviations, yet these lie within energy contour levels closely matching DFT reference results—demonstrating strong configurational–energetic consistency.

Extracting reaction mechanisms from trajectory data and efficiently sampling rare events remain challenging tasks. This work proposes a data-driven flow-matching framework that requires no prior knowledge of the system’s dynamics or stationary distribution. By minimizing a quadratic functional, the method jointly learns a vector flow field and a scalar potential to construct reactive pathways and reaction coordinates. Built upon a weighted Helmholtz–Hodge decomposition and reactive flow analysis, the approach remains well-defined even under non-Markovian collective variables and enables adaptive sampling. Numerical experiments demonstrate that the method generates accurate molecular flow trajectories and reliably computes reaction rate constants.

Traditional discrete path-searching methods (e.g., Nudged Elastic Band, NEB) suffer from failure under poor initial guesses, competing reaction pathways, or complex multi-step mechanisms. To address these limitations, this work proposes a continuous minimum energy path (MEP) representation framework based on implicit neural networks—the first application of implicit neural representations to reaction path modeling. The method integrates an orthogonal gradient constraint loss, adaptive path sampling, and low-dimensional transfer learning. It enables real-time transition-state estimation, escape from local minima, and joint modeling of multiple competing pathways. Experimental results on atomic-scale systems demonstrate substantial improvements over NEB: superior robustness to suboptimal initial guesses, accurate resolution of multi-step reaction mechanisms, and cross-system generalizability—i.e., a single trained model effectively represents MEPs across diverse chemical systems.

Transition path sampling (TPS) in atomic systems suffers from low sampling efficiency and poor generalizability across diverse molecular systems. Method: We propose a zero-shot TPS paradigm leveraging pretrained generative models—specifically, coupling the score functions of denoising diffusion and flow-matching models with stochastic dynamics, thereby formulating high-likelihood path search as an Onsager–Machlup action minimization problem. Contribution/Results: To our knowledge, this is the first approach enabling cross-system zero-shot transfer without fine-tuning or labeled data, overcoming the longstanding bottleneck of task-specific training in conventional TPS. Validated on multiple molecular systems, the generated paths exhibit physical plausibility, high diversity, and significantly improved sampling efficiency. Moreover, the framework is modular and seamlessly integrates with emerging large-scale generative models. This work establishes a general, efficient, and scalable tool for atomic-scale dynamical pathway analysis.

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.

This work addresses the trade-off between physical realism and computational efficiency in molecular conformation generation. Traditional approaches rely on expensive ab initio calculations, while diffusion models are constrained by explicit timestep conditioning. The authors propose Generative Pseudo Force Fields (GPFF), which construct a quadratic pseudo potential energy surface around reference equilibrium geometries, enabling online generation of non-equilibrium training data without ab initio evaluations for perturbed conformations. By integrating the diffusion process with machine-learned force fields, GPFF implicitly encodes noise levels, thereby eliminating explicit timestep dependencies. The framework supports both standard and adaptive sampling and naturally incorporates structural priors and geometric constraints. On QM9, it achieves over 50% validity with only six neural function evaluations and reaches 100% validity at 256 evaluations, demonstrating high-precision real-time generation in drug molecule editing applications.

This work proposes FES-FM, a novel method for efficiently sampling free energy surfaces (FES) to elucidate chemical reaction and conformational transition mechanisms. By introducing reduced flow matching into FES sampling for the first time, the approach constructs a dynamic transport map directly in collective variable (CV) space, enabling efficient and direct sampling of the FES while circumventing costly simulations in high-dimensional configuration space. The method incorporates a prior distribution derived from the potential energy Hessian, enhancing physical plausibility while preserving rotational and translational invariance, and leverages CV projection techniques to improve sampling accuracy. Benchmarking across diverse potential functions and collective variables demonstrates that FES-FM substantially reduces computational cost and achieves superior sampling precision per unit time compared to conventional approaches.

This study addresses significant performance bottlenecks in machine learning force fields (MLFFs) when executing molecular dynamics simulations on GPUs, primarily caused by irregular memory access patterns, low data reuse, and inefficient kernel execution during descriptor computation and neural network inference. For the first time, the work provides a systematic architectural-level analysis of MLFF workloads and introduces a scalability evaluation framework tailored for drug discovery. It establishes a new benchmark based on controllable-scale polyacrylic acid chains and conducts fine-grained GPU performance profiling using hardware-aware analysis tools. The findings uncover key performance-limiting factors and offer critical insights to guide hardware-algorithm co-optimization strategies for accelerating MLFF-based simulations.

Current coarse-grained molecular dynamics neural potentials rely solely on force matching and lack constraints on the second-order curvature of the free energy landscape, limiting their accuracy and transferability. This work proposes an efficient curvature-aware supervision method that incorporates second-order information into the potential without explicitly constructing the full Hessian matrix, by matching stochastic Hessian-vector products. The approach innovatively decomposes the target coarse-grained Hessian into a projected all-atom Hessian and an online covariance correction term, and introduces an unbiased stochastic estimator to enable stable training. Evaluation on nine unseen fast-folding proteins shows that the method outperforms force-matching-only baselines in capturing slow dynamical modes for eight systems, with up to an 85% reduction in KL divergence along the slowest collective variable for the largest system.