Research map / Probabilistic and scientific modeling / Dynamical systems

FlashMD: long-stride, universal prediction of molecular dynamics

NeurIPS 2025 · Probabilistic and scientific modeling › Dynamical systems · approach: Stochastic population dynamics · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FlashMD, a PET-based GNN that directly predicts mass-scaled positions and momenta over strides 1-2 orders of magnitude longer than MD steps, with energy-conservation rescaling, ensemble control, and UQ, trained on water-specific and universal (MAD/PET-MAD) trajectories.

Problem

Machine learning MD acceleration is still limited to tiny integration steps (~1 fs) because the integrator must be run step by step; direct prediction of long-stride MD trajectories is underexplored and prone to failure modes (energy drift, non-symplecticity, chaos, symmetry breaking).

Approach

Directly predict time-evolved mass-scaled positions and momenta over long strides with an E(3)-unconstrained PET graph network, adding momentum-based node inputs, training-time energy-error model selection, inference-time momentum rescaling for energy conservation, random-rotation averaging, and a Liouville-based recipe for combining with thermostats/barostats.

Key results

• Energy-conservation rescaling recovers correct mean temperatures for overall water NVT simulations, while uncorrected models show deviations of tens of K in per-atom-type kinetic energies with SVR thermostat.
• Universal FlashMD reproduces alanine dipeptide Ramachandran basins with strides up to 16 fs (32 times the 0.5 fs MD step).
• The universal Li3PS4 model predicts the superionic transition at about 675 K, within the range reported for PET-MAD MD.

Measured effects reported in the paper

Surprising findings

• Models with similar position and momentum accuracy can yield very different energy conservation, so energy error must be included in model selection.
• Local (Langevin) thermostats were needed because a global SVR thermostat gave spurious per-atom kinetic energy deviations (lack of equipartition), also seen in direct-force models.

Limitations

• Non-symplectic and non-energy-conserving dynamics; energy drift needs inference-time rescaling and thermostats.
• Predictions degrade rapidly with stride; symplecticity violations grow strongly with stride.
• Dynamical properties are distorted by the local thermostats needed for correct equilibrium sampling.

Experimental setup

Datasets: MAD dataset liquid water MD trajectories generated with PET-MAD q-TIP4P/f water trajectories

Benchmarks: liquid water NVT effective temperature liquid water radial distribution functions liquid water density at NPT alanine dipeptide Ramachandran plots Al(110) surface mean square displacement gamma-Li3PS4 Li conductivity

Metrics: effective temperature deviation radial distribution function density mean square displacement ionic conductivity validation RMSE energy conservation symplecticity error

Baselines: PET-MAD MDnet TrajCast conventional velocity Verlet MD

Builds on: Point-Edge Transformer PET-MAD velocity Verlet LLPR MDnet

Closest papers by meaning

• UniSim: A Unified Simulator for Time-Coarsened Dynamics of Biomolecules (ICML 2025 · similarity 0.88)
• The dark side of the forces: assessing non-conservative force models for atomistic machine learning (ICML 2025 · similarity 0.88)
• Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials (NeurIPS 2026 · similarity 0.88)
• Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.87)
• Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (ICLR 2025 · similarity 0.87)
• FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials (ICML 2025 · similarity 0.87)