Molecular dynamics

Liquid densities

Summary

Performance in predicting densities for 61 organic liquids, each system consisting of about 1000 atoms. The dataset covers aliphatic, aromatic molecules, as well as different functional groups and halogenated molecules.

Metrics

  1. Density error

For each system, the density is calculated by taking the average density of an NPT molecular dynamics run. The initial part of the simulation, here 500 ps, is omitted from the density calculation. This is compared to the reference density, obtained from experiment.

Benchmark speed

Multi-day: tests take several days to run on GPU.

Data availability

Input structures:

  • Weber et al., Efficient Long-Range Machine Learning Force Fields for

    Liquid and Materials Properties. arXiv:2505.06462 [physics.chem-ph]

Reference data:

  • Same as input data

  • Experimental

Polymer densities

Summary

Performance in predicting room-temperature amorphous densities for 130 polymers. Starting structures are prebuilt polymer cells, and the reference data are experimental densities.

Metrics

  1. Density error

For each polymer, the density is calculated by averaging over the final 500 ps NPT production stage of a 24-stage Polymatic-style equilibration protocol. The protocol uses a 0.5 fs timestep and runs for about 2.06 ns per model and polymer at the default time prefactor. The predicted density is compared to the experimental reference density.

Benchmark speed

Multi-day: tests are likely to take several days to run on GPU.

Data availability

Input structures:

  • Built from the SimPoly polymer dataset and stored as prebuilt structures for ML-PEG calculations.

Reference data:

  • Simm et al., SimPoly: Simulation of Polymers with Machine Learning Force Fields Derived from First Principles. arXiv:2510.13696 [physics.comp-ph]

  • Experimental

Water density

Summary

Performance in predicting the density of water at temperatures of 270, 290, 300, and 330 K. The water systems consist of 333 molecules.

Metrics

  1. Density error

For each system, the density is calculated by taking the average density of an NPT molecular dynamics run. The initial part of the simulation, here 500 ps, is omitted from the density calculation. This is compared to the reference density, obtained from experiment.

Benchmark speed

Multi-day: tests take several days to run on GPU.

Data availability

Input structures:

  • Weber et al., Efficient Long-Range Machine Learning Force Fields for Liquid and Materials Properties. arXiv:2505.06462 [physics.chem-ph]

Reference data:

  • Same as input data

  • Experimental

Water ethanol density curves

Summary

Benchmark of the density of water-ethanol mixtures for different concentrations of ethanol, compare to experiment. 1 ns of NPT MD on about 120 water/ethanol molecules for 6 concentrations.

Metrics

  1. rms of the density difference.

  2. rms of the excess volume difference.

  3. Concentration of the minimal excess volume.

For each system, the density is calculated by taking the average density of an NPT molecular dynamics run. The initial part of the simulation, here 500 ps, is omitted from the density calculation. This is compared to the reference density, obtained from experiment. The excess volume is computed as the difference between the actual molar volume of the mixture and the ideal molar volume obtained by linear combination of the pure-component molar volumes. The concentration of the minimal excess volume is estimated by fitting a quadratic to the three grid points surrounding the minimum and taking the vertex of the parabola.

Benchmark speed

Multi-day: tests take several days to run on GPU.

Data availability

Input structures: Packmol generated

Reference data: * M. Southard and D. Green, Perry’s Chemical Engineers’ Handbook, 9th Edition. McGraw-Hill Education, 2018. * Experimental

Bond length distribution

Summary

Performance in maintaining physically reasonable covalent bond lengths during molecular dynamics of small organic molecules. For each of a set of molecules covering the C-C, C=C, C#C, C-N, C-O, C=O and C-F bond types, an NVT molecular dynamics simulation is run at 300 K starting from a QM-optimised reference geometry, and the deviation of a tracked bond from its reference length is measured along the trajectory.

Metrics

  1. Bond length deviation

The length of the tracked bond is measured at each frame of the trajectory, and its absolute deviation from the reference bond length is averaged over the trajectory and across all molecules. A well behaved potential keeps bonds close to their reference length, so a lower deviation is better.

A histogram shows the distribution of the sampled bond length deviations for each model.

Benchmark speed

Slow: 8 molecules of 4-13 atoms, one MD simulation each, of 1,000,000 steps, i.e. 1 ns at a 1 fs timestep. The molecules are small, so the cost per step is dominated by per-call overhead rather than by system size, and tests are likely to take a couple of hours per model on GPU. Faster inference can be achieved using the jax-accelerated simulations in MLIP Audit directly.

Data availability

Input structures:

  • MLIP Audit benchmark suite, InstaDeep. Reference geometries selected from the QM9 dataset (Ramakrishnan et al., Scientific Data 1, 140022, 2014).

Reference data:

  • QM-optimised equilibrium bond lengths of the reference geometries.

Ring planarity

Summary

Performance in maintaining planar aromatic rings during molecular dynamics simulations of small organic molecules, evaluated on 6 molecules with aromatic ring systems selected from the QM9 dataset. For each molecule, an NVT molecular dynamics simulation is run at 300 K starting from the QM9-optimised geometry, and the deviation of the ring atoms from their best-fit plane is measured along the trajectory.

Metrics

  1. Planarity deviation

At each frame of the trajectory, the ring atoms are fitted to a plane and the root mean square deviation of the atoms from that plane is calculated. This is averaged over the trajectory and across all molecules. Aromatic rings are planar, so a well-behaved potential keeps this deviation small; a lower deviation is better. If any molecule is unstable (fails with an error or the simulation explodes), the metric is reported as NaN, as is the benchmark score.

  1. Planarity deviation (stable only)

The planarity deviation, averaged only over the molecules whose simulations were stable. This is identical to the metric above unless any molecule was unstable, in which case it still gives a value while the metric above is NaN. It is reported for information only and does not contribute to the benchmark score.

A histogram shows the distribution of the sampled planarity deviations for each model, using only stable simulations.

Benchmark speed

Slow: 6 molecules of 12-16 atoms, one MD simulation each, of 1,000,000 steps, i.e. 1 ns at a 1 fs timestep. The molecules are small, so the cost per step is dominated by per-call overhead rather than by system size, and tests are likely to take a couple of hours per model on GPU. Faster inference can be achieved using the jax-accelerated simulations in MLIP Audit directly.

Data availability

Input structures:

  • MLIP Audit benchmark suite, InstaDeep. Starting geometries selected from the QM9 dataset (Ramakrishnan et al., Scientific Data 1, 140022, 2014).

Reference data:

  • QM-optimised starting geometries from QM9 (\(B3LYP/6-31G(2df,p)\) level of theory).

Stability

Summary

Performance in running stable molecular dynamics across a diverse set of systems: small molecules (containing H/C/N/O, sulfur, and halogens), peptides in vacuum (neurotensin, PDB: 2LNF; cyclic oxytocin, PDB: 7OFG), a protein in vacuum (PDB: 1A7M), and solvated peptides with and without counter-ions. Each system is run with a 100,000 step NVT molecular dynamics simulation at 300 K, with a frame saved every 100 steps.

Metrics

  1. Stability score

The stability score as computed by MLIP Audit. Computed as:

\[\begin{split}S = \begin{cases} 0, & \text{the simulation failed with an error} \\ \frac{1}{2} \frac{f_\mathrm{e}}{N}, & \text{the simulation exploded} \\ \frac{1}{2} + \frac{1}{2} \frac{f_\mathrm{h}}{N}, & \text{no explosion, but a hydrogen was lost} \\ 1, & \text{stable for the whole trajectory} \end{cases}\end{split}\]

where \(N\) is the total number of frames the simulation is run for, \(f_\mathrm{e}\) is the frame where the system exploded, and \(f_\mathrm{h}\) is the frame where a hydrogen drifted at least 2.5 Å away from all heavy atoms. The reported value is the mean of \(S\) over all systems. The mean is between 0 and 1, where higher is better.

A scatter plot below the table shows the stability score for each system and model. Individual models can be toggled via the legend.

Benchmark speed

Very slow: tests are likely to take many hours on GPU. Faster simulation times can be achieved using the jax accelerated simulations in MLIP Audit directly.

Data availability

Input structures:

  • MLIP Audit benchmark suite, InstaDeep. Structures derived from PDB entries 2LNF, 7OFG, and 1A7M.