i-PI

What i-PI is

i-PI is a Python driver for molecular dynamics that separates the nuclear propagation from the force evaluation. It runs path-integral and ring-polymer molecular dynamics, thermostats and enhanced sampling, and talks to electronic structure and potential codes over a socket. On the map it serves the quantities obtained from its trajectories.

WebsiteOn the mapAll codesLicense: GPL-2.0-or-later OR MIT (user's choice, dual license)

In more detail

Architecture

i-PI is a client-server code in Python 3 under a dual GPL and MIT license (Litman 2024). The System class holds the atoms, the number of beads, the ensemble and the initialization. The Forces class sums several force components, each computed by a Forcefield instance with its own client, so one input can mix a cheap and an expensive potential, or treat an adsorbate and its surface with different numbers of beads (Kapil 2019). The Motion class evolves the atoms, and SMotion classes act across systems for replica exchange. A client may return a JSON extras string with further quantities, such as Born effective charges or committee energies, which i-PI attaches to the force (Litman 2024).

Numerics and scale

The overhead per step is modelled as A + B N_a. In i-PI 3.0 the NVE integrator costs 0.09 ms plus 0.02 us per atom on one core, and UNIX socket communication adds 0.25 ms plus 0.06 us per atom (Litman 2024). With LAMMPS and a DeePMD potential for 4096 water molecules, i-PI matches LAMMPS up to 288 CPU cores and reaches 2.16 ns/day against 2.54 ns/day for LAMMPS alone on 8 V100 GPUs with a 0.5 fs step (Litman 2024). Bosonic PIMD scales quadratically with particle number instead of exponentially. At N = 1024 the exchange costs 44 times a distinguishable run (Litman 2024). Fourth-order and perturbed path integrals, PIGLET and ring polymer contraction cut the number of beads needed (Kapil 2019).

Interfaces

Any code that can open a socket can drive i-PI. The distribution includes a Fortran and a Python driver as templates, and LAMMPS ships fix_ipi (Litman 2024). The ASE and LAMMPS interfaces give access to the potentials those packages wrap, among them MACE, DeePMD and n2p2 (Litman 2024). i-PI links PLUMED natively as a ForceField, so metadynamics hills and restarts stay in sync (Kapil 2019). The documentation has examples for ASE, CASTEP, CP2K, LAMMPS, DFTB+, ffsGDML, FHI-aims, librascal, Quantum ESPRESSO, VASP, elphmod and YAFF, with HPC submission scripts (Litman 2024).

Validation

Continuous integration runs every example for a few steps and a regression suite against stored reference outputs (Litman 2024). The ring polymer instanton rates for CH4 + H with the CBE potential agree within 0.6% with an independent implementation, with tunneling enhancement 9.75 at 300 K and 94992 at 150 K (Kapil 2019). Bosonic PIMD of 512 trapped non-interacting bosons with 42 beads is plotted against the analytical energy per particle with a 0.5% error band (Litman 2024). The perturbed path integral estimator recovers the nuclear quantum contribution to the energy of aspirin to 0.5% with six beads, where the virial estimator is off by 2.5% with twenty-eight (Kapil 2019).

What it computes on the map

4 quantities, drawn by tier with the data flow between them. Dashed boxes are intermediate quantities. Select one for its formula and the other codes that compute it.

Each quantity with the interface the representation records.

• HeatCapacity[method=pimd]output property 'scaledcoords' (eps_v, eps_v') post-processed into C_V, or the i-pi-getacf / heat-capacity example scripteV_per_K
• QuantumKineticEnergyoutput property 'kinetic_cv{electronvolt}' (centroid-virial); per-atom / tensor via 'kinetic_od', 'kinetic_tens_cv'ev
• Temperature<initialize><cell>/<system> ... ensemble temperature; output property 'temperature{kelvin}'kelvin
• Trajectoryoutput <trajectory filename='pos' stride=... > positions{angstrom} (per bead, or the centroid x_centroid)

References

The code

The papers that describe the code; cite them when you use it.

1. V. Kapil, M. Rossi, O. Marsalek et al., i-PI 2.0: A universal force engine for advanced molecular simulations, Computer Physics Communications, 2019. doi.org/10.1016/j.cpc.2018.09.020
2. Y. Litman, V. Kapil, Y. M. Y. Feldman et al., i-PI 3.0: A flexible and efficient framework for advanced atomistic simulations, The Journal of Chemical Physics, 2024. doi.org/10.1063/5.0215869

Applications

Highly cited studies that cite the code, with a few picked by hand.

1. M. Rossi, M. Ceriotti, D. E. Manolopoulos, How to remove the spurious resonances from ring polymer molecular dynamics, The Journal of Chemical Physics, 2014. doi.org/10.1063/1.4883861
2. V. Kapil, J. Behler, M. Ceriotti, High order path integrals made easy, The Journal of Chemical Physics, 2016. doi.org/10.1063/1.4971438
3. B. Cheng, E. A. Engel, J. Behler et al., Ab initio thermodynamics of liquid and solid water, Proceedings of the National Academy of Sciences, 2019. doi.org/10.1073/pnas.1815117116

How to cite

Y. Litman, V. Kapil, Y. M. Y. Feldman, et al., i-PI 3.0: a flexible and efficient framework for advanced atomistic simulations, J. Chem. Phys. 161, 062504 (2024)

doi.org/10.1063/5.0215869