Machine Learning Force Fields

forcefield

Overview

Machine Learning Force Fields (MLFFs) offer a novel approach for predicting the energies of arbitrary systems. While they have a slower runtime performance than classical force fields, they hold the potential to provide near-quantum accuracy at a fraction of the cost of DFT. MLFFs generally provide smooth potential energy functions that directly mimic DFT that only depend on the cartesian coordinates of the system, with some approaches also using the net charge and spin as input, allowing them to be applied to ionic and open shell systems.

Current Limitations

Limitations for Molecular Dynamics Calculations:

  • Requires fine-tuned, and generally much smaller timesteps
  • No concepts of bonds/bond potentials means unintended "reactions" can occur, particularly in systems that begin with steric clashes
  • Currently no energy groups, so some analysis/property calculation doesn't work

Limitations for Quantum Mechanics Calculations:

  • Only first derivates are currently available analytically (geometry optimizations), second derivatives are computed by finite difference
  • Excited states are not available
  • Wavefunction properties are not available
    • No coupling integrals, NMR, higher spin states (T1), open shell*, ESP charges

Available Models

MPNICE

Model Name Description Element Support Levels of Theory Application Area Limitations
Organic_MPNICE Trained to organic molecules H C N O S F P Cl Br B (Supported but untested: Li Na Mg Si K Ca Sn I) wB97X-D3BJ/def2-TZVPD Organic molecules and liquids Limited element coverage
Organic_MPNICE_TB Delta learned model utilizing GFN2-xTB H C N O S F P Cl Br B (Supported but untested: Li Na Mg Si K Ca I) wB97X-D3BJ/def2-TZVPD Finite system applications of organic molecules Limited to finite systems
Organic_Crystals_MPNICE Multi-task model trained to finite organic systems and organic crystals H C N O S F P Cl Br PBE-D3 (PW basis), wB97X-D3BJ/def2-TZVPD Organic crystal structure optimization and ranking  
Inorganic_MPNICE Single task model trained to AIMD data from OMAT24 and fine tuned to MPTrj with GFN1-xTB charges Up to Z=94, excluding Po, At, Rn, Fr, Ra PBE-D3/GFN1-xTB charges Bulk Inorganic Materials Untested on surfaces, defects, organometallics. No magnetic properties.
Hybrid_MPNICE Model with a single output head trained to MPtrj energies/forces/stress/GFN1-xTB charges), and the forces and dipoles for finite organic systems Up to Z=94, excluding Po, At, Rn, Fr, Ra wB97X-D3BJ/def2-TZVPD, PBE-D3/GFN1-xTB charges Bulk inorganic materials, organic molecules Interfaces between materials and organic molecules/liquids must be benchmarked versus a reference. Liquid densities for organics can have large error in density in NPT dynamics. Note that there is NO organometallic data in the training set
Hybrid_MPNICE_O Multi task model trained to MPtrj and finite organic systems. This is the wB97X output head, which is slightly more accurate for organic tasks. Does not exactly reproduce wB97X total energies (uses PBE atomic energies in order to support full elemental coverage) Up to Z=94, excluding Po, At, Rn, Fr, Ra wB97X-D3BJ/def2-TZVPD, PBE-D3 Bulk inorganic materials, organic molecules and liquids Interfaces between materials and organic molecules/liquids must be benchmarked versus a reference. Note that there is NO organometallic data in the training set
Hybrid_MPNICE_I Multi task model trained to MPtrj and finite organic systems. This is the PBE-D3 output head, which is slightly more accurate for inorganic tasks such as mechanical properties of bulk inorganic crystals. Up to Z=94, excluding Po, At, Rn, Fr, Ra wB97X-D3BJ/def2-TZVPD, PBE-D3 Bulk inorganic materials, organic molecules and liquids Interfaces between materials and organic molecules/liquids must be benchmarked versus a reference. Note that there is NO organometallic data in the training set

UMA

The UMA models are developed by Meta and display high accuracy, providing near full periodic table coverage for finite systems and accurate reaction energies. They do not have long range interactions and are ~10x more expensive than MPNICE, but are extremely useful in certain use cases. See Meta’s UMA manuscript for more details.

The “small” model (version 1.1) is available in the Schrödinger Suite, which can be run with tasks omol, omc, oc20, odac and omat. You can run UMA with the MLFF_inference.py script, Jaguar, and Desmond. To specify the model name in MLFF_inference.py or Desmond, you can specify your model type as UMA_sm_<task>. Similarly, in Jaguar the model is specified with mlff=uma_sm_<task>.

Using MLFFs in place of Quantum Mechanics Calculations

An MLFF potential can be used for single point energy evaluations, geometry optimizations (constrained, TS optimizations), frequency calculations (numerical only currently), and scans. Jaguar always runs the models in double precision on the CPU.

MLFF potentials can be selected from the Theory text box in Jaguar panels. You can select MLFF after clicking on the filter button to list all available models.

To run Jaguar with an MLFF potential from the command line, the following keywords can be added to Jaguar input file:

&gen
mlff=organic_mpnice_tb
mlff_solv=1 ! Turns on implicit solvent, only available *_tb models


mlff_solvent_model = ALPB ! solvent model to use, only applies if `mlff_solv=1`
&

See External Methods Keywords in the Jaguar Input File (Machine learning methods and xTB) for all keyword values.

MLFFs in the Material Science Suite

MLFFs are available in various panels in the MS Suite:

MD QM