Machine Learning Potential for better framework-adsorbate energy: CO2 in Mg-MOF74

Introduction

Using Machine learning potential (MLP), we can better describe the energetics of CO2 with Mg-MOF74. * Below is an example of using the Allegro MLP model in gRASPA for framework-adsorbate energy

simulation.input

DNN Basic Info

The block below indicates the basic information of the MLP

UseDNNforHostGuest yes
DNNModelName co2-mof74-deployed.pth
DNNMethod Allegro
DNNEnergyUnit eV
MaxDNNDrift 100000.0

• DNNMethod can be chosen from Allegro or LCLin
• MaxDNNDrift
  • this keyword indicates the maximum energy difference acceptable between the classical energy and the MLP energy
  • How gRASPA handles this: the classical energies are calculated first, then the DNN energy is calculated as a "correction"
  • Also consider tuning OverlapCriteria together with this keyword
    • OverlapCriteria controls whether a trial should be rejected because of high vdW energies between a pair of atoms

Component Information

Component 0 MoleculeName             CO2
            IdealGasRosenbluthWeight 1.0
            FugacityCoefficient      1.0
            TranslationProbability   1.0
            RotationProbability      1.0
            ReinsertionProbability   1.0
            SwapProbability          1.0
            DNNPseudoAtoms C_co2 O_co2
            CreateNumberOfMolecules  0

• Special Notes!
  • DNNPseudoAtoms is required.
  • This indicates the pseudo-atoms in the classical simulation to be considered in the MLP.
  • Consider TIP4P water as an example. The fictional charge site should not be considered.
  • see the example here for a fictional atom site in a fake CO2 model