[Q&A] Hybrid DeepMD–Classical Force Field NVT Simulation in CP2K

[drpriyadey]

Question

Is it possible to perform an NVT simulation in CP2K for a hybrid metallic–organic system, where a DeepMD potential is used for the inorganic component and a classical force field is applied to the organic component?

DeePMD-kit Version

DeePMD-kit v3.1.3

Backend and its version

TensorFlow 2.19.1

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

Python 3.12.13
CUDA toolkit (nvcc) 12.0.140
CUDA 12.8
CP2K 2024.3

Details

No response

Reproducible Example, Input Files, and Commands

No response

Further Information, Files, and Links

No response

[Tobi-Adesoye]

Hi @drpriyadey,

Yes, this is completely possible in CP2K, but it requires configuring the simulation engine under a Mixed Force-Field (QM/MM style) framework. Instead of treating DeePMD as a global potential, you partition the system so that CP2K evaluates the organic component using its internal classical MM mechanics, while routing the inorganic/metallic atomic coordinates to DeePMD-kit via an external calculator hook.

Here is how to map this out in your CP2K input script (.inp):

1. Define the Mixed Force Evaluation Structure
   You must instruct CP2K to initialize a MIXED force environment. This tells the core coordinator to combine different calculation blocks for the total energy and forces:

Fortran
&FORCE_EVAL
METHOD Mixed
&MIXED
MIXING_TYPE Linear
&LINEAR
# Combines forces additively across the defined subsystems
V_NAME Force_Metal, Force_Organic
WEIGHT 1.0, 1.0
&END LINEAR
&END MIXED
&END FORCE_EVAL

2. Bind the Inorganic Metallic Subsystem to DeePMD-kit
   Create your first distinct FORCE_EVAL block tracking your inorganic component. You will use EXTERNAL_POTENTIAL to pass your exported TensorFlow graph (.pb file) to the DeepMD C++ API interface embedded in your CP2K build:

Fortran
&FORCE_EVAL
METHOD FIST
SUBsys_NAME Force_Metal
&MM
&FORCEFIELD
&EXTERNAL_POTENTIAL
FUNCTION_TYPE DEEPMD
MODEL_FILE_NAME ./graph.pb
# Explicitly list the atom types assigned to the Deep Potential
ATOM_MAP_NAME Mo, S
&END EXTERNAL_POTENTIAL
&END FORCEFIELD
&END MM
&SUBSYS
# Include coordinate definitions tracking only the inorganic subset
&CELL
ABC 25.0 25.0 25.0
&END CELL
&TOPOLOGY
COORD_FILE_NAME metal_slice.xyz
COORDINATE XYZ
&END TOPOLOGY
&END SUBSYS
&END FORCE_EVAL

3. Bind the Organic Subsystem to your Classical Force Field
   Create your second FORCE_EVAL block using standard FIST mechanics to evaluate your classical force parameters (e.g., AMBER, CHARMM, or OPLS-AA) for the organic component:

Fortran
&FORCE_EVAL
METHOD FIST
SUBsys_NAME Force_Organic
&MM
&FORCEFIELD
PARM_FILE_NAME organic_forcefield.parm
PARMTYPE AMBER
&END FORCEFIELD
&POISSON
&EWALD
EWALD_TYPE SPME
&END EWALD
&POISSON
&END MM
&SUBSYS
&CELL
ABC 25.0 25.0 25.0
&END CELL
&TOPOLOGY
COORD_FILE_NAME organic_slice.xyz
COORDINATE XYZ
&END TOPOLOGY
&END SUBSYS
&END FORCE_EVAL

4. Configure the NVT Ensemble Integration Loop
   Finally, your main MOTION block handles the execution matrix. For a hybrid system, a stochastic velocity rescaling thermostat (CSVR) is highly recommended to prevent temperature spikes or artificial energy transfer across the hybrid boundary:

Fortran
&MOTION
&MD
ENSEMBLE NVT
STEPS 50000
TIMESTEP [fs] 0.5
TEMPERATURE [K] 300.0
&THERMOSTAT
TYPE CSVR
&CSVR
TIMECON [fs] 10.0
&END CSVR
&END THERMOSTAT
&MD
&END MOTION

Ensure that your CP2K executable was compiled with the flags matching your libdeepmd_c.so library linking, or you will throw an unrecognized keyword error when initializing the DEEPMD block type.

[drpriyadey]

Thank you so much @Tobi-Adesoye san. I will try the procedure.

[Tobi-Adesoye]

You are very welcome, @drpriyadey! Don't hesitate to reach back out or reopen the thread if you encounter any convergence anomalies or formatting quirks in your INPUT blocks once the CP2K run kicks off. Best of luck with your hybrid NVT simulation!

…

On Wed, Jun 10, 2026 at 2:39 AM Priya Dey ***@***.***> wrote: Thank you so much @Tobi-Adesoye <https://github.com/Tobi-Adesoye> san. I will try the procedure. — Reply to this email directly, view it on GitHub <#5505?email_source=notifications&email_token=AQSPCWHW6OUJKUTTCOFJ4KD47C333A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZSGQ2DIMJTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17244413>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQSPCWB2PUQH6KSS7ITFU7L47C333AVCNFSM6AAAAACZ6SGCECVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTOMRUGQ2DCMY> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AQSPCWEXY7FDSROYTAUKRXD47C333A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZSGQ2DIMJTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android <https://github.com/notifications/mobile/android/AQSPCWAWQ7T4P727YROPNRT47C333A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZSGQ2DIMJTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.*** com>

[drpriyadey]

Dear @Tobi-Adesoye ,
When i am running the above script with the suitable structure and parameter files, i am getting the "unknown keyword" error from cp2k for some of the lines. Is there anything I am missing? Thank you for your help.

[drpriyadey]

Dear @Tobi-Adesoye
After i am trying to correct the input file, I have the following error:
DBCSR| CPU Multiplication driver XSMM (U)
DBCSR| Multrec recursion limit 512 (U)
DBCSR| Multiplication stack size 1000 (D)
DBCSR| Maximum elements for images UNLIMITED (U)
DBCSR| Multiplicative factor virtual images 1 (U)
DBCSR| Use multiplication densification T (D)
DBCSR| Multiplication size stacks 3 (U)
DBCSR| Use memory pool for CPU allocation F (U)
DBCSR| Number of 3D layers SINGLE (U)
DBCSR| Use MPI memory allocation F (U)
DBCSR| Use RMA algorithm F (U)
DBCSR| Use Communication thread T (U)
DBCSR| Communication thread load 87 (D)
DBCSR| MPI: My process id 0
DBCSR| MPI: Number of processes 1
DBCSR| OMP: Current number of threads 1
DBCSR| OMP: Max number of threads 1
DBCSR| Split modifier for TAS multiplication algorithm 1.0E+00 (U)

**** **** ****** ** PROGRAM STARTED AT 2026-06-22 18:05:18.760
***** ** *** *** ** PROGRAM STARTED ON priya-dey
** **** ****** PROGRAM STARTED BY priya-dey
***** ** ** ** ** PROGRAM PROCESS ID 88936
**** ** ******* ** PROGRAM STARTED IN /home/priya-dey/Project-ML/bt3

CP2K| version string: CP2K version 2024.3
CP2K| source code revision number: git:6712648
CP2K| cp2kflags: omp libint fftw3 libxc libgrpp pexsi elpa parallel scalapack m
CP2K| pi_f08 cosma quip deepmd xsmm plumed2 spglib sirius libvori li
CP2K| ibbqb libtorch libvdwxc hdf5
CP2K| is freely available from https://www.cp2k.org/
CP2K| Program compiled at Tue Apr 7 01:57:03 AM JST 2026
CP2K| Program compiled on priya-dey
CP2K| Program compiled for local
CP2K| Data directory path /home/priya-dey/cp2k/data
CP2K| Input file name tobi.inp

---

• ___ *
• / \ *
• [ABORT] *
• ___/ CPASSERT failed *
• | *
• O/| *
• /| | *
• / \ force_env_types.F:730 *

---

===== Routine Calling Stack =====

        1 CP2K

[Tobi-Adesoye]

Hi @drpriyadey, Thanks for testing this and for sharing the error output. Looking at the messages you posted, I suspect the problem is not with your structure files but with the input syntax itself. The unknown keyword errors, followed by the CPASSERT failed in force_env_types.F, indicate that CP2K is encountering sections or keywords that it does not recognize for your build or version. In hindsight, the example I shared was intended to illustrate the high-level partitioning strategy rather than serve as a drop-in CP2K input file. Some of the block names and keywords may therefore not correspond exactly to valid CP2K syntax. A few things would help narrow this down: 1. Could you post the complete tobi.inp file (or at least the relevant FORCE_EVAL sections)? 2. Was your CP2K executable compiled with DeePMD support? Your banner indicates deepmd is present, which is encouraging, but the exact interface still matters. 3. What is your intended coupling strategy? For example: - DeePMD for one subset of atoms and a classical force field for another subset, or - DeePMD as the sole potential for the entire system. If the goal is a true hybrid DeePMD + classical force-field simulation in CP2K, I would also recommend checking whether the required coupling mechanism is explicitly supported by your CP2K version and build configuration, as the available interfaces can differ between releases. If you share the input file, I'd be happy to take a closer look and help identify the specific source of the assertion failure.

…

On Mon, Jun 22, 2026 at 10:08 AM Priya Dey ***@***.***> wrote: Dear @Tobi-Adesoye <https://github.com/Tobi-Adesoye> After i am trying to correct the input file, I have the following error: DBCSR| CPU Multiplication driver XSMM (U) DBCSR| Multrec recursion limit 512 (U) DBCSR| Multiplication stack size 1000 (D) DBCSR| Maximum elements for images UNLIMITED (U) DBCSR| Multiplicative factor virtual images 1 (U) DBCSR| Use multiplication densification T (D) DBCSR| Multiplication size stacks 3 (U) DBCSR| Use memory pool for CPU allocation F (U) DBCSR| Number of 3D layers SINGLE (U) DBCSR| Use MPI memory allocation F (U) DBCSR| Use RMA algorithm F (U) DBCSR| Use Communication thread T (U) DBCSR| Communication thread load 87 (D) DBCSR| MPI: My process id 0 DBCSR| MPI: Number of processes 1 DBCSR| OMP: Current number of threads 1 DBCSR| OMP: Max number of threads 1 DBCSR| Split modifier for TAS multiplication algorithm 1.0E+00 (U) **** **** ****** ** PROGRAM STARTED AT 2026-06-22 18:05:18.760 ***** ** *** *** ** PROGRAM STARTED ON priya-dey ** **** ****** PROGRAM STARTED BY priya-dey ***** ** ** ** ** PROGRAM PROCESS ID 88936 **** ** ******* ** PROGRAM STARTED IN /home/priya-dey/Project-ML/bt3 CP2K| version string: CP2K version 2024.3 CP2K| source code revision number: git:6712648 CP2K| cp2kflags: omp libint fftw3 libxc libgrpp pexsi elpa parallel scalapack m CP2K| pi_f08 cosma quip deepmd xsmm plumed2 spglib sirius libvori li CP2K| ibbqb libtorch libvdwxc hdf5 CP2K| is freely available from https://www.cp2k.org/ CP2K| Program compiled at Tue Apr 7 01:57:03 AM JST 2026 CP2K| Program compiled on priya-dey CP2K| Program compiled for local CP2K| Data directory path /home/priya-dey/cp2k/data CP2K| Input file name tobi.inp ------------------------------ - ___ * - / \ * - [ABORT] * - ___/ CPASSERT failed * - | * - O/| * - /| | * - / \ force_env_types.F:730 * ------------------------------ ===== Routine Calling Stack ===== 1 CP2K — Reply to this email directly, view it on GitHub <#5505?email_source=notifications&email_token=AQSPCWEVU6HAW6SN2PMY3WL5BDZRNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTHEYDCMBWUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17390106>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQSPCWHS4FFPW43NB23PLEL5BDZRNAVCNFSNUABIKJSXA33TNF2G64TZHMYTCNBQGA3DCOJTHNCGS43DOVZXG2LPNY5TCMBSGE3TAMZRUF3AE> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AQSPCWHEPDACOZ7R4HW6D7L5BDZRNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTHEYDCMBWUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android <https://github.com/notifications/mobile/android/AQSPCWBAQNK7J6RA24WW73D5BDZRNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTHEYDCMBWUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.*** com>

[drpriyadey]

Dear @Tobi-Adesoye
The input is following:
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
!instruct CP2K to initialize a MIXED force environment
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

&FORCE_EVAL
METHOD Mixed
&MIXED
MIXING_TYPE LINEAR_COMBINATION
&LINEAR
!# Combines forces additively across the defined subsystems
LAMBDA 0.5
&END LINEAR
&END MIXED
&END FORCE_EVAL

!!Bind the Inorganic Metallic Subsystem to DeePMD-kit

&FORCE_EVAL
METHOD FIST
&MM

&FORCEFIELD
&NONBONDED
&DEEPMD
ATOMS C CU N S
ATOMS_DEEPMD_TYPE 0 1 2 3
POT_FILE_NAME graph.pb
&END DEEPMD
&END NONBONDED
&END FORCEFIELD
&END MM

&SUBSYS
&CELL
A 18.8886360910 0 0
B 0 19.7198589339 0
C 0 0 55.0091418766
PERIODIC XY
&END CELL

&TOPOLOGY
  COORD_FILE_NAME structure_new.xyz
  COORD_FILE_FORMAT XYZ
&END TOPOLOGY

&KIND C
  ELEMENT C
&END KIND

&KIND Cu
  ELEMENT Cu
&END KIND

&KIND N
  ELEMENT N
&END KIND

&KIND S
  ELEMENT S
&END KIND

&END SUBSYS

&END FORCE_EVAL
#Bind the Organic Subsystem to your Classical Force Field

&FORCE_EVAL
METHOD FIST
#SUBsys_NAME Force_Organic
&MM
&FORCEFIELD
PARM_FILE_NAME y6.prmtop
PARMTYPE AMBER
&END FORCEFIELD
&POISSON
&EWALD
EWALD_TYPE SPME
&END EWALD
&END POISSON
&END MM
&SUBSYS
&CELL
ABC 25.0 25.0 25.0
&END CELL
&TOPOLOGY
COORD_FILE_NAME y6_core.xyz
COORD_FILE_FORMAT XYZ
&END TOPOLOGY
&END SUBSYS
&END FORCE_EVAL
#Configure the NVT Ensemble Integration Loop

&MOTION
&MD
ENSEMBLE NVT
STEPS 50000
TIMESTEP [fs] 0.5
TEMPERATURE [K] 300.0
&THERMOSTAT
TYPE CSVR
&CSVR
TIMECON 1000.0
&END CSVR
&END THERMOSTAT
&END MD
&END MOTION

[drpriyadey]

Dear @Tobi-Adesoye ,
Can you help me understand, what do you mean by "the required coupling
mechanism is explicitly supported by your CP2K version and build
configuration"?