Questions about training potential functions for solid-liquid interface heat transfer using DPGEN

[Xing-Yuan-Yang]

[param.json](https://github.com/user-attachments/files/25733781/param.json) Hello everyone, I want to use DPGEN to train a potential function for measuring the thermal resistance at the copper-water solid-liquid interface. However, I still have some questions about dataset construction and model training. First, I used MS to model the system, as shown in the figure. The model consists of a copper substrate with a (1,1,1) surface cut, a water box with a density of 1 g/ml, and a vacuum layer with a height of 30 Å. There is a 3 Å gap between the water box and the copper substrate. Since this system is somewhat large, I skipped the structure optimization step and directly performed 300 steps of AIMD constant-temperature simulation in VASP based on kspacing at a temperature of 400 K. Then, I trained the model in DPGEN, and the param.json file is provided below. Many parameters were set using the default values from the tutorial. I have always been confused about the settings for model_devi_f_trust_lo/hi. Previously, setting them to 0.15-0.6 resulted in 100% failure, while setting them to 0.05-0.2 allowed some configurations to enter the candidate set. Since the trained potential function is intended for use in LAMMPS simulations of boiling processes, I am not very clear about how to set the exploration parameters during the model_devi stage. Currently, I am performing constant-temperature explorations at several temperatures near the boiling point. I would greatly appreciate any guidance from the experts in the forum.

[dosubot[bot]]

Hi @Big-chicken23! I'm Dosu and I’m helping the dpgen team.

Your experience with model_devi_f_trust_lo/hi values aligns with community best practices. Here are some recommendations:

Trust Parameter Settings

Your values of 0.05-0.2 are appropriate. The general guideline is to set model_devi_f_trust_lo close to your training force RMSE, and model_devi_f_trust_hi at roughly 3-5× higher. The 0.15-0.6 range was likely too permissive, allowing configurations that were too far from your training data to be selected.

Why 100% Failure Occurs

High failure rates typically happen when:

1. MD trajectories run too long in early iterations when model quality is still low
2. The model encounters configurations it hasn't learned (like surface diffusion, clustering, or water-copper interactions not covered in initial data)

Key Recommendations:

• Shorten MD trajectories in early iterations to prevent exploring unphysical configurations before the model is well-trained (related discussion)
• Consider using model_devi_adapt_trust_lo for adaptive sampling

Temperature Exploration for Boiling

For phase transition simulations, note that models trained at low temperatures cannot reliably be used at high temperatures, but models trained at high temperatures can be applied at low temperatures. For boiling simulations, you should:

• Train on temperatures spanning from room temperature to well above your target boiling point
• Incrementally extend exploration trajectory length when moving to higher temperatures
• Consider exploring a range of densities as well, since water systems benefit from broader temperature and density coverage

For Copper-Water Interface Data

Interfaces between different materials typically need to be constructed manually. Make sure your training data covers diverse configurations including: various adsorption sites, surface diffusion pathways, multi-water configurations, and any high-deviation or "failed" structures that represent challenging but physical scenarios.

Would you like more details on any specific aspect, such as how to configure adaptive trust levels or suggestions for initial data coverage?

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[Xing-Yuan-Yang]

@dosu
Could you write an INCAR file for the structural optimization of my copper-water solid-liquid interface?

[dosubot[bot]]

Here's an INCAR for structural optimization of your Cu-water interface. Run this before your AIMD to get a reasonable starting structure:

SYSTEM = Cu-H2O interface relaxation

# Electronic relaxation
ALGO = Normal
PREC = Accurate
ENCUT = 500          # Same as your AIMD/fp
EDIFF = 1E-6         # Tighter convergence for relaxation
NELM = 200
NELMIN = 4

# Smearing
ISMEAR = 0           # Gaussian smearing
SIGMA = 0.05

# Spin
ISPIN = 1            # Non-magnetic

# Ionic relaxation
IBRION = 2           # Conjugate gradient (robust for initial relaxation)
NSW = 200            # Max ionic steps
ISIF = 2             # Relax ions only, fix cell shape/volume
EDIFFG = -0.03       # Force convergence: stop when all forces < 0.03 eV/Å

# Symmetry
ISYM = 0             # No symmetry

# Performance
LREAL = Auto         # Real-space projection for large systems
NPAR = 4             # Adjust based on your cores

# Output
LCHARG = .FALSE.
LWAVE = .TRUE.       # Save WAVECAR for AIMD restart

# Van der Waals
IVDW = 12            # D3-BJ (consistent with AIMD)

# Mixing
AMIX = 0.2
BMIX = 0.0001

Workflow:

1. Relaxation (this INCAR) → Get optimized CONTCAR
2. AIMD (previous INCAR) → Use CONTCAR as POSCAR, copy WAVECAR for faster startup

Important reminders:

• No Selective Dynamics — let all atoms move freely
• If relaxation is slow to converge, try switching to IBRION = 1 (RMM-DIIS, faster near minimum)
• After relaxation, check that water molecules are intact (O-H ≈ 0.96 Å) and Cu layers aren't distorted

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[Xing-Yuan-Yang]

@dosu
While running structure optimization in VASP, the output keeps showing a warning/error like “ZBRENT”. What does this mean and how can I fix it?

[dosubot[bot]]

The ZBRENT warning means VASP's line minimization algorithm (used with IBRION = 1, RMM-DIIS) is having trouble finding the optimal step size along the search direction. This is common with complex interface systems like your Cu-water model.

Quick fix — change IBRION:

IBRION = 2     # Conjugate gradient (more robust, no ZBRENT issues)
POTIM = 0.5    # Reduce step size for safety

The INCAR I provided earlier already uses IBRION = 2, so if you're using that, you shouldn't see ZBRENT. If you switched to IBRION = 1, go back to IBRION = 2.

If you want to keep IBRION = 1, reduce POTIM:

IBRION = 1
POTIM = 0.2    # Smaller step size (default is 0.5)

If ZBRENT persists even with IBRION = 2:

• Your initial structure may have large forces. Try a two-step approach:
  1. First: IBRION = 2, POTIM = 0.3, NSW = 100 (rough relaxation)
  2. Then: IBRION = 1, POTIM = 0.5, NSW = 200 (fine relaxation using CONTCAR from step 1)

Note: ZBRENT warnings are usually not fatal — the optimization often still converges, just less efficiently. Check if your forces are decreasing in the OSZICAR. If they are, you can let it continue.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[dosubot[bot]]

This is impressive work! Your contour maps clearly show the spatial and temporal evolution of the pseudo-boiling transition — the temperature field leading the structural transition (density drop) propagating from the wall into the bulk is a compelling demonstration of the thermally driven mechanism.

This kind of analysis is exactly what a well-trained deep potential enables — capturing the coupled thermal-structural dynamics at the molecular level that classical force fields may not reproduce accurately.

Is this using a deep potential trained with DPGEN, or a classical force field? If you're planning to extend this approach to your Cu-water boiling study, the methodology would transfer nicely — you would similarly analyze temperature, density, and energy fields near the Cu-water interface to characterize the nucleation and growth of vapor phases during boiling heat transfer.

Good progress! Let me know if you need further help with your Cu-water DPGEN training.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.