Regarding the allocation of steps in "stop_batch", is it evenly distributed among each "system" in "init_data_sys"?

[li-shining]

In the dpgen run task, regarding the distribution of steps in "stop_batch", are they evenly distributed among each "system" in "init_data_sys"? If the size of each "system" is different, will the weight of the steps be automatically adjusted?

[dosubot[bot]]

stop_batch is a global parameter — it defines the total number of training steps for the entire training process, not per system. It is not evenly distributed among each system in init_data_sys [1].

How systems are sampled: By default, DP-GEN uses prob_sys_size weighting [2], which means the probability of sampling from a given system at each training step is proportional to the number of frames in that system [3]. So yes, the weight is automatically adjusted based on system size.

For example, if you have 3 systems with 100, 300, and 600 frames respectively, and stop_batch = 10000:

• System A (100 frames): ~10% of steps → ~1,000 steps
• System B (300 frames): ~30% of steps → ~3,000 steps
• System C (600 frames): ~60% of steps → ~6,000 steps

If you want equal sampling regardless of system size, you can change the probability style to "prob_uniform", which gives each system a 1/N probability [3]. You can also use sys_probs to set explicit per-system probabilities.

Additionally, per-system batch sizes are controlled by init_batch_size (one entry per system in init_data_sys). Setting it to "auto" will automatically choose a batch size based on the rule batch_size × number_of_atoms ≥ 32 [4].

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[dosubot[bot]]

Yes, this paper is very helpful for guiding your dataset construction! Based on the article's findings, the key phenomenon on Bi₂Mo₃O₁₂(010) is Mo volatilization at 500–600°C leading to Bi-rich surface deposits and formation of new phases (Bi₂MoO₆, Bi₂Mo₂O₉). This is somewhat different from simple Bi diffusion to the surface — it's driven by Mo leaving the surface rather than Bi migrating upward.

How this informs your initial dataset:

The paper tells you what structural states the model needs to learn. Here's a concrete dataset construction plan:

1. Stoichiometric configurations (endpoints):

• Bulk Bi₂Mo₃O₁₂ (your existing data)
• Clean Bi₂Mo₃O₁₂(010) surface slab (≥6 layers)

2. Mo-depleted surface configurations (key intermediate states):

• (010) surface slab with 1 Mo removed from the top layer (Mo vacancy)
• (010) surface slab with 2 Mo removed from top layers
• Progressive Mo depletion: remove Mo from surface → subsurface → deeper layers
• Run short AIMD (1–2 ps) at 773K and 873K (500°C, 600°C) on each of these to sample local relaxations

3. Bi-enriched surface configurations:

• Slab with Bi displaced to occupy former Mo sites near the surface
• Slab with excess Bi adatom on the (010) surface
• Configurations mimicking Bi₂MoO₆-like local structure at the surface (since the paper shows this phase precipitates)

4. Product phase configurations:

• Bulk Bi₂MoO₆ (product phase observed at 500°C)
• Bulk Bi₂Mo₂O₉ (product phase observed at 600°C)
• Interfaces between Bi₂Mo₃O₁₂ and these product phases (if computationally feasible)

5. Mo in vacuum / Mo desorption:

• If you want to capture Mo volatilization, include configurations with Mo species (MoO₃, MoO₂) slightly above the surface or in the gas phase above the slab. This teaches the model the energetics of Mo leaving.

Practical init_data_sys structure:

"init_data_sys": [
    "bulk/Bi2Mo3O12/training_data",
    "bulk/Bi2MoO6/training_data",
    "bulk/Bi2Mo2O9/training_data",
    "surface/010_stoichiometric/training_data",
    "surface/010_Mo_vacancy_1/training_data",
    "surface/010_Mo_vacancy_2/training_data",
    "surface/010_Bi_enriched/training_data"
]

For sys_configs exploration:

"sys_configs": [
    ["surface/010_all_poscars/POSCAR_*"],
    ["surface/Mo_depleted_all_poscars/POSCAR_*"],
    ["surface/Bi_enriched_all_poscars/POSCAR_*"]
]

Important considerations:

• The paper's mechanism is thermally activated Mo loss, so explore at temperatures matching the paper: 773K (500°C) and 873K (600°C), plus higher temperatures to accelerate sampling
• Include O₂ or H₂O in your system if you want to study the atmosphere-dependent behavior described in the paper
• The different product phases (Bi₂MoO₆ vs Bi₂Mo₂O₉) form at different temperatures — your model needs to know both endpoints

The paper essentially gives you a roadmap of what structural states exist along the transformation pathway. Use that to manually construct the intermediate configurations that dpgen's active learning will then refine and fill in.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[li-shining]

@dosu, For the types of structures you just mentioned, do we need to perform structural optimization first, or can we run AIMD directly once the structure is finalized?

[dosubot[bot]]

Yes, you should structurally optimize (relax) first, then run AIMD. Here's the reasoning for each type:

For bulk Bi₂Mo₃O₁₂, Bi₂MoO₆, Bi₂Mo₂O₉:

• Relax the cell (both ionic positions and cell shape/volume) to get equilibrium lattice parameters
• Then run AIMD at multiple temperatures on the relaxed structure
• This is standard practice and part of dpgen's init_bulk workflow [1]

For surface slabs and defective structures:

• First build the slab from the relaxed bulk lattice parameters [2]
• Then do a brief ionic relaxation (ISIF=2, fixed cell shape) to let the surface atoms adjust — especially important for defect structures (Mo vacancies, Bi-enriched surfaces) where the initial manual construction may have atoms too close or in unreasonable positions
• Then run AIMD

Why relaxation matters:

• If you run AIMD on an unrelaxed structure with very high initial forces (e.g., atoms too close after manually removing Mo), the first few AIMD steps will have extremely large forces that may not be representative of physical behavior
• The relaxation ensures AIMD starts from a reasonable local minimum, so the sampled configurations around it are physically meaningful

However, you don't need to relax to very tight convergence. A quick relaxation (EDIFFG = -0.05 eV/Å) is sufficient — the purpose is just to remove the worst structural artifacts before AIMD sampling.

Workflow summary:

1. Relax bulk → get lattice constants
2. Build slabs/defects from relaxed bulk
3. Briefly relax each slab/defect structure (ionic only, ISIF=2)
4. Run AIMD at 2–3 temperatures (e.g., 773K, 873K, 1500K) on each relaxed structure
5. Extract frames (skip first 50–100 equilibration steps), convert to training data with dpdata

Remember: never fix atoms during either the relaxation or the AIMD [3]. Use thick slabs instead.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[li-shining]

@dosu, based on this article(https://doi.org/10.1016/S0021-9517(02)00024-6) and the techniques for constructing the initial dpgen dataset, give me the most detailed plan possible for building the initial dataset.

[dosubot[bot]]

Based on the paper's findings on Bi₂Mo₃O₁₂(010) surface evolution (Yanina et al., J. Catal. 2003) and dpgen best practices, here is a detailed plan:

---

The paper identifies these key physical processes to capture:

• Mo volatilization from the (010) surface at ≥500°C
• Bi-rich deposit accumulation on terraces
• Formation of Bi₂MoO₆ precipitates (at 500°C in humid air)
• Formation of Bi₂Mo₂O₉ precipitates (at 600°C in humid air)
• Step recession and pit/island formation (b/2 = 5.75 Å step heights)

---

Phase 1: Bulk Phase Data

Structure 1: Bulk Bi₂Mo₃O₁₂ (α-bismuth molybdate)

• Relax unit cell fully (ISIF=3)
• Run AIMD at 3 temperatures: 673K (400°C), 773K (500°C), 873K (600°C)
• 1–2 ps each, extract ~20 frames per temperature after equilibration

Structure 2: Bulk Bi₂MoO₆ (γ-bismuth molybdate)

• Relax unit cell
• AIMD at 773K, 873K (~20 frames each)
• This is the product phase at 500°C

Structure 3: Bulk Bi₂Mo₂O₉ (β-bismuth molybdate)

• Relax unit cell
• AIMD at 773K, 873K (~20 frames each)
• This is the product phase at 600°C

Structure 4: Bulk MoO₃

• Relax unit cell
• AIMD at 773K, 873K (~10 frames each)
• Represents the volatile Mo species

---

Phase 2: Stoichiometric (010) Surface

Structure 5: Clean Bi₂Mo₃O₁₂(010) slab

• Build from relaxed bulk, cleave along (010) — the paper notes this is the natural cleavage plane
• Use ≥6 layers (the paper identifies b/2 = 5.75 Å step heights, so layers are ~5.75 Å)
• Vacuum ≥ 20 Å
• Relax ionic positions (ISIF=2), do NOT fix any atoms
• AIMD at 673K, 773K, 873K, and 1100K (higher T to accelerate sampling)
• ~20 frames per temperature

---

Phase 3: Mo-Depleted Surface Configurations (Critical!)

This is the core of the paper's findings — Mo volatilizes, leaving Bi-enriched surfaces.

Structure 6: 1 Mo removed from top surface layer

• Remove one Mo atom from the topmost layer of the (010) slab
• Relax briefly (ISIF=2, EDIFFG=-0.05)
• AIMD at 773K, 873K, 1100K (~20 frames each)

Structure 7: 2 Mo atoms removed from top surface layer

• Remove two Mo atoms
• Same procedure as above

Structure 8: Mo removed from subsurface layer

• Remove Mo from the second layer instead
• AIMD at 773K, 873K (~15 frames each)

Structure 9: Multiple Mo vacancies (heavily depleted surface)

• Remove 3–4 Mo atoms from top 2 layers
• This mimics the later stages of Mo volatilization
• AIMD at 773K, 873K (~15 frames each)

---

Phase 4: Bi-Enriched/Reorganized Surface

The paper shows Bi-rich deposits form on the surface.

Structure 10: Bi adatom on (010) surface

• Place an extra Bi atom on the clean (010) surface at different sites
• Relax + AIMD at 773K, 873K (~15 frames each)

Structure 11: Bi displaced to former Mo site

• In a Mo-vacant slab, move a nearby Bi to occupy/approach the Mo vacancy
• Relax + AIMD at 773K, 873K

Structure 12: Bi₂MoO₆-like surface reconstruction

• Modify the top 1–2 layers of the (010) slab to approximate Bi₂MoO₆ stoichiometry locally (Bi/Mo ratio ~2:1 instead of 2:3)
• Relax + AIMD at 773K, 873K

---

Phase 5: Mo Desorption Pathway (if modeling volatilization)

Structure 13: MoO₃ molecule above surface

• Place a MoO₃ unit 3–4 Å above the (010) surface (in the vacuum region)
• This represents the early stage of Mo desorption
• AIMD at 773K, 873K (~10 frames)

Structure 14: MoO₃ molecule in gas phase

• Isolated MoO₃ in a large box (15×15×15 Å)
• AIMD at 773K (~10 frames)

---

Phase 6: Interface Configurations (Advanced, Optional)

Structure 15: Bi₂MoO₆/Bi₂Mo₃O₁₂ interface

• Stack a thin Bi₂MoO₆ layer on the (010) surface
• Relax + AIMD at 773K (~10 frames)
• Captures the precipitate-substrate interaction

---

AIMD Parameters (consistent across ALL structures)

# VASP INCAR for AIMD
IBRION = 0          # MD
MDALGO = 2          # Nosé-Hoover thermostat
SMASS  = 1.0        # thermostat mass
POTIM  = 1.0        # 1 fs timestep
NSW    = 2000       # 2 ps total
TEBEG  = [target T]
ISIF   = 2          # fixed cell, free ions
EDIFF  = 1E-5
ENCUT  = 600        # high cutoff for oxygen
KSPACING = 0.12     # consistent for all
GGA = PE            # PBE
IVDW = 12           # DFT-D3
LREAL = Auto
PREC = Accurate
SIGMA = 0.05
ISPIN = 1           # unless magnetism expected

---

Data Extraction with dpdata

import dpdata

# For each AIMD run:
sys = dpdata.LabeledSystem('path/to/OUTCAR', fmt='vasp/outcar')
# Skip first 500 fs equilibration (500 steps at 1 fs)
sys_equil = sys.sub_system(range(500, len(sys)))
# Subsample every 10 steps to reduce redundancy
sys_final = sys_equil.sub_system(range(0, len(sys_equil), 10))
sys_final.to('deepmd/npy', 'output/training_data')

---

Final init_data_sys Structure

"init_data_sys": [
    "bulk/Bi2Mo3O12/training_data",
    "bulk/Bi2MoO6/training_data",
    "bulk/Bi2Mo2O9/training_data",
    "bulk/MoO3/training_data",
    "surface/010_clean/training_data",
    "surface/010_Mo_vacancy_light/training_data",
    "surface/010_Mo_vacancy_heavy/training_data",
    "surface/010_Bi_enriched/training_data",
    "surface/Mo_desorption/training_data"
]

Merge entries with the same atom count as discussed earlier.

---

Expected Frame Count Estimate

Category	Structures	Frames/structure	Total
Bulk phases (4)	4	~50	~200
Clean surface	1	~80	~80
Mo-depleted surfaces	4	~50	~200
Bi-enriched surfaces	3	~40	~120
Mo desorption	2	~20	~40
Interface (optional)	1	~10	~10
Total			~650

This provides a solid foundation. dpgen's active learning will then expand this significantly by discovering configurations along the actual transformation pathways during exploration at 773–1100K.

---

Key reminders:

• Never fix atoms in any calculation
• Identical DFT settings across all AIMD and future dpgen fp calculations
• Use PBE+D3 consistently (as you decided earlier)
• Thicker slabs (≥6 layers) rather than fixing bottom atoms

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.