Multi-Component System Workflows in dpgen/dpgen2#

Multi-component systems—binary alloys, ternary alloys, mixed oxides—require workflow design that accounts for both structural and chemical degrees of freedom. Two code bases support this work: dpgen (classic, param.json-driven) and dpgen2 (second generation, featuring automated alloy structure generation). The key references are the Al-Mg dpgen2 example, the AlloyConfGenerator class, the Mg-Y tutorial for metallic alloys, and GitHub discussion #1876 for the Bi-Mo-O multi-component oxide case study.

dpgen2: Alloy Configuration Generation#

The AlloyConfGenerator in dpgen2 automates starting structure creation for alloy exploration 1. Key parameters:

When concentration is a nested list, dpgen2 randomly selects one concentration row per generated configuration 2. The Al-Mg example demonstrates multi-point sampling:

"concentration": [[1.0, 0.0], [0.5, 0.5], [0.0, 1.0]]

This generates configurations spanning pure Al, equimolar Al-Mg, and pure Mg without manually constructing per-composition datasets. Exploration stages map configurations to temperature/ensemble settings via conf_idx 3. The model seeds from separate Al and Mg AIMD data 4 to avoid starting blind on pure-element chemistry.

dpgen (Classic): Dataset Merging and sys_configs Patterns#

Seeding init_data_sys with Single-Element Data#

The Mg-Y tutorial demonstrates the canonical bootstrap pattern: complete single-element DP-GEN runs for Mg and Y separately, then merge those datasets into init_data_sys alongside multi-component Mg-Y data. This seeds the binary model with element-specific knowledge before exploring alloy chemistry.

Merging by Structure Type, Not Temperature#

For multi-component oxide systems, merge POSCARs from all temperatures into a single directory per environment type, then vary temperature via temps in model_devi_jobs 5. This avoids combinatorial explosion:

"sys_configs": [
    ["o2_adsorption/all_poscars/POSCAR_*"],
    ["o_adsorption/all_poscars/POSCAR_*"]
]

Group by chemistry (O₂ adsorption vs. O adsorption), not by temperature 5. For init_data_sys, entries with the same atom count should also be merged using dpdata 5. dpdata automatically handles differing atom orderings during append()—only atom types and counts must match 5.

Sampling Imbalance and prob_sys_size#

stop_batch is a global parameter; frames are sampled by prob_sys_size (proportional to frame count) by default 5. A real Bi-Mo-O example showed O-atom adsorption at only 6.67% accuracy vs. 90.48% for O₂ adsorption because the O-atom data had fewer frames and received disproportionately fewer training samples 5. Remedies:

• Duplicate underrepresented entries in init_data_sys
• Use "prob_uniform" to give all systems equal weight
• Increase stop_batch (400,000–800,000 for heterogeneous multi-component systems; 100,000 is too low) 5

Active Learning Strategy for Multi-Component Systems#

Staged Temperature Expansion#

The community consensus for complex systems 5:

"model_devi_jobs": [
    {"sys_idx": [0,1,2,3], "temps": [725], ...}, // early: 1 T
    {"sys_idx": [0,1,2,3], "temps": [725, 1100], ...}, // mid: add T
    {"sys_idx": [0,1,2,3], "temps": [725, 1100, 2000], ...} // late: full T
]

Alternatively, the original DP-GEN paper (Cu example) explores the full temperature list from iteration 0 and instead progressively expands sys_idx to add new structure types 5. Both approaches are valid; staged temperature expansion is more conservative for reactive/surface systems.

Initial Dataset for Phase-Transformation Systems#

For systems undergoing surface segregation or volatilization (Bi-Mo-O), bulk + stoichiometric surface alone is insufficient. Required categories 5:

For metallic alloys, the Mg-Y tutorial adds defective structures and amorphous phases using hybrid MC/MD 6.

Critical Rules#

• Never fix bottom-layer atoms in slabs. dpgen strips selective dynamics flags when writing FP inputs; inconsistent force labels degrade the model. Use thick slabs (≥6 layers) instead 5.
• Identical DFT settings (functional, ENCUT, KSPACING, pseudopotentials) across all AIMD and dpgen fp steps 5.
• sel: "auto" is strongly recommended for multi-component systems; inverted sel values cause >99% failed frames 5.
• numb_models must be ≥ 3 for valid ensemble uncertainty 5.

Key Source Files and References#