ChemDX_NEB_MLIP

Machine Learning Interatomic Potential (MLIP) project using the ChemDX database

🏆 Developed as part of the KRICT ChemDX Hackathon 2025

---

🧠 Motivation

Can we systematically supply missing data that greatly improves MLIPs?

• Many machine learning interatomic potentials (MLIPs) struggle with poor transferability and instability when simulations explore configurations far from equilibrium. Traditional datasets are heavily biased toward near-equilibrium or randomly-perturbed structures, leaving transition states and reaction pathways underrepresented.

• This project investigates whether explicitly adding minimum energy path (MEP) data from Nudged Elastic Band (NEB) calculations can systematically enhance MLIP performance, particularly for dynamic simulations.

---

⚙️ Systems Studied

We focused on surface adsorption systems:

System	Surface Type	Adsorbate	Stable Site
Au on Al(100)	Pure metal surface	Au	Hollow
Au on AlPd(100)	Alloy surface	Au	Hollow

These systems provide well-defined diffusion pathways that are ideal for evaluating the impact of NEB data on MLIP performance.

---

🔬 Approach Overview

1️⃣ Data Generation

Atomic configurations were generated using Atomic Simulation Environment (ASE) with three complementary sampling strategies:

Method	Purpose	Configuration Type
Relaxation	Stable geometries	Energy minima
Molecular Dynamics (MD)	Thermal fluctuations	Near-equilibrium
NEB	Diffusion pathways	Transition states

Two datasets were constructed:

Dataset	Sampling Methods	Configuration Space Coverage
Set #1	Relaxation + MD	Near-equilibrium only
Set #2	Relaxation + MD + NEB	Near-equilibrium + Transition states

2️⃣ MLIP Training

We trained neural network potentials using the Atomistic Machine-learning Package (AMP).

Model	Training Data	Purpose
Model #1	Set #1 (Relax + MD)	Baseline MLIP
Model #2	Set #2 (Relax + MD + NEB)	NEB-enhanced MLIP

Both models used identical network architectures to ensure that performance differences arise from data quality, not model complexity.

3️⃣ Model Evaluation

Models were evaluated using ANN-driven Molecular Dynamics simulations.

Evaluation metrics included:

• Force prediction error
• Energy conservation during MD
• Structural stability under finite temperature simulations

---

📈 Key Results

🧩 Dataset Impact

Including NEB configurations significantly expands coverage of high-energy and transition-state regions, which are missing in conventional MD-only datasets.

• Tutorial: Au on Al(100) data generation
• Tutorial: Au on AlPd(100) data generation

⚛️ Force Prediction Accuracy and MD Stability

The NEB-enhanced MLIP demonstrated:

• Lower force errors even for near-equilibrium MD trajectories without diffusion process
• Stable long MD trajectories
• Proper energy conservation

Model Type	Force Error	Energy Conservation
Relax + MD	High	❌
Relax + MD + NEB	Low	✅

• Tutorial: Au on Al(100) trained on set 1
• Tutorial: Au on Al(100) trained on set 2
• Tutorial: Au on AlPd(100) trained on set 1
• Tutorial: Au on AlPd(100) trained on set 2

---

🧠 Key Insight

Strategically adding transition-state data can be more effective than increasing model size.

This work highlights a data-centric pathway to improving MLIPs: identifying physically important but under-sampled regions (like saddle points) and systematically incorporating them into the training set.

---

🛠 Tools Used

Tool	Role
ASE	Structure generation, Relaxation, MD, NEB
AMP	Neural network potential training
ChemDX Database	Initial structures and metadata
Python	Workflow orchestration
NumPy / Matplotlib	Data analysis and visualization

---

📂 Repository Structure

jupyter_notebook/
 ├── Au_on_Al/
 │    ├── 01_data_generation.ipynb
 │    ├── 02_training_MD_set_1.ipynb
 │    └── 02_training_MD_set_2.ipynb
 │
 ├── Au_on_AlPd/
 │    ├── 01_data_generation.ipynb
 │    ├── 02_training_MD_set_1.ipynb
 │    └── 02_training_MD_set_2.ipynb

docs/images/        # Figures and animations used in README

---

🚀 How to Reproduce

1. Install dependencies
2. Run data generation notebooks
3. Train MLIP models for Set #1 and Set #2
4. Run MD evaluation notebooks to compare stability and accuracy

---

🔍 Conclusion

Adding NEB-based transition-state data leads to substantial improvements in MLIP performance by:

• Expanding configuration space coverage
• Reducing force errors in off-equilibrium regions
• Enabling stable and physically reliable MD simulations

This demonstrates that targeted data augmentation is a powerful strategy for building more transferable and stable ML interatomic potentials.

---

🙏 Acknowledgments

Developed during the KRICT ChemDX Hackathon 2025.
We thank the ChemDX team for organizing the event and providing access to the database.

---

📬 Contact

For questions or contributions, please contact:

• In Won Yeu (GitHub: @InWonYeu)
• Jaeseon Kim (GitHub: @kasee0)
• Seojin Jeon (GitHub: @sjjeon6783)