LES Distillation Benchmark

Training data, training scripts, fitted machine-learning interatomic potentials (MLIPs), molecular dynamics (MD) setups, analysis scripts, and raw data for the Latent Ewald Summation (LES) distillation benchmark across three chemical systems.

Systems

Folder	System	Property	Method
System1_water/	Bulk liquid water	IR spectrum, BEC benchmark	MACE two-stage training
System2_HCl/	2 M HCl solution	H₃O⁺ IR difference spectrum	MACE two-stage training
System3_TiO2-water/	TiO₂(110)-water interface	Surface IR spectrum, water density profile	CACE; force-only fine-tuning of MACE-MP-0(L)

Repository layout

les_distill/
├── System1_water/
│   ├── Datasets/                  # train/test xyz (RPBE-D3, UMA-M)
│   ├── MLIP_and_MD_setups/        # MACE training script and MD utilities
│   ├── water_IR/                  # IR spectrum data and plotting
│   ├── water_BEC/                 # BEC benchmark data and notebook
│   ├── water_MLIPs_RPBE-D3_sampled_configs/   # trained MACE models
│   └── water_MLIPs_UMA-M-MD_sampled_configs/  # trained MACE models (learning curve)
├── System2_HCl/
│   ├── Datasets/                  # train/test xyz (2M HCl, UMA-S sampled)
│   ├── MLIP_and_MD_setups/        # MACE training script and MD/BEC analyses
│   ├── MLIPs/                     # trained MACE models 
│   └── HCl_solution_IR/           # H₃O⁺ IR spectra data and plotting
└── System3_TiO2-water/
    ├── MLIP_and_MD_setups/        # CACE training, NVT MD, BEC analysis, opt
    ├── MLIPs/                     # CACE models (direct fit, fine-tuned, student)
    └── surface_water_MD_results/  # IR spectra, density profiles

Requirements

Each system has its own dependency list. In general:

• System 1 & 2: MACE, ASE, NumPy, Matplotlib, SciPy
• System 3: CACE, ASE, PyTorch, NumPy, Matplotlib, SciPy