NequIP

Official Resources

• Homepage: https://github.com/Teoroo-CMC/ParAutomatik
• Source Repository: https://github.com/Teoroo-CMC/ParAutomatik
• License: MIT License

Overview

ParAutomatik is a cutting-edge workflow automation tool designed for the parameterization of Density Functional Tight Binding (DFTB) models using Machine Learning. It addresses one of the biggest bottlenecks in semi-empirical methods: the difficulty of creating accurate parameters. By combining high-throughput DFT calculations with Neural Network training, ParAutomatik automates the fitting of repulsive potentials and electronic parameters, significantly accelerating the development of transferable DFTB models.

Scientific domain: Machine Learning, Parameterization, Tight-Binding Target user community: Method developers, Researchers needing custom potentials for specific chemistries

Theoretical Methods

• SCC-DFTB: Self-Consistent Charge Density Functional Tight Binding.
• Machine Learning: Neural Networks (PyTorch) for regression.
• Active Learning: Iterative improvement of training sets.
• Repulsive Potential Fitting: Fitting the difference between DFT and electronic DFTB energy ($E_{rep} = E_{DFT} - E_{elec}$).
• Spline Interpolation: Generation of traditional spline-based repulsive potentials.

Capabilities (CRITICAL)

• Dataset Generation: Automated sampling of geometries (dimers, trimers, clusters).
• Reference Calculation: Automatic execution of DFT reference runs (via ASE calculators).
• ML Fitting: Training of NNs to predict repulsive energies.
• Validation: Automatic benchmarking against DFT forces and energies.
• Output Generation: Production of .skf files compatible with DFTB+.

Key Strengths

Automation:

• Replaces manual "by-eye" fitting or simple polynomial fits with robust ML workflows.
• Handles the complexity of multi-element parameterization consistency.

Physics-Informed ML:

• Uses Neural Networks to learn the complex environment dependence of repulsion, or to generate improved 2-body splines.

Reproducibility:

• Defines the parameterization protocol as code, making the provenance of potentials clear.

Inputs & Outputs

• Inputs:
  • List of elements (e.g., ["C", "H", "O"]).
  • Reference method definition (e.g., PBE/def2-TZVP).
  • Configuration settings (cutoff distances, NN architecture).
• Outputs:
  • dataset.db: ASE database of training structures.
  • model.pt: Trained PyTorch model.
  • *-*.skf: Final Slater-Koster files ready for DFTB+.
  • report.pdf: Validation report showing RMS errors.

Interfaces & Ecosystem

• ASE: Built entirely around the Atomic Simulation Environment.
• DFTB+: The primary target engine for the resulting parameters.
• PyTorch: The ML backend.
• dftbpara: Compatible with/Alternative to other fitting tools.

Advanced Features

• Delta Learning: Predicting the correction to a baseline model.
• Force Matching: Training on forces (gradients) as well as energies for better dynamics stability.

Performance Characteristics

• Speed: Training takes minutes to hours (GPU supported); Reference data generation is the bottleneck.
• Scalability: Can parameterize complex multi-element sets if training data is sufficient.

Computational Cost

• Moderate: dominated by the cost of the ab initio reference calculations.

Limitations & Known Constraints

• Two-Body Limit: Standard .skf format is strictly two-body; if ParAutomatik is used to generate these, it compresses many-body physics into 2-body terms (loss of accuracy). (Note: Newer versions may support Many-Body Repulsion).
• Data Hungry: Quality depends entirely on the coverage of the training set.

Comparison with Other Codes

• vs TBFIT: TBFIT fits electronic band structures (SK integrals); ParAutomatik typically fits the repulsive part (total energy/forces). They are complementary.
• vs Hotbit: Hotbit is a code that can fit parameters manually/scripted; ParAutomatik is a dedicated ML workflow.
• Unique strength: Bringing modern ML workflows to the tedious task of repulsive potential fitting.

Application Areas

• MOFs/COFs: Creating parameters for novel porous materials.
• Catalysis: Tuning parameters for specific metal-organic interfaces.
• High-Pressure: Fitting potentials for extreme conditions where standard sets fail.

Best Practices

• Span Config Space: Ensure training data covers the distances/angles seen in production runs.
• Check Limits: Verify behavior at short range (avoid hole collapse) and long range (smooth cutoff).
• Validate: Always run a test calculation on a system not in the training set (e.g., a crystal bulk modulus).

Community and Support

• GitHub: Developed by the Teoroo-CMC group (Calcula).
• Documentation: Jupyter notebook tutorials available.

Verification & Sources

Primary sources:

1. Repository: https://github.com/Teoroo-CMC/ParAutomatik
2. Publications by the Teoroo group (e.g., on DFTB parameterization).

Verification status: ✅ VERIFIED

• Source code: OPEN (MIT)
• Functionality: Functional ML workflow.