Machine-Learned Interatomic Potentials for Photophysics

Author: Lucas Sanchez Garay
Institution: Department of Physics, University of Warwick
Date: 13 March 2025

This repository is the code and data companion to my undergraduate dissertation:

Machine-Learned Interatomic Potentials for Photophysics
Submitted: May 2025
PDF available in docs/Dissertation_Final_Report.pdf

---

📖 Project Overview

This project investigates whether a short-range, equivariant graph neural network — MACE-OFF23 — pre-trained on ground-state molecular data can be fine-tuned to accurately predict excited-state potential energy surfaces (PES) for ultrafast photophysical processes.

Two representative case studies were chosen:

• Oxirane angle opening — showcasing a strong S₀/S₁ conical intersection around 108°.
• Thymine dimerisation — exploring covalent C₅–C₅′ bond formation with varying distances and angles, highlighting spin contamination under SF-TDDFT.

This work combines quantum chemistry calculations, machine learning, and molecular dynamics to explore the feasibility of transfer learning from ground- to excited-state PES.

---

🎯 Motivation

Excited-state molecular dynamics is computationally expensive, with ab initio methods like SF-TDDFT or multireference approaches scaling poorly with system size.
Machine-learned interatomic potentials (MLIPs) offer a route to bridge the gap:

1. Train on a small set of high-fidelity data.
2. Reuse pre-trained models to drastically reduce training cost.
3. Preserve ground-state accuracy via replay fine-tuning.

---

🔬 Methods Summary

• Electronic structure calculations:
  • SF-TDDFT (ORCA) for excited states
  • DFT for ground-state forces where required
• ML architecture: MACE-OFF23 equivariant GNN
• Training strategy:
  • Replay fine-tuning with a secondary output head for excited states
  • Active learning: short MD bursts near crossing regions to sample challenging geometries
• Evaluation metrics: RMSE in energies and forces; qualitative PES scans; MD trajectory stability

---

📊 Key Results

Predicted Potential Energy Surface for thymine dimerisation, fine-tuned vs raw models.

Predicted Potential Energy Surface for oxirane's angle opening reaction, fine-tuned vs raw models.

Main findings:

• Fine-tuned models achieve accuracy comparable to models trained from scratch, but with far less data.
• Replay fine-tuning helps mitigate catastrophic forgetting of ground-state accuracy.
• Capturing the topology of conical intersections still requires high-fidelity excited-state data.

---

🗂 Repository Structure

.
├── docs/                       # Dissertation PDF + figure exports
├── mace-suite/                 # Adapted MACE library, datasets and runs for each model/system
├── oxirane/                    # Scripts/data for oxirane case study (isolated geometry optimisations and Spin-Flip energy calculations)
├── thymine/                    # Scripts/data for thymine dimerisation case study (isolated geometry optimisations and Spin-Flip energy calculations)
├── results/                    # Results from testing suite plotted to measure energy RMSE error per state, MD approximation by model and Potential Energy Surface predictions
├── testing-suite/              # MD simulations and ORCA energy calculations on some snapshots as ground truth
├── wrappers/                   # Helper functions and job submission wrappers for ORCA jobs
└── .gitattributes

---

📄 The Dissertation

The full written dissertation can be found at:

docs/Dissertation_Final_Report.pdf

It contains:

• Detailed theoretical background
• Computational setup and parameters
• Full results & discussion
• All original figures

---

⚙️ Running the Code

This repo is not designed as a polished software package — it is a research archive.
However, if you wish to reproduce results or explore the code:

1. Install Python 3.10 and the following packages:

   pip install numpy scipy matplotlib pandas jupyter tqdm
   pip install torch ase mace-torch

2. Place your own ORCA executables path in the wrappers/orca_wrappers in the orca_profile line if regenerating TDDFT data.
3. Explore scripts in oxirane/ and thymine/ for data generation
4. Explore scripts in mace-suite/runs for model training
5. testing-suite/ and results/ contain useful evaluation metrics used

---

📚 Citation

If referencing this work:

Lucas Sanchez Garay (2025), Machine-Learned Interatomic Potentials for Photophysics. Department of Physics, University of Warwick, Coventry CV4 7AL, United Kingdom. Submitted: 13 March 2025.

---

🙏 Acknowledgements

• Supervision & guidance from my project supervisor Professor Nicholas Hine
• MACE developers for the open-source MLIP framework
• ORCA developers for making high-quality quantum chemistry accessible

Bibliography

• Neese, F. The ORCA program system Wiley Interdiscip. Rev.: Comput. Mol. Sci., 2012, 2, 1, 73–78 (DOI: 10.1002/wcms.81)
• Kovács, D. P. et al. MACE‑OFF: Short‑Range Transferable Machine Learning Force Fields for Organic Molecules. J. Am. Chem. Soc. 2025, 147 (21), 17598–17611. DOI: 10.1021/jacs.4c07099
• Ask Hjorth Larsen et al. The Atomic Simulation Environment—A Python library for working with atoms. J. Phys.: Condens. Matter Vol. 29 273002, 2017
• Physical Review Letters Matplotlib style: github.com/hosilva/physrev_mplstyle