Abstract
In the field of computational chemistry, predicting bond dissociation energies (BDEs) presents well-known challenges, particularly due to the multireference character of reactive systems. Many chemical reactions involve configurations where single-reference methods fall short, as the electronic structure can significantly change during bond breaking. As generating training data for partially broken bonds is a challenging task, even state-of-the-art reactive machine learning interatomic potentials (MLIPs) often fail to predict reliable BDEs and smooth dissociation curves. By contrast, simple and inexpensive physics-based models, such as the well-established Morse potential, do not suffer from any such limitations. This work leverages the Morse potential to improve reactive MLIPs by augmenting the training data set with inexpensive Morse data along the dissociation pathways. This physics-constrained data augmentation (PCDA) approach results in MLIPs with smooth bond dissociation curves as well as near coupled-cluster level BDEs, all without requiring any expensive multireference quantum mechanical calculations. A case study for methane combustion demonstrates how the PCDA approach can improve an existing reactive MLIP, namely, ANI-1xnr. Not only are the BDEs and bond dissociation curves for all radicals and molecules significantly improved compared to ANI-1xnr but the PCDA-trained MLIP retains the reliability of ANI-1xnr when performing reactive molecular dynamics simulations.
Introduction
Machine Learning Interatomic Potentials (MLIPs) have revolutionized computational chemistry by providing the accuracy of quantum mechanics (QM) at the speed of classical force fields. However, they face a significant challenge in modeling bond dissociation. Predicting how a bond breaks is notoriously difficult because the electronic structure changes drastically as atoms pull apart—a phenomenon known as multireference character.
To train an ML model to handle these transitions, researchers typically require data from extremely expensive QM methods like CASPT2. Without this data, even state-of-the-art models like ANI-1xnr produce "noisy" dissociation curves, predict unphysical energy spikes, and miss bond dissociation energies (BDEs) by as much as 20 kcal/mol. This makes them unreliable for simulating reactive chemistry, such as combustion or complex catalysis.
Methodology
The researchers introduced PCDA-Morse, a workflow that bridges the gap between equilibrium structures and fully dissociated fragments using the Morse potential. The approach relies on four key pillars:
- The Morse Bridge: Instead of running thousands of expensive QM calculations for partially broken bonds, the Morse potential—a simple analytical function—is used to interpolate between stable molecules and separated radicals.
- Active Learning (AL): The system uses an ensemble of ML models to identify "uncertain" configurations, such as elongated bonds. When uncertainty is high, the workflow generates a synthetic data point using the Morse potential rather than a QM solver.
- Teacher-Student Labeling: For structures near equilibrium, a pre-trained MLIP (the "teacher") provides labels. For the "danger zone" of breaking bonds, the Morse potential takes over.
- Energy Shifting (ΔE): To reach high accuracy, Morse curves are shifted to match target BDEs from high-level coupled-cluster calculations (CCSD(T)) or experimental data, allowing the model to learn "gold-standard" physics from a small number of high-quality data points.
The study focused on the CHNO (Carbon, Hydrogen, Nitrogen, Oxygen) chemical space, specifically targeting methane combustion. Reference data included the existing ANI-1xnr (DFT-level) and ANI-1ccx (Coupled-cluster level) datasets. Three new augmented datasets were generated, ranging from ~35k to ~52k points, to optimize for smoothness and coupled-cluster accuracy.
Key Results and Discussion
The PCDA approach demonstrated that a simple, empirical physics model can act as a superior guide for ML models in rare-event regions compared to low-level QM (like DFT), which often fails qualitatively during bond stretching. Key findings include:
- Drastic Error Reduction: The original ANI-1xnr had a Root Mean Square Error (RMSE) for BDEs of 16.3 kcal/mol. The PCDA-trained ANI-Mcc reduced this to just 1.2 kcal/mol, achieving the threshold of "chemical accuracy."
- Elimination of "Ghost" Barriers: Standard MLIPs often predict unphysical energy peaks just before a bond breaks. PCDA eliminates these artifacts, leading to smoother, more realistic reaction pathways.
- Kinetic and Stability Performance: In gas-phase molecular dynamics, PCDA models predicted reaction rate constants that aligned closely with experimental data from the NIST Kinetics Database. Importantly, adding synthetic Morse data did not degrade the model's performance on standard equilibrium properties.
Conclusion
The PCDA-Morse approach provides a computationally efficient pathway to achieving "gold-standard" accuracy in reactive MLIPs. While the current implementation uses a pair-wise assumption that may oversimplify complex environments and does not yet capture non-adiabatic effects like conical intersections, it represents a significant step forward. Future work aims to scale this approach to larger hydrocarbons and drug-like molecules by leveraging extensive experimental BDE databases.
Summary updated