Improving Reliability of Machine Learned Interatomic Potentials With Physics-Informed Pretraining

1. Problem Statement

Machine Learned Interatomic Potentials (MLIPs) have become essential for Molecular Dynamics (MD) simulations due to their balance of accuracy and computational efficiency compared to ab initio methods like Density Functional Theory (DFT). However, MLIPs suffer from a critical Out-of-Distribution (OOD) reliability crisis:

• Unphysical Behavior: When MD simulations explore configurational spaces far from the training data (e.g., high temperatures, structural deformations, or atomic overlaps), MLIPs often predict unphysical energies and forces.
• Simulation Instability: These unphysical predictions lead to catastrophic simulation failures, such as atoms overlapping (violating the Pauli exclusion principle) or structural collapse, rendering long-timescale simulations unreliable.
• Limitations of Current Solutions: Existing strategies like active learning are computationally prohibitive due to iterative retraining costs, while regularization methods often struggle to balance physical constraints with model flexibility.

2. Methodology

The authors propose a Physics-Informed Pretraining framework that leverages simple, classical empirical potentials to inject physical knowledge into MLIPs before they are fine-tuned on high-fidelity quantum mechanical data.

A. The Physics-Based Teacher: Embedded Atom Method (EAM)

Instead of using DFT for pretraining (which is too expensive), the authors train a custom EAM potential to serve as a "teacher."

• Formulation: The EAM energy is defined as the sum of an embedding energy term (electron density), a pairwise interaction term (Morse potential), and an atomic energy term.
• Physical Constraints: Unlike pure data-driven models, the EAM functional form inherently satisfies physical limits: $E(r \to 0) \to \infty$ (preventing atomic overlap) and $E(r \to \infty) \to 0$.
• Training: The EAM parameters are optimized via backpropagation on the reference DFT data to minimize Mean Absolute Error (MAE) for energy, forces, and stress.

B. Pretraining-Finetuning Pipeline

The workflow consists of two main phases:

1. Data Augmentation & Labeling:
   • The original quantum mechanical dataset is perturbed using a two-step algorithm:
     • Targeted Compression: Systematically reducing distances between close atom pairs to create "hard" examples.
     • Random Perturbation: Adding random displacements to atoms.
   • These augmented structures are labeled using the trained EAM potential (not DFT), creating a large dataset of physically consistent but lower-fidelity labels.
2. Pretraining & Finetuning:
   • Pretraining: The target MLIP architecture is trained on the EAM-labeled augmented dataset. This initializes the model weights with physical priors (e.g., correct repulsive behavior).
   • Finetuning: The model is subsequently fine-tuned on the original, high-fidelity DFT data using weight sharing. This preserves the physical intuition learned during pretraining while achieving DFT-level accuracy.

C. Benchmarking Suite for Trajectory Physicality

To evaluate stability, the authors introduced two specific metrics to detect unphysical MD trajectories:

1. Overlapping Atoms: Detects if any atom pair falls below a dynamically determined threshold (based on initial equilibrium distances), indicating a failure of short-range repulsion.
2. Lindemann Index: Measures the root-mean-square fluctuation of atomic positions. High values indicate excessive structural disorder or melting artifacts, signaling instability.

3. Key Contributions

1. Novel Pretraining Strategy: Demonstrated that pretraining MLIPs on simple, physics-constrained empirical potentials (EAM) significantly improves robustness in OOD scenarios without the cost of DFT labeling.
2. Robust Evaluation Metrics: Developed a benchmarking suite specifically designed to detect unphysical artifacts (overlaps and structural collapse) in MD trajectories, moving beyond standard energy/force error metrics.
3. Generalizability: Validated the approach across three diverse material systems (Phosphorus, Silica, and a Lithium-Manganese-Oxygen-Phosphorus subset from Materials Project) and three distinct MLIP architectures (CGCNN, M3GNet, and TorchMD-NET).

4. Results

The study evaluated the method against baselines (vanilla MLIPs) and a Stronger Baseline (SAM optimization).

• MD Stability (Trajectory Physicality):
  • The physics-informed pretraining approach consistently reduced unphysical behaviors.
  • In the Silica and MPtrj datasets, the pretraining method achieved zero atom overlaps in several test cases where baselines failed (e.g., M3GNet on MPtrj dropped from 100% failure rate to 0% for atom overlaps).
  • The Lindemann index violations were significantly reduced, indicating more stable thermal dynamics.
• Prediction Accuracy (EFS):
  • The method was minimally invasive. The fine-tuned models maintained competitive accuracy for Energy, Forces, and Stress (EFS) compared to baselines.
  • In many cases, the pretraining approach actually improved prediction accuracy (lower MAE) compared to the baseline, suggesting that the physical priors helped the model converge to better solutions.
• Structural Fidelity (RDF):
  • Radial Distribution Functions (RDFs) generated by the pretrained models showed superior agreement with DFT reference trajectories, particularly for non-bonded interactions where baselines often produced unphysical short-distance peaks.
• Computational Efficiency:
  • The total overhead of the method (EAM training + augmentation + pretraining) was approximately 17 hours on a single GPU.
  • This is negligible compared to the cost of generating DFT labels for the augmented dataset, which would require prohibitive computational resources.

5. Significance and Conclusion

This work addresses a fundamental bottleneck in computational materials science: the reliability of MLIPs in dynamic, non-equilibrium environments.

• Theoretical Insight: It proves that injecting "soft" physical constraints via a pretraining phase can guide neural networks to learn physically consistent representations, particularly for repulsive interactions that are often underrepresented in equilibrium training data.
• Practical Impact: The framework offers a cost-effective pathway to build robust MLIPs for large-scale MD simulations, reducing the risk of simulation crashes and unscientific results.
• Future Directions: While the current implementation relies on EAM, the authors suggest that future iterations could use more sophisticated "foundational" MLIPs as teacher models to handle systems with higher elemental complexity, provided those teachers also adhere to physical constraints.

In summary, the paper establishes that physics-informed pretraining is a highly effective strategy for enhancing the reliability, stability, and accuracy of machine learning potentials for molecular dynamics simulations.