ML Interatomic Potential Simulations

• Machine learning interatomic potential-based simulations are atomistic methods that use ML surrogate models trained on quantum-mechanical data to predict energies, forces, and stresses with near-first-principles accuracy.
• They decompose total energy into local atomic contributions using invariant descriptor schemes and various regression models, ensuring robustness and scalability across diverse materials systems.
• These simulations enable efficient studies of phase transformations, defect dynamics, and material behaviors under extreme conditions, validated against both quantum and experimental benchmarks.

Machine Learning Interatomic Potential-Based Simulations

Machine learning interatomic potential (MLIP)–based simulations are a class of atomistic simulation methodologies that leverage machine-learned surrogate models, trained on quantum-mechanical reference data, to provide energies, forces, and stresses in molecular dynamics (MD) or Monte Carlo (MC) simulations with near-first-principles accuracy and orders-of-magnitude lower computational cost. This paradigm has transformed large-scale and long-timescale simulations across diverse areas, including materials under extreme conditions, defect dynamics, phase transformations, and spectroscopic property predictions.

1. Theoretical Foundations and Model Architectures

All MLIPs decompose the total potential energy of a system into the sum of local atomic contributions, $E_{\mathrm{tot}} = \sum_i E_i$, where $E_i$ is a function of a local descriptor characterizing the chemical environment of atom $i$ within a cutoff radius. Essential invariances—translational, rotational, and permutational—are built in via mathematical descriptors. Principal frameworks include:

• Descriptor Schemes:
  • Behler–Parrinello symmetry functions: radial and angular environment encoding; see, e.g., La–Si–P ANN-ML potential (Tang et al., 10 Jun 2025).
  • SOAP and turboSOAP: rotationally invariant representations via neighbor densities expanded in spherical harmonics and radial bases (Deringer et al., 2016, Hamedani et al., 8 Oct 2025).
  • Bispectrum/SNAP: 4D hyperspherical harmonics and third-order bispectrum invariants, e.g., SNAP carbon (Willman et al., 2022).
  • Moment Tensor Potentials (MTP): expansion in radial Chebyshev polynomials and contracted angular moment tensors (Thomazini et al., 28 Dec 2025).
  • Atomic Cluster Expansion (ACE): linear expansion in systematically complete basis functions (Leimeroth et al., 5 May 2025).
• ML Regression Models:
  • Kernel regression (GAP, SOAP): kernel ridge regression on descriptors, yielding sparse, nonparametric models (Deringer et al., 2016, Hamedani et al., 8 Oct 2025).
  • Feed-forward neural networks (ANN, Deep Potential, UF3): mapping from descriptors to energies via multilayer perceptrons (Tang et al., 10 Jun 2025, Taormina et al., 11 Nov 2025, Thong et al., 2022, Zhang et al., 2024).
  • Graph neural networks (MACE, NequIP, HIPNN): message-passing architectures for geometric equivariance and higher-body correlations (Leimeroth et al., 5 May 2025, Brunken et al., 28 May 2025, Alzate-Vargas et al., 23 Jul 2025).
  • Hybrid and physics-augmented models: e.g., SNAP+ZBL for short-range repulsion in irradiation environments (Bhardwaj et al., 5 Feb 2025), PINN (Mishin, 2021).

2. Training Strategies and Active Learning Protocols

MLIP construction involves assembling a comprehensive quantum-mechanical reference database (energies, forces, and optionally stresses/virials), spanning the relevant phase space:

• Sampling of Atomic Environments:
  • Equilibrium lattice and polymorphs, strained and distorted configurations, surfaces and interfaces, point and extended defects, and high-temperature liquids and amorphous states (Willman et al., 2022, Deringer et al., 2016, Hamedani et al., 8 Oct 2025, Alzate-Vargas et al., 23 Jul 2025).
  • For systems with complex chemistry (e.g., AlN or KNbO₃), genetic algorithms, random structure search, AIMD, and normal-mode sampling are used (Taormina et al., 11 Nov 2025, Thong et al., 2022).
  • Coverage of far-from-equilibrium, cascade-generated, or mixed-phase environments is essential for robustness (e.g., SiC collision cascades (Hamedani et al., 8 Oct 2025), Nb cascade simulations (Bhardwaj et al., 5 Feb 2025)).
• Active Learning and Dataset Curation:
  • Iterative “on-the-fly” discovery of underrepresented configurations via uncertainty quantification (committee models, predictive variance) and selection for additional ab initio labeling (Zhang et al., 2024, Bhatia et al., 16 Jun 2025, Alzate-Vargas et al., 23 Jul 2025).
  • Farthest-point sampling and genetic search to maximize configurational diversity (Zhang et al., 2024, Taormina et al., 11 Nov 2025).
  • Ensemble knowledge distillation (EKD) for force learning when forces are unavailable directly from quantum chemistry (Matin et al., 18 Mar 2025).
• Regression and Optimization:
  • Weighted least-squares for linear models (SNAP, UF3, MTP), with explicit group weight optimization via genetic algorithms or Bayesian search (Willman et al., 2022, Taormina et al., 11 Nov 2025, Thomazini et al., 28 Dec 2025).
  • Nonlinear neural networks trained via stochastic (Adam) optimization, with regularization and early stopping (Alzate-Vargas et al., 23 Jul 2025, Thong et al., 2022).
  • Hyperparameter tuning for cutoff radii, expansion orders, basis sizes, network depths, and regularization (Willman et al., 2022, Taormina et al., 11 Nov 2025, Leimeroth et al., 5 May 2025).

3. Performance Metrics, Validation, and Transferability

Robust MLIPs are benchmarked through direct comparison to quantum-mechanical and experimental observables:

• Accuracy Metrics:
  • RMSEs on energies and forces (e.g., ≲1 meV/atom and ≲0.05 eV/Å for best-in-class models (Willman et al., 2022, Hamedani et al., 8 Oct 2025)).
  • Elastic constants, lattice and cohesive energies, phonon dispersions, and thermal properties converge closely to DFT/experiment (Hamedani et al., 8 Oct 2025, Taormina et al., 11 Nov 2025, Zhang et al., 2024, Thong et al., 2022).
  • Defect formation energies and migration barriers, e.g., in UC, Nb, SiC, and AlN (Alzate-Vargas et al., 23 Jul 2025, Bhardwaj et al., 5 Feb 2025, Hamedani et al., 8 Oct 2025, Taormina et al., 11 Nov 2025).
• Physical Properties:
  • Phase diagrams and transition/melting lines reproduced to within a few percent of QMD/experiment over wide P–T ranges (e.g., carbon up to 5 TPa/20,000 K (Willman et al., 2022), alumina to 200 GPa (Zhang et al., 2024)).
  • Dynamic/kinetic simulations: domain walls, phase transitions in perovskites, melting and recrystallization, irradiation cascades, epitaxial growth (Thong et al., 2022, Robredo-Magro et al., 21 Nov 2025, Hamedani et al., 8 Oct 2025, Taormina et al., 11 Nov 2025).
  • Spectroscopic properties with ML-predicted dipoles and IR spectra at DFT fidelity and 100× speedup (Bhatia et al., 16 Jun 2025).
• Transferability and Limitations:
  • Many MLIPs exhibit smooth potential energy surfaces and retain stability over nanosecond-to-microsecond MD timescales and millions to billions of atoms (Willman et al., 2022, Chen et al., 2023, Zhang et al., 2024).
  • Extrapolation to high P–T, amorphous, or defective regimes is critically dependent on training data diversity; failure modes include unphysical forces or PES “blow-up” at unsampled conditions (Leimeroth et al., 5 May 2025, Robredo-Magro et al., 21 Nov 2025).
  • Extensions to new chemistries require full retraining; high descriptive orders or species expansion may be necessary for complex materials (Willman et al., 2022, Taormina et al., 11 Nov 2025, Zhang et al., 2024).

4. High-Performance Simulation and Scalability

Modern MLIP-based MD can match or exceed classical force fields in computational performance through algorithmic and hardware innovations:

• Algorithmic Efficiency:
  • O(N) scaling is achieved via local environment decomposition and neighbor-list construction (Willman et al., 2022, Deringer et al., 2016).
  • Kernel and descriptor optimizations (e.g., turboSOAP, rational function interpolation) significantly accelerate evaluation times (Chen et al., 2023, Hamedani et al., 8 Oct 2025).
  • GPU and many-core parallelization: e.g., linear scaling on OLCF Summit and Sunway supercomputer, enabling direct simulation of 10⁶–10⁹ atoms, 31 ps/step/atom for 52×10⁹ atoms (Willman et al., 2022, Chen et al., 2023).
• Distributed Inference and Graph-Level Parallelism:
  • DistMLIP implements zero-redundancy, graph-level parallelization for modern GNN-based MLIPs, bypassing the cubic-scaling ghost-atom bottleneck of spatial domain decomposition (Han et al., 28 May 2025).
  • Scalability in nanosecond–100 ns regimes for glueball-sized systems, supporting MACE, CHGNet, TensorNet, and eSEN on multi-GPU clusters (Han et al., 28 May 2025).
• Software Ecosystem:
  • Integration with LAMMPS, VASP, ASE, and JAX-MD; plug-in support for major MLIP models; user-friendly workflows for training, validation, and MD execution (Brunken et al., 28 May 2025, Chen et al., 2023, Han et al., 28 May 2025).
  • Open-source frameworks for training, data management, and workflow reproducibility (Ceriotti, 2022, Brunken et al., 28 May 2025).

5. Applications and Case Studies

MLIP-based simulations have enabled decisive progress across multiple frontiers:

Material/System	MLIP Type	Applications	Reference
Carbon (extreme P–T)	SNAP	Phase diagram, melting, shock Hugoniot	(Willman et al., 2022)
Amorphous Carbon	GAP (SOAP)	Liquid/amorphous structure, surface energy, recon.	(Deringer et al., 2016)
SiC (3C)	GAP (turboSOAP)	Radiation damage, threshold energies, melting	(Hamedani et al., 8 Oct 2025)
Uranium Monocarbide	HIP-NN	Equations of state, defects, diffusion	(Alzate-Vargas et al., 23 Jul 2025)
Alumina/Al₂O₃	NEP (ACE/NN)	Phase diagram, amorphous structure, thermal props	(Zhang et al., 2024)
AlN (epitaxy)	UF3 (splines)	Epitaxial growth, dislocation core, surface energy	(Taormina et al., 11 Nov 2025)
Perovskites	DP-NN/GAP/Allegro	Phase transitions, domain walls, vortices, phonon	(Thong et al., 2022Robredo-Magro et al., 21 Nov 2025)
Nb (irradiation)	SNAP (bispectrum)	Cascade simulation, SIAs, defect statistics	(Bhardwaj et al., 5 Feb 2025)
La–Si–P system	ANN–ML (BP)	Melting, nucleation, growth kinetics	(Tang et al., 10 Jun 2025)
Organics (IR spectra)	MACE (active AL)	High-throughput anharmonic IR spectrum prediction	(Bhatia et al., 16 Jun 2025)

These case studies demonstrate systematic recovery of quantum accuracy (energies, forces, and derived properties) and enable explorations (e.g., microsecond-scale kinetics, high-pressure phase transitions, defect evolution under irradiation, epitaxial morphologies) that were formerly intractable to ab initio simulation.

6. Practical Guidance, Limitations, and Future Directions

Best Practices:

• Design training datasets by stratified or active sampling spanning all relevant structures, phases, and processes; employ ensemble/uncertainty metrics to identify gaps (Zhang et al., 2024, Willman et al., 2022, Bhatia et al., 16 Jun 2025).
• Validate on both interpolation (test-set error) and extrapolation (PES smoothness, dynamical stability, out-of-sample benchmarks) (Leimeroth et al., 5 May 2025, Robredo-Magro et al., 21 Nov 2025).
• Use MLIPs appropriate for the task: linear ACE/MTP for large, fast MD; equivariant GNNs (MACE, NequIP) for highest accuracy; hybrid and physics-augmented schemes for improved extrapolation (Leimeroth et al., 5 May 2025, Mishin, 2021).

Limitations:

• Extrapolation to unsampled high-energy or electronic environments (e.g., high-temperature plasma, explicit electronic degrees of freedom) remains challenging; quantum and ML extension to include electronic entropy/generalized free-energy terms is ongoing (Willman et al., 2022).
• Omission of long-range Coulomb terms limits applicability to ionic/ferroelectric materials unless explicitly modeled or included in descriptors (Robredo-Magro et al., 21 Nov 2025, Thong et al., 2022).
• Some MLIPs require extensive retraining for multicomponent or off-stoichiometry chemistries, though frameworks for transfer learning and multi-fidelity hybridization are emerging (Matin et al., 18 Mar 2025, Leimeroth et al., 5 May 2025).

Prospects:

• Integration of property-predictive models beyond energetics (e.g., dipoles, dielectric response, spectra) for full many-body trajectory observables (Ceriotti, 2022, Bhatia et al., 16 Jun 2025).
• Further automation and scale-up with distributed inference (multi-GPU, multi-node) for MD of 10⁶–10⁹ atom systems, including out-of-domain active learning (Han et al., 28 May 2025, Chen et al., 2023).
• Coupling with physics-based models (hybrid MLIP+force-field, PINN) for robust transferability and interpretability (Mishin, 2021).
• Advanced deployment in inverse design, structure search, and in operando spectroscopy, harnessing MLIP-driven large-scale dynamics as a standard scientific tool (Zhang et al., 2024, Bhatia et al., 16 Jun 2025).

---

References

• (Willman et al., 2022) Machine Learning Interatomic Potential for Simulations of Carbon at Extreme Conditions
• (Hamedani et al., 8 Oct 2025) SiC-TGAP: A machine learning interatomic potential for radiation damage simulations in 3C-SiC
• (Thong et al., 2022) Machine-learning interatomic potential for molecular dynamics simulation of ferroelectric KNbO3 perovskite
• (Taormina et al., 11 Nov 2025) Machine-learning interatomic potential for AlN for epitaxial simulation
• (Zhang et al., 2024) Exploring the energy landscape of aluminas through machine learning interatomic potential
• (Robredo-Magro et al., 21 Nov 2025) Minimalist machine-learned interatomic potentials can predict complex structural behaviors accurately
• (Alzate-Vargas et al., 23 Jul 2025) Atomistic modeling of uranium monocarbide with a machine learning interatomic potential
• (Deringer et al., 2016) Machine-learning based interatomic potential for amorphous carbon
• (Bhardwaj et al., 5 Feb 2025) A Robust Machine Learned Interatomic Potential for Nb: Collision Cascade Simulations with accurate Defect Configurations
• (Chen et al., 2023) TensorMD: Scalable Tensor-Diagram based Machine Learning Interatomic Potential on Heterogeneous Many-Core Processors
• (Leimeroth et al., 5 May 2025) Machine-learning interatomic potentials from a users perspective: A comparison of accuracy, speed and data efficiency
• (Thomazini et al., 28 Dec 2025) A Simple and Efficient Non-DFT-Based Machine Learning Interatomic Potential to Simulate Titanium MXenes
• (Tang et al., 10 Jun 2025) Developing a Neural Network Machine Learning Interatomic Potential for Molecular Dynamics Simulations of La-Si-P Systems
• (Brunken et al., 28 May 2025) Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems
• (Matin et al., 18 Mar 2025) Ensemble Knowledge Distillation for Machine Learning Interatomic Potentials
• (Bhatia et al., 16 Jun 2025) Leveraging active learning-enhanced machine-learned interatomic potential for efficient infrared spectra prediction
• (Mishin, 2021) Machine-learning interatomic potentials for materials science
• (Ceriotti, 2022) Beyond potentials: integrated machine-learning models for materials
• (Morrow et al., 2021) Indirect Learning of Interatomic Potentials for Accelerated Materials Simulations
• (Han et al., 28 May 2025) DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials

---