Spectral Neighbor Analysis Potential (SNAP/qSNAP)

• SNAP/qSNAP are advanced machine-learned interatomic potentials that encode local atomic environments using rotationally and permutationally invariant bispectrum descriptors.
• They employ a linear energy model and a quadratic extension to capture many-body interactions, achieving near-DFT accuracy in simulating metals, alloys, and extreme conditions.
• Validated in applications from nanoparticle thermodynamics to phase stability at extreme pressures, these models balance computational efficiency with quantum-level fidelity.

The Spectral Neighbor Analysis Potential (SNAP) and its quadratic extension (qSNAP) are machine-learned interatomic potential frameworks that rigorously encode local atomic environments using high-order, rotationally and permutationally invariant bispectrum descriptors. Designed to interpolate between ab initio quantum accuracy and the computational efficiency of empirical force fields, SNAP/qSNAP have demonstrated quantum-level fidelity in complex materials modeling tasks ranging from metallic alloys and high-entropy ceramics to extreme-pressure phases and nanoparticle thermodynamics (Bideault et al., 2024, Wood et al., 2017, Willman et al., 2022).

1. Mathematical Foundations and Descriptor Construction

SNAP encodes the atomic environment of each atom $i$ as a local neighbor density,

$$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$

where $f_c$ is a smooth cutoff function vanishing at $r_{ij} > r_{cut}$, and the sum includes all neighbors within the cutoff radius. To ensure rotation and permutation invariance and to capture many-body correlations, $\rho_i$ is projected onto a basis of 4D hyperspherical (SO(4)) harmonics $U^j_{mm'}(\theta, \phi, \chi)$: $$\rho_i(\mathbf r) = \sum_{n} u^{(i)}_{n}\,R_n(r)\,U^j_{mm'}(\Omega)$$ where $n \equiv (j, m, m')$, $R_n(r)$ is an orthonormal radial basis, and $J_{max}$ sets the maximum angular resolution (typically $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$0 defines the band limit).

From the expansion coefficients $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$1, rotationally invariant bispectrum components are constructed via Clebsch–Gordan–type couplings: $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$2 The full bispectrum feature vector $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$3 collects all distinct components up to the truncation.

For multi-element systems, explicit multi-element SNAP (EME-SNAP) generalizes this expansion by defining partial neighbor densities for each chemical species, leading to bispectrum terms $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$4 which scale as $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$5 in descriptor count (Cusentino et al., 2020).

2. Energy Model: Linear SNAP and Quadratic Extension (qSNAP)

Linear SNAP

The atomic energy within the SNAP framework is given as a linear function of the bispectrum descriptors: $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$6 where $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$7 and $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$8 are fitting coefficients determined via a weighted least-squares regression over a DFT reference dataset; the total energy sums all atomic contributions. Analytical derivatives with respect to atomic positions provide forces and stress tensors required for MD.

Quadratic SNAP (qSNAP)

The quadratic SNAP (qSNAP) augments the expressiveness of SNAP by including quadratic terms in the bispectrum components: $$\rho_i(\mathbf r)=\sum_{j} f_c(r_{ij})\,\delta(\mathbf r-\mathbf r_{ij})$$9 The quadratic coefficients $f_c$0 systematically capture many-body effects, allowing improved treatment of strongly anharmonic, defect-rich, or undercoordinated environments. The increase in parameter count ($f_c$1) necessitates larger, more diverse training datasets and robust regularization against overfitting (Wood et al., 2017).

3. Training and Optimization Workflow

SNAP and qSNAP potentials are trained against a database of DFT-calculated energies, forces, and (optionally) stresses for structurally and chemically diverse configurations. The fit minimizes a weighted objective function: $f_c$2 Group-wise weights $f_c$3 are tuned (often via an outer-loop genetic algorithm or cross-validation) to reflect the intended application targets (e.g., elastic properties, defect energetics, phase stability). The hyperparameters $f_c$4 and $f_c$5 are chosen to minimize validation RMSE.

For qSNAP, the data volume must be sufficient to avoid overfitting; random holdout cross-validation, principal component analysis, and genetic algorithm optimization of group weights are employed to ensure generalization (Wood et al., 2017, Bideault et al., 2024).

4. Validation: Accuracy and Transferability

Bulk, Defect, and Surface Properties

SNAP/qSNAP potentials have demonstrated near-DFT accuracy across a wide set of observables:

• Energy RMSE and Force RMSE: For example, for Co q-SNAP yields 4.3 meV/atom (train), 28.6 meV/atom (test) and forces within 54.2–92.7 meV/Å (Bideault et al., 2024).
• Phonon dispersions: Deviation from DFT <0.5 THz in hcp and fcc Co (Bideault et al., 2024).
• Surface energies: Maximum error 47 mJ/m$f_c$6 in Co, compared to >270 mJ/m$f_c$7 for empirical EAM (Bideault et al., 2024).
• Elastic moduli, formation energies, melting points: Systematically within 5–10% (often lower) of DFT.

Nanoparticle Thermodynamics and Size-Scaling

The flexibility and stability of qSNAP permit long-time MD simulations of large systems (e.g., 9201-atom Co nanoparticles, 100 ns trajectories), enabling calculation of converged heat capacity, melting temperatures, and vacancy formation energies, reproducing both DFT and experimental scaling laws (e.g., Gibbs–Thomson scaling of nanoparticle melting $f_c$8) (Bideault et al., 2024).

Extreme Conditions and Extrapolation

qSNAP provides robust performance outside the training regime (e.g., up to 5 TPa and 20,000 K in carbon), with phase boundaries and Hugoniot curves accurate to within 3% versus quantum molecular dynamics and experiment (Willman et al., 2022). However, as the number of species or descriptors increases, numerical stability and the risk of overfitting become significant, particularly for qSNAP applied to large, chemically diverse datasets (>3000 structures) (Choyal et al., 2023).

5. Computational Efficiency and Implementation

The computational cost of SNAP scales as $f_c$9 per MD step due to the bispectrum calculation and analytic force derivation. The transition from linear SNAP to qSNAP (for fixed $r_{ij} > r_{cut}$0) increases cost only by a factor of ≲2, as the quadratic form is efficiently computed without new neighbor loops (Wood et al., 2017). For practical MD (e.g., 10$r_{ij} > r_{cut}$1–10$r_{ij} > r_{cut}$2 atoms, 100 ns), qSNAP achieves throughput comparable to optimized EAM within LAMMPS, benefitting from vectorization, mixed-precision kernels, and efficient MPI/OpenMP scaling (Bideault et al., 2024, Willman et al., 2022). Memory footprint and runtime scale with the descriptor count; for $r_{ij} > r_{cut}$3, typical $r_{ij} > r_{cut}$4 is 50–80.

6. Applications and Extensions

SNAP/qSNAP have enabled quantum-accurate, long-time, large-scale simulations for:

• Elemental and multi-component alloys (Ni–Mo, W–Be, W–ZrC, Li–TMO$r_{ij} > r_{cut}$5) (Li et al., 2018, Cusentino et al., 2020, Sikorski et al., 2022, Choyal et al., 2023).
• Nanoparticle melting, heat capacity, structural transitions, size-dependent stability (Bideault et al., 2024).
• Phase diagrams at extreme conditions (carbon under GPa–TPa pressures) (Willman et al., 2022).
• Radiation damage and defect evolution in metals (Nb, W) (Bhardwaj et al., 5 Feb 2025, Wood et al., 2017).
• High-throughput materials screening (ternary convex hulls, multicomponent alloys) (Minotakis et al., 2023, Rossignol et al., 2023).

Systematic extensions include explicit multi-element descriptors (Cusentino et al., 2020), spin- and vector-field coupling for magnetic materials (Domina et al., 2022), and neural network or Gaussian-process regression atop the SNAP bispectrum for further flexibility (but at increased training and computational cost) (Zagaceta et al., 2020).

7. Limitations and Practical Considerations

• qSNAP parameter inflation: Quadratic coupling increases the number of parameters by $r_{ij} > r_{cut}$6, making the approach data-hungry and susceptible to overfitting unless large, appropriately diverse training sets are provided (Wood et al., 2017, Choyal et al., 2023).
• Numerical stability: qSNAP can suffer from ill-conditioning and convergence failures with very large descriptor sets or insufficient regularization; cross-validation and data scaling are essential safeguards (Choyal et al., 2023).
• Computational cost: While much faster than quantum methods, SNAP is typically 10–100× slower than simple EAM/MEAM, particularly at high expansion order (Chen et al., 2017, Wood et al., 2017).
• Transferability: Linear SNAP is often preferred in high species-count, disordered, or high-entropy systems due to its stability and modest data requirements (Choyal et al., 2023). qSNAP is recommended where sub-meV/atom accuracy is critical and sufficient training data are available.

---

References:

• (Bideault et al., 2024) Polyvalent Machine-Learned Potential for Cobalt: from Bulk to Nanoparticles
• (Wood et al., 2017) Extending the Accuracy of the SNAP Interatomic Potential Form
• (Choyal et al., 2023) Constructing and evaluating machine-learned interatomic potentials for Li-based disordered rocksalts
• (Willman et al., 2022) Machine Learning Interatomic Potential for Simulations of Carbon at Extreme Conditions
• (Cusentino et al., 2020) Explicit Multi-element Extension of the Spectral Neighbor Analysis Potential for Chemically Complex Systems
• (Bhardwaj et al., 5 Feb 2025) A Robust Machine Learned Interatomic Potential for Nb: Collision Cascade Simulations with accurate Defect Configurations
• (Minotakis et al., 2023) Machine-Learning Surrogate Model for Accelerating the Search of Stable Ternary Alloys
• (Sikorski et al., 2022) Machine Learned Interatomic Potential for Dispersion Strengthened Plasma Facing Components
• (Li et al., 2018) Quantum-Accurate Spectral Neighbor Analysis Potential Models for Ni-Mo Binary Alloys and FCC Metals