High-Dimensional Neural Network Potentials

• High-Dimensional Neural Network Potentials (HDNNPs) are machine-learned models that decompose the total energy into atomic contributions, achieving near quantum-mechanical accuracy.
• They employ atom-centered symmetry functions and per-element neural networks to capture local chemical environments and many-body interactions effectively.
• HDNNPs enable scalable simulations for metals, semiconductors, and magnetic systems, validated against high-level quantum methods with low mean errors.

A high-dimensional neural network potential (HDNNP) is a machine-learned, atomistic potential-energy model that combines the first-principles accuracy of electronic structure theories with the computational efficiency of classical interatomic potentials. Following the Behler–Parrinello formalism, HDNNPs decompose the total potential energy of a system into atomic contributions, each predicted by an atomic neural network operating on local symmetry function descriptors. This approach enables highly transferable, accurate, and scalable force fields for multielement chemical and materials systems, including metals, semiconductors, transition-metal oxides, and magnetic materials.

1. Fundamental Principles and Energy Decomposition

HDNNPs represent the total energy of an $N$-atom system as a sum of local atomic energy contributions:

$E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$

where $E_i$ is the output of an atomic neural network that depends on a descriptor vector $\mathbf{G}_i$ encoding the geometric environment of atom $i$ within a finite cutoff radius. This additive Ansatz ensures extensivity and computational scalability while capturing intricate many-body effects (Schran et al., 2021, Weinreich et al., 2020, Omranpour et al., 2024, Eckhoff et al., 2021, Ko et al., 2020).

For more advanced (third- and fourth-generation) HDNNPs, $E_i$ may also depend on global variables such as atomic partial charges or spins, obtained via auxiliary neural networks and charge/spin equilibration schemes (Ko et al., 2020, Kocer et al., 11 Feb 2025, Eckhoff et al., 2021).

2. Symmetry Function Descriptors

Atomic environments are mapped to high-dimensional, fixed-length feature vectors using atom-centered symmetry functions (ACSFs):

• Radial functions probe the distance distribution to neighbors:

$$G_i^2 = \sum_{j\neq i} \exp[-\eta (R_{ij} - R_s)^2]\, f_c(R_{ij})$$

where $\eta$ and $R_s$ control the width and center, and $f_c$ is a smooth cutoff.

• Angular functions encode three-body correlations and angular selectivity:

$E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$0

Typical ACSF parameterizations employ dozens of distinct functions per element/species to efficiently span relevant regions of chemical space, ensuring invariance under permutations, translations, and rotations (Schran et al., 2021, Weinreich et al., 2020, Omranpour et al., 30 Aug 2025, Omranpour et al., 2024, Minamitani et al., 2019).

Extensions include spin-dependent ACSFs (sACSFs) for collinear magnets, where additional spin coordinates enable atomic neural networks to distinguish ferro- and antiferromagnetic orders (Eckhoff et al., 2021).

3. Neural Network Construction and Training

Separate feed-forward neural networks are constructed for each chemical element (and, if needed, for each spin type (Eckhoff et al., 2021)), with shared weights across all atoms of the same species. Typical architectures (per atom) include:

Element	# Input Neurons	Hidden Layers	Output
H, O, Cu	30–100 (ACSFs)	2–3 layers, 15–500 neurons/layer (tanh activ.)	Scalar $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$1

Networks are trained to reproduce reference quantum-mechanical energies and forces (and, for higher-generation models, partial charges) using extended Kalman filter or Adam optimizers (Weinreich et al., 2020, Omranpour et al., 2024, Minamitani et al., 2019). The training loss is typically a weighted sum of energy and force mean squared errors, with explicit monitoring against independent test sets to prevent overfitting (Schran et al., 2021, Omranpour et al., 30 Aug 2025).

For 4G-HDNNPs, electronegativity neural networks are included to predict environment-dependent $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$2, which enter a global charge equilibration (QEq) problem producing system-wide atomic charges. These charges augment atomic input features in the energy network, enabling correct treatment of long-range charge transfer, protonation states, and non-local electrostatics (Ko et al., 2020, Kocer et al., 11 Feb 2025).

4. Methodological Developments: Generations and Extensions

HDNNPs have advanced through distinct methodological "generations":

1. 2G-HDNNPs: Purely local models with sum-over-atom energies based only on ACSFs (Schran et al., 2021, Weinreich et al., 2020, Minamitani et al., 2019).
2. 3G-HDNNPs: Incorporate local, environment-dependent charges into energy prediction, with fixed-range Coulombic interactions.
3. 4G-HDNNPs: Couple neural-network-predicted electronegativities with global charge equilibration; the total energy includes both local atomic energies (functions of ACSFs and $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$3) and an explicit, system-wide Coulombic term. These models accurately describe redox events, protonation, and long-range electron transfer (Ko et al., 2020, Kocer et al., 11 Feb 2025).
4. mHDNNPs with sACSFs: Address variable magnetic configurations by embedding canonicalized spin variables into ACSFs, enabling ab initio thermodynamics of magnetic solids (Eckhoff et al., 2021).

Iterative solvers for QEq (iQEq) further lower the scaling of 4G-HDNNP evaluations to $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$4 per MD step and integrate efficiently with production-scale classical MD codes (Kocer et al., 11 Feb 2025).

5. Performance and Validation Benchmarks

HDNNPs achieve high quantitative accuracy compared to both DFT/Hartree–Fock and high-level wavefunction methods:

• Mean absolute/Root mean square errors:
  • He–H$E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$5O$E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$6: MAD(train) = $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$7 kJ/mol; MAD(test) = $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$8 kJ/mol vs. CCSD(T) (Schran et al., 2021).
  • $E_{\text{tot}} = \sum_{i=1}^N E_i(\mathbf{G}_i)$9-Brass nanoparticles: Energy RMSE = $E_i$0 meV/atom; Force RMSE = $E_i$1 meV/Å (test) (Weinreich et al., 2020).
  • Co$E_i$2O$E_i$3 bulk: Energy RMSE = $E_i$4 meV/atom; Force RMSE = $E_i$5 eV/Å (Omranpour et al., 2024).
  • Magnetic MnO: Energy RMSE = $E_i$6 meV/atom (test) (Eckhoff et al., 2021).
  • Si/GaN: Force RMSE $E_i$7 meV/Å (Si), $E_i$8 meV/Å (GaN); energy accuracy $E_i$9 error in $\mathbf{G}_i$0 over 200–1000 K (Minamitani et al., 2019).

Validation is typically performed using energy/force parity plots, phonon densities of states, comparison to experimental lattice constants, defect/diffusion barriers, and, where applicable, magnetic exchange parameters and ordering temperatures.

6. Applications Across Materials and Chemistry

HDNNPs have been successfully applied across a broad range of condensed-matter and chemical systems:

• Transition-metal oxides (Co$\mathbf{G}_i$1O$\mathbf{G}_i$2, Li$\mathbf{G}_i$3Mn$\mathbf{G}_i$4O$\mathbf{G}_i$5, MnO) for catalysis, battery materials, spin glasses (Eckhoff et al., 2020, Omranpour et al., 2024, Omranpour et al., 30 Aug 2025, Eckhoff et al., 2021).
• Interface and solvation: Water at catalytic surfaces, protonated cations in superfluid helium, capturing quantum solvation features (Schran et al., 2021, Omranpour et al., 30 Aug 2025).
• Metals and alloys: $\mathbf{G}_i$6-brass nanoparticles (Cu-Zn), including alloy energetics, melting, and surface phenomena at up to $\mathbf{G}_i$7 atoms (Weinreich et al., 2020).
• Semiconductors: Si and GaN, predicting phonons and lattice thermal conductivity at DFT accuracy with orders-of-magnitude computational speedup (Minamitani et al., 2019).
• Global charge transfer: Protonation/deprotonation, charge-induced structural transitions, supported metal nanoclusters, redox (Ko et al., 2020, Kocer et al., 11 Feb 2025).

Extended timescale MD and MC simulations of nanosecond scale (and up to microseconds in some studies), and nanometer to micron spatial scales, become feasible while maintaining near-DFT accuracy.

7. Limitations and Future Directions

Current HDNNP frameworks are limited by the locality of descriptors (for 2G/3G) and by the computational cost of QEq (for 4G), as well as the non-uniqueness of partial charges and restrictions to collinear spin order:

• Descriptor limitations: Conventional ACSFs are blind to long-range electronic structure and noncollinear spin effects. Fourth-generation models address electrostatic nonlocality but not higher multipolar or explicit polarizability effects (Ko et al., 2020, Kocer et al., 11 Feb 2025).
• Scaling and implementation: Direct QEq solvers scale as $\mathbf{G}_i$8; iterative (iQEq) approaches achieve $\mathbf{G}_i$9 (Kocer et al., 11 Feb 2025).
• Transferability: Each HDNNP is only as accurate as its reference DFT/higher-level data and the configurational, stoichiometric, or charge/spin space covered by the training set. Extrapolation is nontrivial and may require active-learning/adaptive retraining (Omranpour et al., 2024, Eckhoff et al., 2020).
• Spin treatment: Present models handle only collinear magnetic order with discrete spin variables; noncollinear extensions and orbital degrees of freedom remain an open field (Eckhoff et al., 2021).

Planned extensions include explicit noncollinear and continuous spin descriptors, integration with $i$0-learning for post-DFT corrections, polarizability-aware descriptors, and hybrid quantum–machine-learning force fields for electronic excited states and quantum nuclear motion.

---

References:

• "High-dimensional neural network potentials for solvation: The case of protonated water clusters in helium" (Schran et al., 2021)
• "Properties of $i$1-Brass Nanoparticles I: Neural Network Potential Energy Surface" (Weinreich et al., 2020)
• "A High-Dimensional Neural Network Potential for Co$i$2O$i$3" (Omranpour et al., 2024)
• "High-Dimensional Neural Network Potentials for Magnetic Systems Using Spin-Dependent Atom-Centered Symmetry Functions" (Eckhoff et al., 2021)
• "A Fourth-Generation High-Dimensional Neural Network Potential with Accurate Electrostatics Including Non-local Charge Transfer" (Ko et al., 2020)
• "Iterative charge equilibration for fourth-generation high-dimensional neural network potentials" (Kocer et al., 11 Feb 2025)
• "Simulating lattice thermal conductivity in semiconducting materials using high-dimensional neural network potential" (Minamitani et al., 2019)
• "Insights into the Structure and Dynamics of Water at Co$i$4O$i$5(001) Using a High-Dimensional Neural Network Potential" (Omranpour et al., 30 Aug 2025)
• "Closing the gap between theory and experiment for lithium manganese oxide spinels using a high-dimensional neural network potential" (Eckhoff et al., 2020)