CHGNet Foundation Potential

• CHGNet Foundation Potential is a charge-aware graph neural network model that encodes both atomic and electronic degrees of freedom to achieve near-DFT accuracy at reduced computational cost.
• It leverages radial Bessel and Fourier-based angular expansions with multiple message-passing layers, and incorporates magnetic moment prediction to capture complex charge effects.
• The model is efficiently fine-tuned for specific materials, integrating seamlessly into NEB/MD simulation workflows to accelerate the discovery of fast-ion conductors and novel functional materials.

The CHGNet Foundation Potential is a universal, charge-aware graph neural network (GNN) machine learning interatomic potential designed to encode both atomic and electronic degrees of freedom within a single framework. Its core architecture, data-driven pretraining, and fine-tuning strategies position it as a foundation model for modeling chemically and structurally diverse materials systems with near-density functional theory (DFT) accuracy and vastly reduced computational cost. CHGNet’s integration of magnetic moments as proxies for local orbital occupancies allows it to capture charge effects and phase behavior inaccessible to conventional MLIPs.

1. Architecture and Mathematical Formulation

CHGNet represents each atom as a node in a graph, with edges encoding pairwise distances and angular information. The initial node embeddings are derived from atomic numbers, and edges use radial Bessel expansions of interatomic distances. Multiple message-passing layers update atom and edge features through nonlinear functions and linear transformations. The architecture also processes three-body angular terms using a bond (angle) graph, with angular features expanded in a Fourier basis.

Mathematically, after $T$ message-passing iterations, each atom $i$ has a latent descriptor $h_i = v_i^{(T)} \in \mathbb{R}^{64}$, processed through a feed-forward “energy head” to yield atomic energy contributions $\epsilon_i$. The total energy is $E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$. Forces and stress tensors are obtained by autodifferentiation with respect to atomic positions and strain, respectively.

A unique feature is the explicit prediction of site magnetic moments $m_i = L_m(v_i^{(3)})$, which regularizes the node embeddings to encode orbital occupancy and charge effects. These magnetic moments are not static inputs but are part of the output heads during multitask training, encouraging correlation between predicted atom environments and their electronic configuration (Deng et al., 2023, Christiansen et al., 28 Feb 2025).

2. Foundation Model Pretraining

CHGNet is pretrained on the Materials Project Trajectory (MPtrj) dataset, which contains $\sim$1.5 million inorganic structure snapshots encompassing DFT-computed energies, forces, stresses, and magnetic moments for $\sim$94 chemical elements. The pretraining loss is a weighted mean squared error (or Huber loss) over these four targets: $$\mathcal{L} = \text{MSE}(E^\text{pred}, E^\text{DFT}) + w_F\,\text{MSE}(\mathbf{F}^\text{pred}, \mathbf{F}^\text{DFT}) + w_{\sigma}\,\text{MSE}(\sigma^\text{pred}, \sigma^\text{DFT}) + w_m\,\text{MSE}(m^\text{pred}, m^\text{DFT})$$ with typical weights $w_F = 1$, $i$0, and $i$1. Optimization is performed using Adam or RAdam with initial learning rates near $i$2 (Deng et al., 2023, Lian et al., 3 Jul 2025).

On the full MPtrj test set, pretrained CHGNet attains mean absolute errors of 30 meV/atom (energy), 77 meV/Å (forces), 0.348 GPa (stress), and 0.032 $i$3 (magmom) (Deng et al., 2023).

3. Fine-Tuning and Domain Adaptation

The foundation CHGNet can be efficiently fine-tuned for specific materials classes or underrepresented chemical environments. When applied to high-throughput discovery of Li-ion conductors, CHGNet is further refined on $i$42,800 DFT-computed transition-state geometries (generated by running CHGNet-NEB on Li-quaternary systems, filtered for relevant chemistries and symmetries) (Lian et al., 3 Jul 2025).

Fine-tuning retains the GNN architecture and loss structure, using the same MSE-based targets with normalized weights, a RAdam optimizer (learning rate $i$5), and training for 500 epochs. Final held-out test performance is sharply improved relative to the parent model, achieving MAE$i$6 meV/atom, MAE$i$7 meV/Å, and MAE$i$8 mGPa. Critically, the mean absolute error for predicted migration barriers on the NEB-derived transition state dataset drops from 0.24 to 0.07 eV (train) and 0.23 to 0.09 eV (test), with $i$9 increasing from 0.94 to 0.98 (Lian et al., 3 Jul 2025).

Systematic softening (i.e., force underestimation) sometimes observed in pretrained uMLIPs is mitigated by fine-tuning on targeted DFT data. For e.g., in layered WS$h_i = v_i^{(T)} \in \mathbb{R}^{64}$0 and MoS$h_i = v_i^{(T)} \in \mathbb{R}^{64}$1, fine-tuning on $h_i = v_i^{(T)} \in \mathbb{R}^{64}$2100 DFT structures brings force errors down to 100 meV/Å and energy errors to $h_i = v_i^{(T)} \in \mathbb{R}^{64}$31 meV/atom, verifying performance against both DFT and experimental EXAFS spectra (Žguns et al., 10 Sep 2025).

4. Integration with Atomistic Simulation Workflows

CHGNet can be deployed as a plug-in force provider for atomistic simulation packages (e.g., ASE) in both NEB and MD workflows.

NEB workflow (automated):

1. Structure generation: CIF→POSCAR conversion, $h_i = v_i^{(T)} \in \mathbb{R}^{64}$410 Å supercell construction.
2. Enumeration: All $h_i = v_i^{(T)} \in \mathbb{R}^{64}$5 hops among $h_i = v_i^{(T)} \in \mathbb{R}^{64}$6 inequivalent Li Wyckoff sites, both forward and reverse.
3. Path interpolation: Image-dependent pair potential (IDPP).
4. Force evaluation: 7-image CI-NEB with CHGNet calculator, using spring forces and convergence for $h_i = v_i^{(T)} \in \mathbb{R}^{64}$7 eV/Å.
5. Post-processing: Extraction of migration barriers, statistical analysis over all distinct hops.

MD workflow:

• Ensemble: NVT with Nosé–Hoover thermostat, 1 fs (CHGNet) or 2 fs (AIMD) timestep.
• Conductivity: 3$h_i = v_i^{(T)} \in \mathbb{R}^{64}$8200 ps trajectories at each $h_i = v_i^{(T)} \in \mathbb{R}^{64}$9, splitting trajectory for mean square displacement (MSD) analysis, extracting diffusion coefficient $\epsilon_i$0 and Arrhenius activation energy. Room-temperature ionic conductivity is estimated via the Nernst–Einstein relation $\epsilon_i$1 (Lian et al., 3 Jul 2025).

CHGNet’s deployment yields $\epsilon_i$2100$\epsilon_i$3 speedup for NEB (10$\epsilon_i$4 s/path DFT+CI-NEB vs 10$\epsilon_i$5 s CHGNet-NEB) and $\epsilon_i$6200$\epsilon_i$7 greater sampling in MD (1.7 h/1,000 AIMD steps vs 2%%%%37$i$138%%%% CHGNet steps in 3.2 h) (Lian et al., 3 Jul 2025).

5. Benchmarking, Validation, and Performance Comparison

Energy and force metrics: Fine-tuned CHGNet reduces energy MAE from 9 → 2 meV/atom and force MAE from 24 → 13 meV/Å (transition-state set) (Lian et al., 3 Jul 2025).

Migration barrier prediction: For NEB-calculated Li$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$0 migration in Li$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$1MnO$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$2, fine-tuning lowers MAE from 0.21 (pretrained) to 0.056 eV; for LiTi$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$3(PO$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$4)$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$5, fine-tuned CHGNet achieves 0.40 eV vs DFT 0.38 eV (Lian et al., 3 Jul 2025).

Diffusivity and conductivity: In NASICON Li$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$6Al$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$7Ti$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$8(PO$E_{\text{CHGNet}}(R) = \sum_{i=1}^N \epsilon_i$9)$m_i = L_m(v_i^{(3)})$0, CHGNet-MD gives $m_i = L_m(v_i^{(3)})$1 eV ($m_i = L_m(v_i^{(3)})$2 eV) and direct MD $m_i = L_m(v_i^{(3)})$3(300 K)$m_i = L_m(v_i^{(3)})$4 mS/cm (vs 5.1 mS/cm from AIMD), whereas the pretrained model overestimates mobility ($m_i = L_m(v_i^{(3)})$5 eV, $m_i = L_m(v_i^{(3)})$6 mS/cm) (Lian et al., 3 Jul 2025).

Experimental validation: In layered chalcogenides (WS$m_i = L_m(v_i^{(3)})$7, MoS$m_i = L_m(v_i^{(3)})$8), fine-tuning on 50–100 DFT frames delivers EXAFS agreement at the level of experimental uncertainty for Debye–Waller factors and spectral amplitudes, requiring only minor computational resources for retraining (Žguns et al., 10 Sep 2025).

Phonon properties: CHGNet achieves force RMSE (131.76 meV/Å) comparable to state-of-the-art uMLPs but exhibits higher errors for second/third-order IFCs (7.79 eV/Å$m_i = L_m(v_i^{(3)})$9 and 15.98 eV/Å$\sim$0) and large errors predicting lattice thermal conductivity (MAE$\sim$11 dex versus DFT/experiment). This suggests that, in its current form, CHGNet underperforms as a universal force provider for phonon-mediated thermal transport predictions relative to models such as EquiformerV2 (Anam et al., 3 Sep 2025).

6. Discovery and Screening of New Functional Materials

High-throughput CHGNet-accelerated NEB/MD calculations enable enumeration and rapid screening of candidate fast-Li$\sim$2 ion conductors. In a materials discovery campaign, 66 compounds with at least one Li$\sim$3 hop barrier <0.5 eV were identified. Notably, orthorhombic Pnma polymorphs (LiMgPO$\sim$4 with 0.27 eV and LiTiPO$\sim$5 with 0.28 eV barriers), as well as their aliovalent-doped derivatives (Li$\sim$6Mg$\sim$7Al$\sim$8PO$\sim$9, Li$\sim$0TiPO$\sim$1F$\sim$2), display room-temperature ionic conductivities (0.19 mS/cm, 0.024 mS/cm) and formation energies close to the hull ($\sim$330–50 meV/atom) (Lian et al., 3 Jul 2025).

This workflow, combining high-throughput NEB, efficient fine-tuning, and accurate barrier estimation, allows systematic mapping of migration pathways and conductivity metrics, establishing a paradigm for ML-accelerated exploration of novel battery electrolytes.

7. Capabilities, Limitations, and Perspectives

Strengths:

• Broad chemical and structural transferability due to extensive pretraining on $\sim$41.6 million DFT configurations.
• Charge-awareness via magnetic-moment proxy, enabling electronic structure effects and phase complexity to be encoded in the model.
• Rapid, automated integration into NEB/MD pipelines, with speedups enabling previously intractable disorder and kinetic pathway sampling.
• Efficient fine-tuning (<1k DFT structures) corrects softening and restores DFT accuracy for both equilibrium and transition-state configurations (Lian et al., 3 Jul 2025, Žguns et al., 10 Sep 2025).

Limitations:

• Underestimation of high-energy images (10–50 meV/atom) persists, motivating growth of transition-state training data.
• Pretrained and fine-tuned models are specialized to Li-containing materials; extension to, e.g., Na or Mg systems, requires new labeled data (Lian et al., 3 Jul 2025).
• Global charge neutrality and explicit long-range electrostatics are not enforced; charge is inferred via magnetic moments, which are not applicable to nonmagnetic species (Deng et al., 2023). For explicit charge transfer, potentials such as QET, which directly solve for atomic charges, may provide superior accuracy for systems dominated by long-range Coulomb interactions (Ko et al., 10 Nov 2025).
• For phonon-dominated properties—second/third-order IFCs, lattice thermal conductivity—CHGNet currently exhibits systematically larger errors compared to leading uMLPs (Anam et al., 3 Sep 2025).

A practical perspective is that, for equilibrium and activated-process modeling in charge- and spin-active solids, CHGNet now functions as a robust, general-purpose foundation potential. Where domain-specialized accuracy is required in poorly represented chemistries or electronic regimes, fine-tuning or $\sim$5-learning using CHGNet’s internal atomic embeddings via Gaussian process regression remains computationally inexpensive and effective (Christiansen et al., 28 Feb 2025).

---

Key References:

• “High-Throughput NEB for Li-Ion Conductor Discovery via Fine-Tuned CHGNet Potential” (Lian et al., 3 Jul 2025)
• “CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling” (Deng et al., 2023)
• “Benchmarking CHGNet Universal Machine Learning Interatomic Potential Against DFT and EXAFS: Case of Layered WS2 and MoS2” (Žguns et al., 10 Sep 2025)
• “A Comprehensive Assessment and Benchmark Study of Large Atomistic Foundation Models for Phonons” (Anam et al., 3 Sep 2025)
• “$\sim$6-model correction of Foundation Model based on the models own understanding” (Christiansen et al., 28 Feb 2025)
• “A Fast, Accurate, and Reactive Equivariant Foundation Potential” (for comparison of charge-awareness paradigms) (Ko et al., 10 Nov 2025)