Fine-Tuned CHGNet Potentials

• Fine-tuned CHGNet potentials are domain-adapted versions of a universal charge-informed graph neural network that enhances DFT-level predictions by incorporating specialized ab initio data.
• They employ systematic fine-tuning strategies using small sets of DFT configurations to correct force softening and stability mispredictions in complex materials.
• These potentials demonstrate significant accuracy improvements in modeling vibrational, structural, and transport properties across varied inorganic compounds such as layered dichalcogenides and fast-ion conductors.

Fine-tuned CHGNet potentials are domain-adapted instances of the Crystal Hamiltonian Graph Network (CHGNet), a universal machine learning interatomic potential (uMLIP) based on a charge-informed graph neural network. Fine-tuning refers to the systematic retraining or augmentation of this pretrained, universal model with a relatively small set of ab initio (typically DFT) data specific to a target chemical system or application. This addresses common out-of-domain errors such as force softening and stability misprediction, and realizes near–ab initio accuracy for both static and dynamic material properties with substantial computational efficiency. The fine-tuned CHGNet workflow has demonstrably advanced the modeling of vibrational, structural, and transport phenomena in complex inorganic compounds, including layered dichalcogenides, fast-ion conductors, and disordered halides.

1. Underlying Architecture and Pretraining of CHGNet

CHGNet is a charge- and spin-informed graph neural network, wherein each atom is represented as a node and interatomic distances define the graph edges. Initial atomic, edge, and (optionally) bond-angle features are embedded via basis expansions (e.g., SmoothRBF, Fourier angular basis) and processed through a sequence of message-passing blocks. The model predicts total energy, forces, stresses, and site magnetic moments, leveraging multi-task learning to encode critical electronic and structural information. Magnetic moments, incorporated as an auxiliary head after the third message-passing block, regularize spin and oxidation state information.

Pretraining is completed using the Materials Project Trajectory Dataset (>1.5 million DFT relaxation trajectories across ≳250,000 inorganic compounds). The default loss comprises squared or Huber terms on energies, atomic forces, stresses, and magnetic moments, with task-specific weighting (default: $w_E=1$, $w_F=1$, $w_S=0.1$, $w_m=0.1$). Optimization exploits the Adam optimizer with batch sizes of 32–40, learning rate scheduling (e.g., CosineAnnealing), and early stopping (Deng et al., 2023).

2. Theory and Implementation of Fine-Tuning

Fine-tuning CHGNet involves minimizing a composite loss function based on DFT labels for energies ($E_n^\mathrm{DFT}$), forces ($\mathbf{F}_{n,i}^\mathrm{DFT}$), and optionally stresses, using the following form:

$$\mathcal{L} = w_E \sum_{n=1}^{N_\mathrm{frames}} (E_n^\mathrm{ML} - E_n^\mathrm{DFT})^2 + w_F \sum_{n=1}^{N_\mathrm{frames}}\sum_{i=1}^{N_\mathrm{atoms}} \|\mathbf{F}_{n,i}^\mathrm{ML} - \mathbf{F}_{n,i}^\mathrm{DFT}\|^2 + w_S \sum_{n=1}^{N_\mathrm{frames}} \|S_{n}^{\mathrm{ML}}-S_{n}^{\mathrm{DFT}}\|^2$$

Weights are typically set as $w_E=1$, $w_F=1$, $w_S=1$ for maximum parity with the original architecture. Fine-tuning strategies vary based on dataset size: for $N\lesssim100$, all but the final message-passing layers or read-out heads are frozen; for larger datasets ($N\gtrsim500$), all layers are retrained (Žguns et al., 10 Sep 2025, Deng et al., 2023).

Common protocols include:

• Dataset Construction: DFT-relaxed structures or short molecular dynamics/relaxation trajectories under various strains or temperatures, typically $\mathcal{O}(100)$ configurations for efficiency–accuracy tradeoff.
• Hyperparameters: Adam optimizer, learning rate $10^{-3}$–$10^{-2}$, batch size 4–32, 5–20 epochs.
• Validation: 8:1:1 or 9:1 train/validation split; held-out test for MAE reporting.
• Computational demands: On the order of $10^3$ s for $\sim1000$ structures on high-performance CPUs or GPUs (Žguns et al., 10 Sep 2025, Böhm et al., 10 Oct 2025).

3. Quantitative Gains from Fine-Tuning

Fine-tuning CHGNet on system-specific DFT data yields rapid and significant improvements in predictive accuracy. Table 1 summarizes the performance for layered 2Hc-WS$_2$ as a prototypical system (Žguns et al., 10 Sep 2025):

Dataset size	Energy MAE (meV/atom)	Force MAE (meV/Å)	Stress MAE (GPa)
Pretrained	– (shifted; slope 0.60)	337	1.0
1 frame	0.9 (slope 0.89)	178	0.7
70 frames	0.6 (slope 0.94)	102	0.15
350 frames	0.5	72	0.08
7000 frames	0.3	37	0.04

Analogous improvements are observed for NEB barriers and dynamical properties in solid electrolytes (Lian et al., 3 Jul 2025, Böhm et al., 10 Oct 2025). A plausible implication is that $\sim$100 DFT-rich frames suffice to converge force and structural metrics to near-DFT limits, and that further expansion reduces DFT error but does not necessarily improve agreement with experiment (e.g., due to implicit DFT functional bias or structural artefacts).

4. Application Domains and Workflow Integration

Fine-tuned CHGNet potentials have been systematically validated in diverse materials classes:

• Thermal disorder and vibrational spectra (layered dichalcogenides): Fine-tuned models reproduce DFT phonons, elastic tensors, and EXAFS spectra to experimental noise floor, with force MAE $\sim$100 meV/Å for $\sim$100 fine-tuning structures (Žguns et al., 10 Sep 2025).
• High-throughput battery materials screening: In NEB/MD workflows for Li-ion conductor discovery, fine-tuned CHGNet models reduce MAE in migration barriers from $\sim$0.24 eV (vanilla) to $\lesssim$0.1 eV (fine-tuned), with speedups of 100–1000$\times$ over pure DFT (Lian et al., 3 Jul 2025).
• Ionic transport in halide solid electrolytes: Iterative (active) fine-tuning on progressively higher-temperature MD snapshots ensures numerical stability up to 800 K and achieves energy/force MAEs of 2.5 meV/atom and 45 meV/Å, reproducing both static structures and dynamical Li$^+$ diffusion properties across full compositional series (Böhm et al., 10 Oct 2025).
• Surface energetics and cluster stability: For Ag(111)–O surface reconstructions, fine-tuned CHGNet restores DFT-grounded energetic orderings when at least $N_{sub}\gtrsim500$ DFT configurations are available (Pitfield et al., 2024).

5. Alternative and Complementary Augmentation: Δ-Model Corrections

For scenarios where fine-tuning is data- or compute-limited, an alternative is the Δ-model approach: constructing an additive correction $\Delta U(\mathcal{S})$ via Gaussian Process Regression (GPR) operating on local descriptors (e.g., atomic SOAP or CHGNet embeddings). This requires only CPU-based training, is highly data-efficient (even $\sim$10–50 DFT points can yield correct energetic orderings), and is robust in low-data regimes (Christiansen et al., 28 Feb 2025, Pitfield et al., 2024). The corrected potential is

$$U_\text{total}(\mathcal{S}) = U_\text{CHGNet}(\mathcal{S}) + \Delta U(\mathcal{S})$$

with practical advantages for active-learning loops and rapid deployment to new chemical regimes.

6. Limits, Best-Practices, and Recommended Protocols

While fine-tuned CHGNet potentials offer broad transferability, their accuracy is ultimately determined by the quality and coverage of the fine-tuning dataset. For new chemistries, the recommended procedure is:

• Generate $\sim$100 DFT-based snapshots sampling thermally accessible or strain-deformed states.
• Fine-tune CHGNet for 5–10 epochs (learning rate $10^{-2}$, batch size 4–32), freezing most layers for small datasets.
• Validate against both DFT and experimental observables (e.g., EXAFS, phonons).
• For extremely sparse data, apply Δ-learning instead.

It is necessary to recognize that excessive fine-tuning on large DFT sets can reduce absolute DFT error but may degrade agreement with experiment if the DFT functional systematically mispredicts lattice parameters or interlayer interactions (Žguns et al., 10 Sep 2025). In high-throughput workflows, the seamless integration of fine-tuned or Δ-corrected CHGNet into NEB, MD, and optimization pipelines has proven computationally effective and technically robust (Lian et al., 3 Jul 2025, Böhm et al., 10 Oct 2025, Pitfield et al., 2024).

7. Significance and Outlook

Fine-tuned CHGNet potentials have established a scalable paradigm for domain adaptation of universal MLIPs, enabling chemically accurate modeling of materials well beyond CHGNet’s original training manifold, with applications ranging from phonon-driven phenomena to ionic conduction and structural prediction. The framework is inherently modular, supporting augmentation by fine-tuning, Δ-model corrections, or active-learning loops. Ongoing research is directed at extending transferability to multivalent systems, explicit long-range electrostatics, and complex grain-boundary or amorphous contexts (Böhm et al., 10 Oct 2025, Lian et al., 3 Jul 2025).

These developments reinforce the central role of fine-tuned CHGNet in computational materials discovery, providing a foundation for next-generation high-throughput screening and predictive modeling workflows at near–ab initio accuracy and computational efficiency.