Multi-Fidelity ML Force Field Framework

• The paper introduces a multi-fidelity ML force field framework that integrates diverse data sources via fidelity-aware neural network architectures.
• It employs a GNN backbone with dedicated output heads and a composite loss function to harmonize low-cost and high-accuracy data.
• The approach achieves significant reductions in computational expense and prediction errors for energies and forces in complex materials.

A multi-fidelity machine learning force field (MF-MLFF) framework is a data-driven approach that unifies information from multiple sources of varying accuracy—typically distinguished by computational cost, theoretical rigor, or experimental provenance—to construct interatomic potentials. This paradigm aims to achieve chemical accuracy with minimal reliance on prohibitively expensive reference data by simultaneously leveraging abundant low-fidelity and scarce high-fidelity data streams within a unified neural network architecture. MF-MLFF models are especially impactful where high-fidelity data, such as spin-polarized density functional theory (DFT) or coupled cluster calculations, are difficult to generate at scale, as is common in complex materials systems including lithium-ion cathode materials, mixed-metal oxides, and molecular condensed phases (Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024, Gardner et al., 17 Jun 2025, Bukharin et al., 2023, Röcken et al., 2023, Thürlemann et al., 2023).

1. Motivation and Theoretical Underpinnings

The core motivation for multi-fidelity machine-learning force fields arises from the trade-off between cost and accuracy in generating reference data. High-fidelity quantum simulations (e.g., spin-polarized DFT, CCSD(T)) are essential for quantitative predictions but are computationally expensive and sometimes hard to converge for systems with correlated or magnetic electrons. Conversely, low-fidelity methods (e.g., non-magnetic DFT, empirical potentials, or GGA-level DFT) yield fast, stable results but introduce systematic bias in total energies, forces, and derived properties.

MF-MLFF frameworks formalize the simultaneous use of these data sources by extending standard equivariant graph neural network (GNN) architectures to be explicitly fidelity-aware. Through shared representations and fidelity encodings, the model learns transferable geometric and chemical features from low-fidelity data while calibrating them to match high-fidelity reference points in regions of configurational or compositional interest. This approach exploits the greater configurational diversity available cheaply, substantially boosting sample efficiency for high-accuracy learning (Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024).

2. Mathematical Framework and Architecture

A generic MF-MLFF model comprises:

• GNN Backbone with Fidelity Encodings: Per-atom feature vectors $\mathbf v_i^0$ are constructed by concatenating element embeddings $\mathbf X_{Z_i}$ and fidelity embeddings $\mathbf X_f$. At each message-passing or MLP block, one-hot fidelity vectors are injected along with atomic features, i.e.,

$$\mathbf a^{(k)} = [\mathbf W^{(k,k-1)},\,\mathbf W^{(k,k-1)}_F] \begin{pmatrix} \mathbf a^{(k-1)}\\mathbf f_g \end{pmatrix} + \mathbf b^{(k)},$$

allowing the network to learn both fidelity-invariant and incrementally fidelity-specific representations (Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024).

• Output Heads and Composition Model: The total predicted energy for structure $\mathbf r$ of fidelity $f$ is:

$$E_{\mathrm{pred}}(\mathbf r;f) = E_c^f + \sum_{i=1}^N L_{\text{readout}}^f(\mathbf v_i^{n_L}),$$

with $E_c^f$ a composition-dependent bias term for each fidelity and $L_{\text{readout}}^f$ a fidelity-specific MLP (Dong et al., 14 Nov 2025). Forces and (where applicable) magnetic moments or stress are obtained by analytic differentiation and additional MLP heads.

• Loss Function: A weighted composite loss pools all fidelity levels and task targets (energy, force, stress, moment), assigning per-fidelity weights $\lambda_f$ and per-task weights $w_*$. The archetypal objective is:

$$\mathcal L_{\text{tot}} = \sum_{f=1}^{n_F} \lambda_f \sum_{s \in \mathcal D_f} \big[ w_E \ell_H(E_{\mathrm{pred}}^s - E_s) + w_F \ell_H(F_{\mathrm{pred}}^s - F_s) + w_\sigma \ell_H(\sigma_{\mathrm{pred}}^s - \sigma_s) + w_m \mathbf{1}_{f=2} \ell_H(m_{\mathrm{pred}}^s - m_s) \big],$$

where $\ell_H$ is the Huber loss (Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024).

• Training Regime: All available training data are batched with mixed fidelities. No alternating schedule is necessary; model updates simultaneously propagate gradients based on both low- and high-fidelity examples. For “universal MLIP” applications, high-fidelity structures may be up-weighted 7:1 to emphasize accuracy in the target regime (Kim et al., 12 Sep 2024).

3. Multi-Fidelity Training Strategies

Several distinct training strategies have emerged:

• Concurrent Joint Training: All fidelities enter the loss in a single training loop with shared and fidelity-specific parameters. This approach yields data efficiency and avoids catastrophic forgetting or overfitting to scarce high-fidelity points (Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024).
• Transfer Learning (Pre-training and Fine-tuning): The model is first pre-trained on massive low-fidelity data, then fine-tuned on high-fidelity examples either by updating the entire backbone or only the output heads. This is effective when low- and high-fidelity datasets overlap structurally but often underperforms the fully joint approach, especially in compositional or geometric regimes not represented in high-fidelity data (Gardner et al., 17 Jun 2025, Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024).
• Multi-Headed Architecture: The backbone GNN is shared, but each fidelity is assigned a separate output head, whose losses are summed according to a mixing parameter. This structure allows method-agnostic latent representations and is advantageous if users wish to deploy the same model as a "foundation potential" supporting multiple label sources or when datasets are non-overlapping (Gardner et al., 17 Jun 2025, Kim et al., 12 Sep 2024).
• Bias-Aware Curriculum (ASTEROID): The bias of low-fidelity data is estimated using a small high-fidelity subset; heavily biased low-fidelity points are downweighted during pre-training. Subsequent fine-tuning on the high-fidelity set corrects residual errors, maximizing stable transfer while minimizing contamination by systematic low-fidelity errors (Bukharin et al., 2023).
• Hybrid Classical/ML Correction: Baseline classical force fields (e.g., multipole electrostatics, D3 dispersion) are refined by localized ML corrections trained against high-level quantum data, with separate corrections for intra- and intermolecular contributions. This modular approach enables rapid extension to new chemical domains (Thürlemann et al., 2023).

4. Data Selection, Fidelity Definitions, and Generalization

A central challenge is defining and encoding different fidelities. Typical hierarchies include:

Fidelity Level	Source	Example Use
High-fidelity	Spin-polarized DFT, meta-GGA, CCSD(T)	Benchmark, calibration
Low-fidelity	Non-magnetic DFT, GGA, empirical FFs	Broad sampling, geometry exploration
Experimental/Other	Lattice constants, elastic moduli (EXP)	Correction, reality alignment

The framework effectively interpolates and, in some cases, extrapolates in both composition and geometric domain. For example, in Li₆PS₅Cl and InₓGa₁₋ₓN, bespoke MF-MLIPs achieved near-benchmark accuracy (MAE ≈ 5.5 meV/f.u.; R²=0.98) in alloy mixing energies even when high-fidelity data were absent for key compositions, outperforming transfer learning and Δ-ML (Kim et al., 12 Sep 2024).

When high-fidelity data covers only a sparse subset of phase space, MF-MLFFs robustly generalize by sharing learned physical features across fidelities, as validated in cathode materials and broad-benchmark crystalline datasets (Dong et al., 14 Nov 2025, Kim et al., 12 Sep 2024). In fusion with experimental data, the same formalism allows correction of DFT errors in mechanical and structural properties without degrading off-target performance (Röcken et al., 2023).

5. Empirical Performance and Data Efficiency Gains

Quantitative benchmarks consistently demonstrate that MF-MLFFs can dramatically reduce the need for expensive high-fidelity calculations. For LiMnFePO₄ cathode materials, up to a 40% reduction in force mean absolute error (MAE) was obtained using mixed-fidelity training, with similarly significant gains in energy and stress prediction. Adding more low-fidelity frames monotonically improved accuracy at fixed high-fidelity budget (Dong et al., 14 Nov 2025). In Li₆PS₅Cl, the multi-fidelity MLIP matched high-fidelity SCAN-MLIP conductivities within 10%, with high-fidelity data requirements cut by a factor of five (Kim et al., 12 Sep 2024).

Hybrid ML/classical models for molecular condensed phases and crystals achieve lattice energy MAEs ∼2.9 kJ/mol for X23 molecular crystals and force MAEs on par with state-of-the-art DFT, all while training exclusively on monomer and dimer ab initio reference data (Thürlemann et al., 2023).

In settings with extreme cost asymmetry (e.g., CCSD(T):DFT at 40:1), bias-aware training and MF frameworks (e.g., ASTEROID) reduced force prediction MAEs by up to 56%—approaching the performance of DFT-trained sGDML, but at a fraction of data cost (Bukharin et al., 2023).

6. Ablation Studies and Comparative Analyses

Systematic ablation studies confirm the necessity of fidelity-aware modules at multiple locations within the architecture (embedding, message passing, readout, composition). Disabling any such component degrades accuracy, especially for properties only labeled at the high-fidelity level (e.g., magnetic moments) (Dong et al., 14 Nov 2025).

Transfer learning and Δ-learning approaches, while simpler, suffer from catastrophic forgetting when high-fidelity data are limited or not structurally representative, and often fail to interpolate across compositional gaps. In contrast, MF-MLFFs maintain performance in inductive domains due to their parameter-sharing and explicit fidelity conditioning (Kim et al., 12 Sep 2024, Gardner et al., 17 Jun 2025).

Multi-headed models enable seamless ingestion of non-overlapping datasets and can serve as extensible "foundation" potentials. However, a modest tradeoff in ultimate high-fidelity accuracy remains when compared to sequential fine-tuning strategies (Gardner et al., 17 Jun 2025).

7. Extensions, Limitations, and Future Directions

The multi-fidelity framework naturally generalizes to more than two fidelity tiers, accommodating distinct functionals, calculation settings, or experimental modalities. Expansion to coupled-cluster data and experimental observables is feasible and has demonstrated success in both molecular and extended systems (Kim et al., 12 Sep 2024, Röcken et al., 2023).

Scalability is enhanced by design choices such as shared GNN backbones, compact per-fidelity subspaces, and bias-aware data weighting. The approach is compatible with differentiable MD engines and advanced GNNs (e.g., CHGNet, DimeNet++, SevenNet/NequIP), making it suited for universal or bespoke force field construction.

Open challenges include optimal mixing of fidelity weights, automated selection of structurally informative low-fidelity conformers, and systematic quantification of uncertainty due to fidelity mismatches. The ongoing development of public multi-fidelity pre-trained potentials for application domains such as battery materials, polyanion cathodes, and transition-metal oxides is a natural direction (Dong et al., 14 Nov 2025).

---

Key References:

• “Toward Multi-Fidelity Machine Learning Force Field for Cathode Materials” (Dong et al., 14 Nov 2025)
• “Data-efficient multi-fidelity training for high-fidelity machine learning interatomic potentials” (Kim et al., 12 Sep 2024)
• “Understanding multi-fidelity training of machine-learned force-fields” (Gardner et al., 17 Jun 2025)
• “Machine Learning Force Fields with Data Cost Aware Training” (Bukharin et al., 2023)
• “Accurate machine learning force fields via experimental and simulation data fusion” (Röcken et al., 2023)
• “Hybrid Classical/Machine-Learning Force Fields for the Accurate Description of Molecular Condensed-Phase Systems” (Thürlemann et al., 2023)