Scalar machine learning of tensorial quantities -- Born effective charges from monopole models

Technical Summary: Scalar Machine Learning of Tensorial Quantities – Born Effective Charges from Monopole Models

Problem Statement
Predicting tensorial properties, such as the Born effective charge (BEC) tensor, using machine learning (ML) typically necessitates carefully designed tensorial descriptors or equivariant architectures to ensure physical consistency under rotations and translations. While equivariant frameworks (e.g., symmetry-restricted neural networks, equivariant graph convolutions) have successfully addressed this by encoding symmetry constraints directly into the model, they introduce significant complexity. A central challenge remains: can simpler, invariant scalar descriptors achieve comparable accuracy in learning tensorial quantities without explicit tensorial equivariance, thereby facilitating easier integration into existing scalar ML frameworks?

Methodology
The authors propose an alternative strategy that learns the BEC tensor using purely scalar (monopole) descriptors by exploiting the physical definition of the BEC as the derivative of polarization with respect to atomic displacements.

1. Theoretical Decomposition:
   The BEC tensor ($Z^*_{j,\alpha\beta}$) is decomposed into a local rigid-ion term (scalar) and a charge-redistribution term (derivative of a scalar). By representing the charge density as a set of localized monopoles ($q_k$) dependent on atomic configurations, the BEC is expressed as:
   $$Z^*_{i,\alpha\beta} = q_i \delta_{\alpha\beta} + \sum_{j \neq i} \frac{dq_j}{dr^\beta_i} (r^\alpha_j - r^\alpha_i) + \sum_k \frac{\partial p^\alpha_k}{\partial r^\beta_i}$$
   where the first term is the rigid-ion contribution, the second is the non-local charge redistribution (monopole derivative), and the third accounts for dipole contributions if included. Crucially, the tensorial nature emerges implicitly through the differentiation of scalar charges with respect to atomic positions, rather than being encoded in the kernel or network architecture itself.

2. Model Implementations:

   • Kernel Methods: The authors employ Smooth Overlap of Atomic Positions (SOAP) descriptors with linear kernel regression. They compare three models: a monopole-only model ($q$), a dipole-only model ($p$) using $\lambda$-SOAP (vector descriptors), and a combined monopole-dipole model ($q+p$).
   • Neural Networks: The monopole formulation is implemented within the MACE (Message Passing Atomic Cluster Expansion) architecture. Here, the scalar output of the graph represents the monopole charge $q_i$, and the full BEC tensor is constructed via automatic differentiation of these charges with respect to atomic positions.

3. Datasets and Training:
   The models were trained and tested on four datasets: liquid water ($H_2O$), liquid NaCl, various phases of methylammonium lead iodide ($MAPbI_3$), and zirconia ($ZrO_2$). The reference BECs were computed using Density Functional Perturbation Theory (DFPT). Charge neutrality was explicitly enforced during training to ensure physical consistency.

Key Contributions

• Scalar Strategy for Tensors: The paper demonstrates that tensorial quantities can be accurately learned using invariant scalar descriptors by leveraging the derivative relationship between polarization and atomic displacement. This avoids the need for explicit equivariant kernels or complex tensorial output layers.
• Monopole-Dipole Decomposition: The work formalizes a framework where the BEC is reconstructed from learned monopole charges and their derivatives, separating local and non-local contributions.
• MACE Integration: The authors show that embedding this scalar monopole approach into a more expressive architecture like MACE yields performance superior to linear kernel methods, even outperforming combined monopole-dipole kernel models in certain regimes.

Results

• Accuracy: For linear regression models, the combined monopole-dipole model generally yields the lowest test set error. The dipole-only model is highly data-efficient but struggles to reduce errors below ~5% of the standard deviation in its linear form. The scalar monopole model approaches similar accuracy to the dipole model but requires larger training sets to converge.
• MACE Performance: The scalar MACE model consistently outperforms the kernel-based approaches across all systems (except water, where the difference is marginal), suggesting that the expressiveness of the message-passing architecture compensates for the lack of explicit tensorial descriptors.
• Spectroscopic Prediction: When applied to compute finite-temperature infrared (IR) spectra via molecular dynamics, all three modeling approaches (monopole, dipole, and combined) produce virtually indistinguishable results that agree well with experimental data for liquid water and $MAPbI_3$. This indicates that instantaneous errors in the scalar monopole model are effectively averaged out during thermodynamic integration.
• Charge Interpretability: The authors note that the fitted monopole charges are not uniquely defined physical quantities. Their values depend sensitively on regularization schemes and hyperparameters. While they reproduce the correct BECs, they do not necessarily correspond to true spatial charge density changes or formal oxidation states.

Significance and Claims
The paper claims that the scalar monopole approach offers a robust, scalable, and easily deployable alternative for predicting tensorial properties in complex materials. Its primary significance lies in its conceptual simplicity and compatibility with existing scalar ML infrastructures, removing the barrier of implementing complex equivariant architectures.

The authors emphasize that while the individual atom-centered monopole charges are model-dependent parameters without direct physical meaning, the framework successfully captures the necessary physics to reproduce correct polarization responses and IR spectra. They conclude that for large-scale simulations where computational efficiency and ease of integration are paramount, the monopole-based formulation is a viable and effective strategy, capable of matching the performance of more elaborate tensorial models when paired with expressive architectures like MACE.