Tangent-Plane Evidential Uncertainty in Active Learning for Magnetic Interatomic Potentials

arXiv:2605.12353 · physics.comp-ph · Submitted 2026-05-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Tangent-Plane Evidential Uncertainty in Active Learning for Magnetic Interatomic Potentials".

Mira: Here's the gist of the paper. Magnetic interatomic potentials need to account for coupled lattice and spin degrees of freedom, yet constructing reliable training sets remains costly because noncollinear first-principles labels are expensive.

Kai: Here we introduce Equivariant Evidential Deep Learning for Spin-Lattice Potentials (e2SLP), a magnetic uncertainty-learning method that formulates the projected spin-force likelihood and the corresponding epistemic uncertainty in the tangent plane orthogonal to the local spin direction.

Mira: This geometry is consistent with the constrained-moment labels, which contain no independent radial response, and yields a structure-level projected-spin-force uncertainty score for candidate selection.

Kai: Using bulk BiFeO3 and monolayer CrTe2 as benchmark systems, we show that this score correlates strongly with prediction errors and that the full e2SLP selection, training, and fine-tuning workflow lowers energy, force, and projected spin-force errors relative to random sampling.

Paper summary: Kai: So wrapping up the paper "Tangent-Plane Evidential Uncertainty in Active Learning for Magnetic Interatomic Potentials." The authors introduced e2SLP to solve the problem of building magnetic potentials that need to handle coupled lattice and spin degrees of freedom effectively.

Mira: The main point is that by using this tangent plane consistent uncertainty score, they create a selection signal that strongly correlates with prediction errors, meaning we can pick the most informative configurations for labeling.

Lev: It’s about using a single model uncertainty framework to guide an active learning loop, which is more efficient than running multiple models for committee-based criteria.

Kai: The authors used bulk BiFeO3 and monolayer CrTe2 as test cases, showing that the full e2SLP workflow lowers energy, force, and projected spin force errors compared to random sampling.

Mira: And they found that this method works across both bulk and monolayer systems because it provides a consistent geometry for uncertainty in the tangent plane orthogonal to the local spin direction.

Lev: The limitation is that their current tests are only on those two benchmark materials, so future work should explore different magnetic orders or materials with different spin lattice coupling mechanisms.

Conclusion: Kai: So, we've been looking at this paper about "Tangent-Plane Evidential Uncertainty in Active Learning for Magnetic Interatomic Potentials." It tackles how to build reliable magnetic simulations where the lattice structure and the magnetism are both moving around at the same time.

Mira: Right. The authors are using a method called e2SLP. Basically, they take a single model approach to figuring out uncertainty when you're learning about these coupled systems, which is usually really hard because you need separate labels for each thing.

Lev: From an error correction standpoint, that single model idea sounds promising for real hardware. But if the geometry is wrong, it just adds noise instead of helping us find the right answer.

Kai: Exactly. They formulate this uncertainty in a way that matches what we actually care about—the projected spin force in the tangent plane orthogonal to the local spin direction. It's not just any uncertainty; it’s uncertainty tied to the physics of how spin moves relative to the lattice.

Mira: That means they create a score for every potential configuration. This score is supposed to tell you which ones you need more expensive first-principles calculations for, and these scores actually correlate strongly with how wrong the model's predictions are.

Lev: If that correlation holds up when we move beyond two simple materials like Bismuth Iron Oxide and Chromium Telluride, then this could become a real way to build magnetic potentials faster than traditional methods. But the next thing we need to know is if that score still works when things get more complex, like different magnetic structures or stronger spin-lattice coupling.

Yang Cheng, Hongyu Yu, *, Hongjun Xiang‡

Fudan University

physics.comp-ph

Submitted: 2026-05-12

Updated: 2026-07-25

Journal ref: Physical Review Materials 2026

DOI: 10.1103/9hcx-wmmm

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: Magnetic interatomic potentials need to account for coupled lattice and spin degrees of freedom, yet constructing reliable training sets remains costly because noncollinear first-principles labels…

Key concepts

Spin-Lattice Potentials
These are machine learning models designed to accurately simulate materials where both the atomic structure (lattice) and the magnetic spin state are important. They are difficult to build because they require training data covering both structural changes and different magnetic configurations, often needing expensive first-principles calculations.
Projected Spin-Force ($ ext{fs},_ot$)
This is a specific type of force calculated from the energy derivative with respect to the local spin direction. It represents the force component that is perpendicular (projected) to the material's current magnetic spin axis. This quantity is physically relevant for understanding how small changes in spin affect the system.
Tangent-Plane Evidential Uncertainty
This method uses a specialized way to measure uncertainty based on a Gaussian likelihood model projected onto a tangent plane perpendicular to the local spin direction. Instead of measuring uncertainty in all three dimensions, it focuses on the relevant two dimensions where the spin-force prediction is most informative for learning.

Terminology

Summary

Magnetic interatomic potentials need to account for coupled lattice and spin degrees of freedom, yet constructing reliable training sets remains costly because noncollinear first-principles labels are expensive. p. 1 Here we introduce Equivariant Evidential Deep Learning for Spin-Lattice Potentials (e2SLP), a magnetic uncertainty-learning method that formulates the projected spin-force likelihood and the corresponding epistemic uncertainty in the tangent plane orthogonal to the local spin direction. p. 1 This geometry is consistent with the constrained-moment labels, which contain no independent radial response, and yields a structure-level projected-spin-force uncertainty score for candidate selection. p. 1 Using bulk BiFeO3 and monolayer CrTe2 as benchmark systems, we show that this score correlates strongly with prediction errors and that the full e2SLP selection, training, and fine-tuning workflow lowers energy, force, and projected spin-force errors relative to random sampling. p. 1

Methodology Overview

The method combines an uncertainty-guided active-learning workflow with the e2SLP model. Starting from an initial constrained-DFT-labeled dataset, the evidential model evaluates candidate magnetic configurations generated by spin-lattice dynamics and selects configurations with large projected-spin-force uncertainty for additional first-principles labeling. p. 3 p. 3 The e2SLP model uses an SOC-aware magnetic energy backbone and a tangent-plane evidential uncertainty branch. p. 3 p. 3

Uncertainty Formulation

The e2SLP formulation adapts the equivariant evidential construction introduced in e2IP to the magnetic spin-force setting. The predictive mean is identified with the projected energy gradient with respect to the local spin-direction unit vector, γi = −Pi ∂E/∂ˆsi. p. 6 The uncertainty is quantified using a multivariate Student-t negative log-likelihood evaluated in two-dimensional tangent coordinates with d = 2. p. 6

Performance and Validation

The tangent-plane-consistent epistemic uncertainty of the spin-force branch is used as the basis for active-learning acquisition. The structure-level scalar proxy for spin-force epistemic uncertainty is defined as the magnetic-site-averaged trace, U sf epi(C) = 1/Cmag X i∈Cmag tr U sf 2d epi i. p. 7 Across both bulk BiFeO3 and monolayer CrTe2, U sf epi correlates strongly with prediction error and the full e2SLP workflow lowers energy, force, and projected spin-force errors relative to random sampling. p. 3

Ablation Analysis

The smaller four-seed ablation further shows that U sf epi-guided selection lowers the mean errors relative to random selection with the same e2SLP training and fine-tuning route. The other possible source is the model and training route, as e2SLP training followed by fine-tuning improves over random selection with non-evidential training. p. 1

Conclusion

The present tests are still limited to two benchmark systems, one SLD-based candidategeneration protocol, and one structure-level selection score. Further work should test other magnetic orders, materials with different spin–lattice coupling mechanisms, alternative selection rules, and longer downstream spin-lattice simulations. p. 1

--- Page 1 ---

Tangent-Plane Evidential Uncertainty in Active Learning for Magnetic Interatomic Potentials Yang Cheng, Hongyu Yu,∗ and Hongjun Xiang† Key Laboratory of Computational Physical Sciences (Ministry of Education), Institute of Computational Physical Sciences, State Key Laboratory of Surface Physics, and Department of Physics, Fudan University, Shanghai, 200433, China 1 arXiv:2605.12353v2 [physics. p. 0comp-ph] 25 Jul 2026

--- Page 4 ---

I. INTRODUCTION Magnetic materials are central to information storage, spintronics, magneto-optical technologies, and finite-temperature materials modeling[1, 2]. p. 1 In many systems, especially those with strong coupling among lattice, spin, and magnetic anisotropy, reliable simulation requires explicit magnetic degrees of freedom rather than an effective nonmagnetic approximation[3–12]. p. 1 Machine-learning interatomic potentials (MLIPs) extend firstprinciples accuracy to larger length and time scales[13–18], and recent magnetic MLIPs, including SpinGNN[10], mMTP, and mMACE, have begun to incorporate local magnetic moments or spin degrees of freedom directly into the model[7, 11, 19–23]. p. 1 Their construction is particularly demanding because the training data need to cover both structural and spin configurational spaces, while reference labels often require expensive constrained noncollinear density-functional-theory calculations[24–26]. p. 2 This data-construction bottleneck has motivated active-learning strategies for interatomic potentials[27–34]. p. 2 Existing workflows use extrapolation grades, committee or variance estimates, descriptor-space diversity, or hybrid acquisition rules to reduce manual dataset curation and identify informative configurations[27, 28, 31–34]. p. 2 Active learning has also been applied to training-set construction for magnetic interatomic potentials[8]. p. 2 Practical indicators such as extrapolation grades can be effective[28], but they are not formulated as uncertainty models for the projected magnetic-response targets considered here. p. 2 Committeebased criteria require training multiple models and therefore add substantial cost to iterative SOC-aware active-learning loops [29, 30, 35, 36]. p. 2 A closely related development is Equivariant Evidential Deep Learning for Interatomic Potentials (e2IP), a single-model uncertainty framework for vector-valued interatomic-potential targets such as atomic forces[37]. p. 2 In this framework, the model predicts not only the force vector but also a full rotation-consistent 3 × 3 symmetric positive-definite covariance tensor, and uses a Normal–Inverse–Wishart evidential prior to separate epistemic and aleatoric uncertainty [37–40]. p. 2 This makes e2IP attractive for ranking candidate configurations in active learning, because uncertainty can be obtained from a single forward pass rather than from an ensemble. p. 2 However, directly applying this framework to magnetic spin-force learning would be geometrically inconsistent. p. 2 In the present magnetic setting, the learned spin-response target is not a fully unconstrained three-dimensional vector but the projected spin-force target fs,⊥ in the tangent plane orthogonal to the local spin direction[11, 21]. p. 2 Because the constrained-moment labels contain no independent radial response, a three-dimensional uncertainty model would introduce an uncertainty component without a corresponding physical target. p. 2 Using the evidential idea for magnetic interatomic potentials therefore requires a new uncertainty geometry that matches the tangent-space structure of spin-force prediction. p. 2

--- Page 7 ---

II. METHOD Our method combines an uncertainty-guided active-learning workflow with the e2SLP model. p. 3 Starting from an initial constrained-DFT-labeled dataset, the evidential model evaluates candidate magnetic configurations generated by spin-lattice dynamics and selects configurations with large projected-spin-force uncertainty for additional first-principles labeling. The e2SLP model uses an SOC-aware magnetic energy backbone and a tangent-plane evidential uncertainty branch.

--- Page 8 ---

B. Energy-derived forces and projected spin forces Atomic forces are defined as the negative energy gradient with respect to atomic coordinates, whereas spin-force-related quantities are defined from the energy derivative with respect to the local spin-direction unit vector. p. 4 Specifically, the spin coordinate used in the derivative is the dimensionless direction ˆsi of the local magnetic moment, not the magneticmoment magnitude. p. 4 The raw spin-force quantity is therefore fs,i = −∂E/∂ˆsi and has units of energy. p. 4 The physically relevant projected spin force is fs,⊥ = Pfs, P = I − ˆsˆs⊤ where ˆs = s/∥s∥, (3) Denoting the raw spin-force quantity by fs, the physically relevant projected spin force is fs,⊥ = Pfs. p. 4

--- Page 9 ---

C. e2SLP formulation for magnetic spin-force uncertainty To quantify predictive uncertainty, e2SLP adapts the equivariant evidential construction introduced in e2IP to the magnetic spin-force setting. p. 5 Following Wang et al. p. 5[37] and the broader evidential-regression perspective of Amini et al. p. 5[38], each target vector yi is described through a Gaussian likelihood with latent mean and covariance, p(yi µi, Σi) = N (yi µi, Σi), (4) while the pair (µi, Σi) is assigned a Normal–Inverse–Wishart prior, Σi ∼ W−1(Ψi, νi), µ i Σ i ∼ N γ i, 1 κ i Σ i. p. 5 After marginalizing the latent variables, the predictive distribution becomes multivariate Student-t, and the uncertainty decomposes into aleatoric and epistemic parts. p. 5

--- Page 10 ---

D. Tangent-plane-consistent uncertainty geometry and training objective Let ˆs denote the normalized local spin direction. p. 6 The projector P removes the component parallel to the spin axis. p. 6 A three-dimensional covariance or uncertainty tensor Σ is then mapped to the tangent space through Σ⊥ = PΣP. p. 6 To evaluate the likelihood in minimal coordinates, we introduce a local orthonormal tangent basis Ei = [ei,1, ei,2] ∈ R 3×2 (12) with ei,1 and ei,2 orthonormal to ˆsi and to each other. p. 6 The predictive mean γi is not an independent vector head: it is identified with the projected energy gradient with respect to the local spin-direction unit vector, γi = −Pi ∂E/∂ˆsi. p. 6

--- Page 11 ---

E. Structure-level uncertainty score for active learning The tangent-plane-consistent epistemic uncertainty of the spin-force branch is used as the basis for active-learning acquisition. p. 7 At the node level, we work directly with the twodimensional epistemic covariance U sf 2d epi i = νiΣ2d 0i κi(νi − d − 1), (16) and define the structure-level scalar proxy for spin-force epistemic uncertainty for a configuration C as the magnetic-site-averaged trace, U sf epi(C) = 1 Cmag X i∈Cmag tr U sf 2d epi i. p. 7

--- Page 13 ---

A. Correlation between U sf epi and prediction error To assess the usefulness of U sf epi as a configuration-selection signal, we compare it with atomic-force and projected-spin-force prediction errors on held-out test configurations. p. 7 BiFeO3 Fig. p. 8 2 shows that the tangent-plane-consistent uncertainty score is strongly aligned with both force error and projected spin-force error, with Spearman correlation coefficients of 0. p. 8919 and 0.879, respectively. p. 8

--- Page 14 ---

B. Active learning performance on bulk BiFeO3 Bulk BiFeO3 is a room-temperature multiferroic with coupled lattice and G-type antiferromagnetic degrees of freedom[7, 51, 52], providing a benchmark for testing uncertaintyguided selection in a coupled spin–lattice system. p. 9 The e2SLP active-learning strategy consistently improves all reported error metrics relative to random sampling. p. 9

--- Page 17 ---

C. Active learning performance on monolayer CrTe2 Monolayer CrTe2 provides a complementary two-dimensional magnetic benchmark with strong spin–lattice coupling[11, 53, 54]. p. 9 The comparison starts from a shared initial dataset and uses the same temperature-stratified candidate pool at each expansion round. p. 9

--- Page 18 ---

D. Discussion The improvement in accuracy observed here has two possible sources. p. 7 One is configuration selection. p. 7 The correlation tests show that U sf epi ranks configurations by prediction error in bulk BiFeO3 and CrTe2. p. 9 When this score is used for configuration selection, the final errors decrease relative to random sampling. p. 1

--- Page 19 ---

FIG. 4. Monolayer CrTe2: structure and active-learning performance (a) Representative monolayer CrTe2 structure, with Cr atoms (blue), Te atoms (gold), and local spin vectors (red arrows). p. 1

--- Page 10 ---

FIG. 1. e2SLP active-learning framework. p. 8 (a) Uncertainty-guided active-learning workflow: starting from an initial dataset, reference DFT calculations produce a labeled dataset for model training; the trained model is then used together with candidate magnetic configurations generated by spin-lattice dynamics (SLD) to evaluate the structure-level uncertainty score U sf epi, select highuncertainty configurations for additional labeling, and retrain the model on the expanded dataset. p. 8

--- Page 11 ---

FIG. 3. Bulk BiFeO3: spin-lattice structure and magnetic transition (a) Representative BiFeO3 structure extracted directly from the final spin-lattice-dynamics frame and shown as a pseudocubic 2 × 2 × 2 region without symmetry refinement or structural idealization. p. 7

--- Page 14 ---

TABLE I. Performance comparison between random sampling and active learning for the two benchmark magnetic systems, measured by mean absolute error. p. 9 p. 9 p. 9 The lower error in each metric for a given system is highlighted in bold. p. 3

--- Page 15 ---

FIG. 4(b) Monolayer CrTe2: per-atom energy MAE versus active-learning round for random sampling, four-model committee disagreement, and the e2SLP active-learning branch. p. 3 p. 3

--- Page 16 ---

TABLE S1. Round-by-round monolayer CrTe2 active-learning controls. p. 7 p. 7 p. 7 Dataset size denotes the number of training configurations. p. 2 p. 2 p. 2 Energy MAE is reported in meV/atom, force MAE in meV/˚A, and projected spin-force MAE in meV. p. 9

--- Page 17 ---

TABLE S2. CrTe2 selection and training-route ablation on a 500-configuration initial set, evaluated on an independent 500-configuration test set. p. 9 p. 9 The columns denote the number of added configurations. p. 6 p. 6

--- Page 18 ---

TABLE S1. Round-by-round monolayer CrTe2 active-learning controls. Dataset size denotes the number of training configurations.

--- Page 20 ---

TABLE I. Performance comparison between random sampling and active learning for the two benchmark magnetic systems, measured by mean absolute error.

--- Page 19 ---

FIG. 4(b) Monolayer CrTe2: per-atom energy MAE versus active-learning round for random sampling, four-model committee disagreement, and the e2SLP active-learning branch.

--- Page 30 ---

TABLE S1. Round-by-round monolayer CrTe2 active-learning controls. Dataset size denotes the number of training configurations.

--- Page 31 ---

TABLE S2. CrTe2 selection and training-route ablation on a 500-configuration initial set, evaluated on an independent 500-configuration test set. The columns denote the number of added configurations.

--- Page 32 ---

FIG. S1. Bulk BiFeO3: structure-level epistemic uncertainty versus prediction error for the projected-spin-force indicator U sf epi (top row) and the force-branch indicator U f epi (bottom row). p. 9

--- Page 33 ---

FIG. S2. Monolayer CrTe2: structure-level epistemic uncertainty versus prediction error for the projected-spin-force indicator U sf epi (top row) and the force-branch indicator U f epi (bottom row). p. 9

--- Page 34 ---

TABLE I. Performance comparison between random sampling and active learning for the two benchmark magnetic systems, measured by mean absolute error.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Active learning for magnetic interatomic potentials can be improved by using Equivariant Evidential Deep Learning for Spin-Lattice Potentials (e2SLP) to select candidate configurations with large structure-level projected spin-force epistemic uncertainty, which is defined as the magnetic site-averaged trace of the two dimensional epistemic covariance.

  2. The improved system can perform data efficient construction of magnetic machine learning interatomic potentials by prioritizing first principles labeling for configurations with the highest U sf epi score, which correlates strongly with prediction errors.

  3. The system can achieve lower energy, force, and projected spin-force errors relative to random sampling when using the full e2SLP selection, training, and fine-tuning workflow on bulk BiFeO3 or monolayer CrTe2.

  4. The model architecture can be based on a SpinGNN++ inspired backbone combining magnetic symmetry preserving (MSENN) and time reversal equivariant (TENN) channels to construct total energy from Etotal = Estruct + EJ + Eani + Ebiquad + ESOC.

  5. The system can utilize the tangent-plane consistent uncertainty geometry, where the covariance matrix is projected into the local tangent plane Σ0,i,⊥ = PiΣ0,iPi, ensuring that the uncertainty evaluation corresponds only to the transverse response supplied by constrained-moment DFT.

Sources

Related papers