[논문 리뷰] Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications

Naoya Kuroda, Kenji Ishihara외 2명|arXiv (Cornell University)|2026. 03. 03.

Machine Learning in Materials Science인용 수 0

한 줄 요약

저자들은 HE26로 무거운 원소를 포함하는 DFT 데이터셋(8개의 악티나이드 포함)을 구축하고, 97개 원소를 다루는 보편적 MLIP인 MACE-Osaka26를 학습시켜 크로스-도메인 정확도와 악티나이드의 격자 매개변수 및 열전도도를 포함한 특성을 시연합니다.

ABSTRACT

Machine learning interatomic potentials (MLIPs) evaluate potential energy surfaces orders of magnitude faster while maintaining accuracy comparable to first-principles calculations, and universal MLIPs that cover most of the periodic table are becoming increasingly commonplace. However, existing large-scale datasets have limited or no coverage of heavy elements such as minor actinides crucial in the nuclear field, and universal MLIPs are typically limited to 89 elements. Here, we constructed a heavy element dataset HE26 containing minor actinides, based on experimental and computational literature data. By integrating this with existing molecular and crystal datasets, we developed an open-source universal MLIP covering 97 elements, the broadest elemental coverage to date. The resulting model showed strong performance on the inorganic MPtrj and organic OFF23 test sets and promising accuracy on HE26. The dataset and model open a pathway toward the development of energy resources and the design of novel materials, such as actinide-based high-entropy ceramics, in the nuclear field.

연구 동기 및 목표

• Expand elemental coverage of universal MLIPs to heavy elements including minor actinides.
• Integrate new HE26 data with existing MPtrj and OFF23 datasets under a total-energy alignment protocol.
• Train and publicly release a 97-element universal MLIP (MACE-Osaka26).
• Demonstrate model accuracy across crystals, molecules, and heavy-element compounds, and evaluate lattice thermal conductivity in actinide oxides.

제안 방법

• Assemble HE26, a first-principles heavy-element dataset with Basic Heavy Element (BHE), Complex Heavy Element (CHE), and Compositionally Complex Fluorite Oxides (CCFO) subsets.
• Integrate HE26 with MPtrj and OFF23 via total-energy alignment to create a 97-element universal dataset.
• Train MACE-Osaka26 using the MACE framework with a 6.0 Å cutoff and spin-polarized reference energies for heavy elements.
• Apply a ZBL short-range correction and Agnesi transformation for consistency across elements.
• Evaluate cross-domain accuracy on crystalline (MPtrj) and molecular (OFF23) data, and perform detailed HE26 subset analyses (BHE, CHE, CCFO).
• Compute lattice thermal conductivity κL via the phonon Wigner transport equation using third-order IFCs derived from the MLIP.

Figure 1: Overview of the 97-element universal dataset and the MACE-Osaka26 model. (a) Periodic table showing the element distribution (log scale) of the combined dataset consisting of MPtrj, OFF23, and our HE26 dataset. Red boxes indicate the heavy elements newly added in this study. (b) Schematic

실험 결과

연구 질문

• RQ1 How well does a 97-element universal MLIP generalize across crystalline, molecular, and heavy-element chemical spaces?
• RQ2 Can HE26 enable accurate predictions for actinide-containing compounds and fluorite oxides including multi-component systems?
• RQ3 What is the accuracy of lattice parameters and lattice thermal conductivity predictions for actinide oxides and their solid solutions using MACE-Osaka26?

주요 결과

Dataset	Split	# Structures	Structures	Model	Energy MAE (RMSE)	Force MAE (RMSE)	Stress MAE (RMSE)
MPtrj	Train	1502422	MACE-Osaka24	31.4 (74.6)	47.4 (117.0)	1.7 (10.3)
MPtrj	Test	77973	MACE-Osaka24	33.7 (77.2)	61.9 (144.3)	2.2 (45.8)
MPtrj	Train	1502422	MACE-Osaka26	26.8 (60.5)	47.3 (148.3)	1.7 (10.4)
OFF23	Train	951005	MACE-Osaka24	5.7 (26.6)	39.9 (78.1)	N/A
OFF23	Train	951005	MACE-Osaka26	4.9 (11.9)	44.5 (88.0)	N/A
OFF23	Test	50195	MACE-Osaka24	5.8 (25.2)	41.2 (130.7)	N/A
OFF23	Test	50195	MACE-Osaka26	5.0 (12.3)	45.7 (137.1)	N/A
HE26	Train	59438	MACE-Osaka24	N/A a	N/A a	N/A a
HE26	Train	59438	MACE-Osaka26	44.7 (117.3)	26.3 (73.7)	1.3 (7.4)

• MACE-Osaka26 improves crystalline energy prediction (MPtrj) from MAE 31.4 to 26.8 meV/atom and RMSE 74.6 to 60.5 meV/atom.
• On OFF23, energy RMSE drops from 78.1 to 88.0 meV/atom for MACE-Osaka26, with improved force predictions (notably, ~11.9 meV/Å in training set vs ~26.6 for MACE-Osaka24).
• HE26 training shows heavy-element capabilities: energy RMSE 117.3 meV/atom and force RMSE 73.7 meV/Å, with stress RMSE 7.4 meV/Å3, indicating robust learning for Am, Cm, Cf and other heavy elements (training/validation only for HE26).
• Lattice parameters for BHE subset achieve RMSE 0.4488 Å (actinide solids/oxides); CHE subset attains RMSE 0.0477 Å across 64 elements, showing strong generalization to multi-component heavy-element systems.
• CCFO results show actinide series trends (Th–Cf) captured with RMSE 0.0067 Å against DFT trends and RMSE 0.0124 Å for fluorite oxides, with MAE ≈ 0.005 Å across complexity levels (binary to quinary).
• κL predictions for actinide oxides across Th–Cf show agreement with experiments across 300–1500 K, and Am-doped UO2 demonstrates substantial thermal conductivity degradation captured by the model

Figure 2: Accuracy comparison between MACE-Osaka24 and MACE-Osaka26 on cross-domain systems. (a) Evaluation on the MPtrj training set representing crystalline systems, and (b) the OFF23 training set representing molecular systems, using both MACE-Osaka24 (orange) and MACE-Osaka26 (blue) models. (c)