Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications Preprint: APS/123-QED
Naoya Kuroda, Kenji Ishihara, Tomoya Shiota, Wataru Mizukami · 2026
Read Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications Preprint: APS/123-QED by Naoya Kuroda, Kenji Ishihara, Tomoya Shiota, Wataru Mizukami (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03223.txt.
Introduction
Naoya Kuroda, Kenji Ishihara, Tomoya Shiota, and Wataru Mizukami stretch a shared atom model to ninety-seven elements. Such learned models mimic first-rate physics at a sliver of the cost. But heavy elements key to nuclear work, above all minor actinides, were missing from the training data.
The team built a new heavy-element dataset from past lab and computing work. They merged it with known molecule and crystal sets. The open model then scored well on standard tests of crystals and organic matter. It also showed promise on the heavy elements themselves. With actinides at last inside a shared model, teams can screen new nuclear fuels and ceramics far faster than before.
Catalog record
| CREATOR | Naoya Kuroda, Kenji Ishihara, Tomoya Shiota, Wataru Mizukami |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50922 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03223
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications Preprint: APS/123-QED Naoya Kuroda, Kenji Ishihara, Tomoya Shiota, Wataru Mizukami 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. I Introduction 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 of the training pipeline integrating molecular (OFF23) and crystal (MPtrj, HE26) data via the total energy alignment protocol. Minor actinide elements, including transuranium elements such as Am, Cm, and Cf, are not merely nuclear waste but key resources for the sustainability of the nuclear fuel cycle 19 ; 51 ; 61 ; 42 ; 60 , and are emerging as crucial components in advanced applications ranging from americium oxide for space exploration radioisotope thermoelectric generators 3 ; 15 ; 89 to actinide-bearing high-entropy ceramics designed for extreme environments 88 ; 94 ; 50 . Unfortunately, due to their strong radioactivity and toxicity, experimental measurements of important thermophysical properties, such as lattice thermal conductivity, linear thermal expansion, and viscosity, are extremely difficult and expensive, leaving the thermophysical and thermochemical properties of many minor actinide compounds largely unexplored 55 ; 58 ; 14 ; 16 ; 41 ; 83 ; 75 ; 33 ; 92 ; 74 ; 79 ; 95 . Theoretical and computational approaches are therefore essential for characterizing these materials. Computational prediction of thermophysical properties relies on the accuracy of the underlying interatomic potential. First-principles calculations based on density functional theory (DFT) can model interatomic forces accurately, but their computational cost becomes prohibitive for multi-component systems: the reduced crystal symmetry requires large supercells, and the number of atomic displacement configurations needed for higher-order anharmonic force constants grows rapidly 82 . Machine learning interatomic potentials (MLIPs) offer a practical route to overcome this bottleneck 11 ; 6 ; 97 ; 34 ; 69 ; 22 . MLIPs evaluate potential-energy surfaces orders of magnitude faster than DFT while maintaining near-DFT accuracy. System-specific MLIPs, trained on first-principles data for particular systems, have recently predicted the high-temperature thermophysical properties of nuclear materials such as \text{UO}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} and \text{ThO}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} with accuracy approaching that of DFT 77 ; 23 ; 39 ; 2 ; 99 ; 54 . However, because these models are trained on individual systems, they are not suitable for screening across broad compositional spaces, such as actinide-containing mixed oxide fuels 19 ; 51 ; 61 ; 42 ; 60 . Universal MLIPs, also known as atomistic foundation models, are emerging to address this limitation as transferable interatomic potentials that cover a wide range of elements 34 . Architectural developments, such as graph neural networks (GNNs) 68 and E(3) -equivariant message passing 10 ; 56 ; 8 , together with advances in distributed scalability 63 and more efficient equivariant architectures 98 , have driven recent progress. Pretraining with large-scale open datasets 21 ; 5 ; 67 ; 66 ; 53 and heterogeneous training datasets 70 ; 91 ; 9 ; 78 ; 59 ; 37 ; 49 has further improved these models, enabling application to molecules, crystals, and surface systems. However, existing universal MLIPs typically do not cover transuranium elements such as Am, Cm, and Cf 70 ; 59 ; 78 ; 91 ; 7 ; 9 ; 37 ; 49 . This gap comes from the limited or absent coverage of heavy elements in current first-principles datasets 35 ; 21 , which restricts the applicability of pretrained universal models not only to actinide mixed oxides but even to many heavy-element compounds. To expand the elemental coverage of a universal MLIP, we develop a DFT dataset, Heavy Element 2026 (HE26). HE26 spans 65 elements in total and contains eight heavy elements (Am, Cm, Cf, Fr, At, Ra, Po, and Rn) that are rarely present in existing large-scale datasets. The HE26 dataset consists of three subsets: (1) Basic Heavy Element (BHE): elemental solids and binary oxides; (2) Complex Heavy Element (CHE): diverse multicomponent systems; and (3) Compositionally Complex Fluorite Oxides (CCFO): binary-to-quinary fluorite-type mixed oxides of actinides, lanthanides, and rare earth elements, targeting actinide-based high-entropy ceramics and minor-actinide mixed-oxide fuels. We integrate HE26 with the OFF23 43 and MPtrj 7 datasets unified via the total-energy alignment (TEA) protocol 70 . Using this combined dataset, we train MACE-Osaka26, an open-source universal MLIP covering 97 elements. We demonstrate the model’s applicability by predicting lattice parameters and thermal conductivity of actinide fluorite oxides, including mixed oxides containing minor actinides. The remainder of this paper is organized as follows. Section II presents results in three parts: cross-domain accuracy of MACE-Osaka26 across crystalline, molecular, and heavy element datasets; detailed evaluation on the HE26 dataset; and lattice thermal conductivity predictions for actinide fluorite oxides. Section III discusses the origins of the model’s performance and remaining limitations. Section IV summarizes our findings and outlook. Section V describes the dataset construction, MLIP training, and thermal conductivity calculation procedures. II Results We first evaluate the cross-domain generalization capability of the MACE-Osaka26 model across molecules, crystals, and heavy-element compounds spanning 97 elements. Table 1 and Figure 2 compare the prediction accuracies of MACE-Osaka24 and MACE-Osaka26 across crystalline (MPtrj), molecular (OFF23), and heavy-element (HE26) datasets. Appendix B provides detailed descriptions of the prediction accuracies for energies on the test sets, and for forces and stresses on both the training and test sets. As shown in Table 1 , the prediction accuracies for the training and test sets exhibit generally similar trends across the MPtrj and OFF23 datasets. Furthermore, due to the limited total number of structures in the HE26 dataset, a separate test set was not partitioned. For these reasons, the parity plots illustrating the energy prediction accuracies in Figure 2 are presented using the training sets for all three datasets. For the crystalline MPtrj dataset, MACE-Osaka26 shows consistent improvements in energy prediction on the test set, with the mean absolute error (MAE) decreasing from 33.7 to 29.3 meV/atom and the root-mean-square error (RMSE) from 77.2 to 62.0 meV/atom. While the force MAE slightly increased, the force RMSE and stress RMSE showed meaningful reductions, indicating better handling of outlier configurations in bulk systems. The most striking improvement is observed in the molecular OFF23 test set. MACE-Osaka26 significantly enhanced energy accuracy compared