Explainable Attention CNN for Predicting Properties of Heusler Alloys
Source abstract
Abstract
Machine learning implemented with multiple architectures exhibits a superior predictive capability. The spatial attention mechanism efficiently extracts position-related information and inherently aligns with the physical rules governing such properties. We applied the spatial attention mechanism to investigate material properties and developed a deep learning model termed Attention Convolutional Neural Networks (CNN). This model demonstrated remarkable performance for Heusler alloys, achieving R 2 values of 0.9988 for ground state energy ( E 0 ) and 0.9690 for the energy above the convex hull ( E hull ), alongside a 27.9% reduction in mean square error compared to standard CNN. Attention CNN also offers significant interpretability. The embeddings for magnetic moment ( m s ) elucidate atomic characteristics, reflecting the electron configuration and positions within the periodic table. Analysis of the elemental masked attention matrix reveals the contributions from individual elements and the relationships between elements. The single-element contribution analysis concurs with findings from embeddings, while the analysis of two-element contributions is consistent with the Slater–Pauling rule. Moreover, using Attention CNN, we discover a few compounds (LiClMnPt, TiCoMnIr, TiCoFeIr, and TiRhMnIr) with high saturated magnetic moment ranking in potentially metastable Heulser alloys.
Source: openalex_inverted_index · Quality: baseline
Research context
OpenAlex topics: Welding Techniques and Residual Stresses · Machine Learning in Materials Science · Non-Destructive Testing Techniques
SciVX discovery tags are terminology-based. OpenAlex topics are source-provided classifications.
Affiliations
State Key Laboratory of Low-Dimensional Quantum Physics, Department of Physics
Tsinghua University
Frontier Science Center for Quantum Information
Funding
Major Research Plan · Tsinghua University
References (49)
Show source-provided references
1. 10.1038/s41524-022-00723-910.1038/s41524-022-00723-9 ↗
2. 10.1038/s41524-024-01358-810.1038/s41524-024-01358-8 ↗
3. 10.1038/s41524-019-0249-110.1038/s41524-019-0249-1 ↗
4. 10.1038/s41524-023-01084-710.1038/s41524-023-01084-7 ↗
5. 10.1038/s42256-023-00716-310.1038/s42256-023-00716-3 ↗
6. 10.1038/s41524-023-01130-410.1038/s41524-023-01130-4 ↗
7. 10.1038/s41467-023-43720-210.1038/s41467-023-43720-2 ↗
8. 10.1038/s41467-022-29939-510.1038/s41467-022-29939-5 ↗
9. 10.1038/s43588-022-00349-310.1038/s43588-022-00349-3 ↗
10. 10.1038/s41524-024-01343-110.1038/s41524-024-01343-1 ↗
11. 10.1021/acs.jpca.3c0386010.1021/acs.jpca.3c03860 ↗
12. 10.1038/s41524-023-01188-010.1038/s41524-023-01188-0 ↗
13. 10.1016/j.commatsci.2021.11096710.1016/j.commatsci.2021.110967 ↗
14. 10.1038/s41524-024-01433-010.1038/s41524-024-01433-0 ↗
15. 10.1021/acs.jpclett.4c0317910.1021/acs.jpclett.4c03179 ↗
16. 10.1063/1.494002610.1063/1.4940026 ↗
17. Batatia I. · 2022 · 11423
18. 10.1038/s41524-025-01623-410.1038/s41524-025-01623-4 ↗
19. 10.1038/s41524-024-01332-410.1038/s41524-024-01332-4 ↗
20. 10.1038/natrevmats.2016.3210.1038/natrevmats.2016.32 ↗
21. 10.1088/1361-6463/aaaa5810.1088/1361-6463/aaaa58 ↗
22. 10.1088/1361-648X/ac268c10.1088/1361-648X/ac268c ↗
23. 10.1016/j.progsolidstchem.2011.02.00110.1016/j.progsolidstchem.2011.02.001 ↗
24. 10.1080/14686996.2020.181236410.1080/14686996.2020.1812364 ↗
25. 10.1016/j.mtphys.2024.10134310.1016/j.mtphys.2024.101343 ↗
26. 10.1038/s41578-018-0036-510.1038/s41578-018-0036-5 ↗
27. 10.1103/PhysRevMaterials.6.L09140210.1103/PhysRevMaterials.6.L091402 ↗
28. 10.1103/PhysRevMaterials.2.12380110.1103/PhysRevMaterials.2.123801 ↗
29. 10.1021/acs.chemmater.6b0272410.1021/acs.chemmater.6b02724 ↗
30. 10.1021/acs.chemmater.3c0132310.1021/acs.chemmater.3c01323 ↗
31. 10.1039/C8SC02648C10.1039/C8SC02648C ↗
32. 10.1016/j.matchemphys.2023.12803010.1016/j.matchemphys.2023.128030 ↗
33. 10.1103/PhysRevX.4.01101910.1103/PhysRevX.4.011019 ↗
34. 10.1016/j.jmmm.2022.16981810.1016/j.jmmm.2022.169818 ↗
35. 10.1063/5.023230810.1063/5.0232308 ↗
36. 10.1007/978-3-030-01234-2_110.1007/978-3-030-01234-2_1 ↗
37. 10.1038/s41524-021-00545-110.1038/s41524-021-00545-1 ↗
38. 10.1103/PhysRevB.109.08530610.1103/PhysRevB.109.085306 ↗
39. 10.1016/j.commatsci.2012.02.00210.1016/j.commatsci.2012.02.002 ↗
40. 10.1007/s11837-013-0755-410.1007/s11837-013-0755-4 ↗
41. 10.1038/s41524-019-0223-y10.1038/s41524-019-0223-y ↗
42. He K. · 2016 · 77010.1109/CVPR.2016.90 ↗
43. 10.1038/s41598-021-92030-410.1038/s41598-021-92030-4 ↗
44. 10.1088/1361-648X/ab6e9610.1088/1361-648X/ab6e96 ↗
45. 10.1016/j.commatsci.2022.11183610.1016/j.commatsci.2022.111836 ↗
46. 10.1103/PhysRev.49.93110.1103/PhysRev.49.931 ↗
47. 10.1103/PhysRev.54.89910.1103/PhysRev.54.899 ↗
48. 10.1126/sciadv.160224110.1126/sciadv.1602241 ↗
49. 10.1021/acs.chemmater.7b0523010.1021/acs.chemmater.7b05230 ↗