Machine-learning potentials for structurally and chemically complex MAB phases: strain hardening and ripplocation-mediated plasticity
Nikola Koutná, Shuyao Lin, Lars Hultman, Davide G. Sangiovanni, Paul H. Mayrhofer · 2025
Read Machine-learning potentials for structurally and chemically complex MAB phases: strain hardening and ripplocation-mediated plasticity by Nikola Koutná, Shuyao Lin, Lars Hultman, Davide G. Sangiovanni, Paul H. Mayrhofer (2025) 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-2503-16018.txt.
Introduction
Nikola Koutná and four co-authors study a tough layered material in 2025, using machine learning models that stand in for costly physics math. The material, called a MAB phase, stacks metal boride sheets with layers of a second element. Such stacks are hard to test by hand across many mixes, heats, and shapes.
The team trained three learning models and checked them on millions of snapshots from first-principles runs. They then pulled virtual samples to find strength limits at room heat and at high heat. Pulling along the layers gave high strength and shear failure, while pulling across them gave a softer response with kinks. One mix held off cracks through bent layers and twin-like shifts, even when hot.
Catalog record
| CREATOR | Nikola Koutná, Shuyao Lin, Lars Hultman, Davide G. Sangiovanni, Paul H. Mayrhofer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-16018
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine-learning potentials for structurally and chemically complex MAB phases: strain hardening and ripplocation-mediated plasticity Nikola Koutná, Shuyao Lin, Lars Hultman, Davide G. Sangiovanni, Paul H. Mayrhofer Abstract Though offering unprecedented pathways to molecular dynamics (MD) simulations of technologically-relevant materials and conditions, machine-learning interatomic potentials (MLIPs) are typically trained for “simple” materials and properties with minor size effects. Our study of MAB phases (MABs)—alternating transition metal boride (MB) and group A element layers—exemplifies that MLIPs for complex materials can be fitted and used in a high-throughput fashion: for predicting structural and mechanical properties across a large chemical/phase/temperature space. Considering group 4–6 transition metal based MABs, with A = Al \text{A}=\text{Al} and the 222, 212, and 314 type phases, three MLIPs are trained and tested, including lattice and elastic constants calculations at temperatures T ∈ { 0,300 , 1200 } T\in\{0,300,1200\} K, extrapolation grade and energy (force, stress) error analysis for ≈ 3 ⋅ 10 6 \approx{3\cdot 10^{6}} ab initio MD snapshots. Subsequently, nanoscale tensile tests serve to quantify upper limits of strength and toughness attainable in single-crystal MABs at 300 K as well as their temperature evolution. In-plane tensile deformation is characterised by relatively high strength, {110} ⟨ 001 ⟩ \langle 001\rangle type slipping, and failure by shear banding. The response to [001] loading is softer, triggers work hardening, and failure by kinking and layer delamination. Furthermore, W 2 AlB 2 able to retard fracture via ripplocations and twinning from 300 up to 1200 K. Keywords : MAB phase; Machine-learning interatomic potentials; Mechanical properties; Molecular dynamics 1 Introduction The swift development in the field of machine-learning interatomic potentials [ 1 ] (MLIPs) opens the door to molecular dynamics simulations (MD) of complex materials under experiment-relevant conditions, reaching near ab initio accuracy at greatly reduced computational costs [ 2 , 3 ] . MLIPs are non-empirical models that learn atomic properties from a reference ab initio dataset, using mathematically convenient and possibly physically-motivated descriptors (e.g., moment tensors [ 4 ] , atomic clusters [ 5 ] ) to encode atomic environments. MLIPs are improvable by active learning exploiting certain extrapolation control (e.g., the MaxVol grade via the D-optimality criterion [ 6 ] ). However, fundamental challenges—such as (i) construction of robust and transferable training sets, and (ii) validation of MLIP-MD simulations beyond ab initio length scales—make efficient MLIP training for structurally and/or chemically complex materials and defects difficult. Especially MLIPs for simulations including wide range of mechanical strains and strain-activated defects are non-trivial, and we attempt making a step forward in this direction. Already MLIP development for elastic constants calculations is far from routine [ 7 , 8 ] . Modelling nucleation and dynamics of defects governing macroscopic mechanical behaviour poses even bigger challenges [ 9 , 10 ] , as extended defects are typically too large to be explicitly labelled by ab initio calculations and included in the training database. Examples of MLIPs tailor-trained for accurate simulations of dislocations [ 11 ] and grain boundaries [ 12 , 13 ] most often concern elemental metals. MLIP-MD mechanistic simulations at the nanoscale have been sparse; including tensile testing and fracture toughness predictions for C-based materials [ 14 , 15 , 9 , 16 ] , BaZrO 3 [ 17 ] , TiB 2 [ 18 ] , as well as nanoindentation simulations for W [ 19 ] , Si, AlN, SiC, and BC 2 N [ 20 ] . Routine predictions of strength and toughness related quantities and mechanisms, however, have been hindered by the lack of MLIPs for structurally complex materials (e.g., multicomponent alloys, superlattices) and/or their unproved transferability (especially in case of universal pre-trained potentials such as M3GNET [ 21 ] , CHGNET [ 22 ] , or MACE [ 23 ] ). MAB phases [ 24 ] (MABs) alternate atomically-thin layers of ceramic-like transition metal boride, M–B, with mono- or bi-layers of an A-element (A = Al, Si, Ga, In,…). The modulated charge density distribution, with predominantly covalent–ionic bonds separated by metallic bonds, renders a combination of properties typical for ceramics (e.g., high-temperature strength and oxidation resistance) as well as for metals (e.g., high fracture toughness and damage tolerance). Thus, MABs are attractive for high-temperature structural applications, as protective coatings, sensors, low friction surfaces, electrical contacts, tunable damping films for microelectromechanical systems, or as a basis of 2D materials for new-generation nanodevices [ 25 , 26 , 27 , 28 ] . With typical formula M n+1 AB 2n ( n=1,2,3,\ldots ), MABs are rich in structures, chemistry, and bonding motifs; including five experimentally known hexagonal or orthorhombic phase prototypes: MAB (222 type, Cmcm [ 29 ] ), M 2 AB 2 (212 type, Cmmm [ 30 ] and P\overline{6}m2 [ 26 ] ), M 3 AB 4 (314 type, Pmmm [ 31 ] ), M 4 AB 6 (416 type, Cmmm [ 32 ] ), and other stable phases predicted by ab initio calculations [ 33 , 34 , 35 ] . The vast chemical and phase space together with numerous competing phases makes systematic screening of this material class difficult and results in relatively few MABs synthesised till today (Ti 2 InB 2 [ 26 ] , Cr 2 AlB 2 [ 36 , 31 , 32 ] , Cr 3 AlB 4 [ 31 , 32 ] , Cr 4 AlB 6 [ 32 ] , MoAlB [ 29 , 37 , 38 , 39 , 40 ] , WAlB [ 29 , 41 ] , Fe 2 AlB 2 [ 42 ] , Mn 2 AlB 2 [ 30 ] ). Mechanical response of MABs has been investigated by 0 K density functional theory (DFT) calculations including phenomenological descriptors of strength and ductility [ 43 , 44 , 45 , 34 ] , toughness and damage tolerance [ 46 , 47 ] . Experimental studies concerned indentation hardness (e.g., MoAlB [ 38 , 39 ] , Mn 2 AlB 2 [ 30 ] ) and fracture toughness measurements (Fe 2 AlB 2 [ 48 ] ), as well as crack deflection and crack healing (MoAlB [ 49 , 50 ] , Fe 2 AlB 2 [ 51 ] ). Currently lacking atomistic simulations could help understanding how mechanical properties and atomic-scale deformation mechanisms of MABs depend on their chemistry and structure. They could also provide hints for a very fundamental question regarding deformation mechanisms; namely, whether MABs can nucleate so-called ripplocations [ 52 ] . Ripplocations are atomic layer ripples claimed to be the universal deformation mechanism in layered materials upon edge-on basal plane compression, observed at nm-to-km length scales (in MAX phases as well as geological formations) [ 53 ] . Despite the laminated character of MABs and their similarity to MAX phases, no ripplocations have been reported, Chen et al. [ 37 ] even suggested that MABs do not ripplocate. In this study, we present a systematic screening of 222, 212, and 314 type MABs containing group 4–6 transition metals, M = (Ti, Ta, W), and A = Al, focusing on structural parameters, mechanical strength, toughness, and fracture mechanisms as a function of temperature. The chosen MABs have been shown dynamically stable [ 34 ] , some are isoelectronic to already synthesised MABs