A high-dimensional neural network potential for finite-temperature phenomena in NiTi martensite
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "A high-dimensional neural network potential for finite-temperature phenomena in NiTi martensite".
Mira: Here's the gist of the paper. We present a high-dimensional neural network potential (HDNNP) for the martensitic phase of the NiTi shape-memory alloy trained to density functional theory (DFT) data.
Kai: The HDNNP accurately describes the relative stability of the B19′ and B33 phases, including subtle energy differences on the order of meV/atom.
Mira: The predicted stacking-fault energy landscape is strongly anisotropic and reveals a preferential shear pathway, providing atomistic insight into deformation and twinning mechanisms.
Kai: Finite-temperature molecular dynamics simulations further enable the investigation of unconstrained structural evolution as a function of temperature.
Paper summary: Kai: To recap, this paper presents the high-dimensional neural network potential for finite-temperature phenomena in NiTi martensite. The core thesis is that they built an HDNNP trained on DFT data to accurately model the relative stability between the B19' and B33 phases, capturing subtle energy differences and showing a strongly anisotropic stacking fault energy landscape with a preferential shear pathway.
Mira: They also emphasize their systematic validation against DFT for equilibrium crystal structures, elastic constants, generalized stacking fault energies, and vibrational properties to confirm the model's accuracy. The dataset construction involved iteratively expanding seven thousand three hundred DFT-based structures using an active learning strategy based on force discrepancies from molecular dynamics trajectories.
Lev: The training focused specifically on the B19' phase and its relation to the B33 structure, and they showed that at zero temperature, the HDNNP reproduces the relative stability of these phases as well as the pronounced elastic anisotropy characteristic of B19'.
Kai: Moving into finite temperature dynamics, they used molecular dynamics in NpT ensemble to show a structural evolution from configurations near B33 at lower temperatures toward configurations approaching B19' with increasing temperature. This shift is driven by a shear-like mechanism, suggesting the B33 phase has a temperature-driven instability manifested by the gradual evolution of the monoclinic angle.
Mira: So, for someone just listening to the show, it means this potential gives us a way to theoretically predict how NiTi martensite will change its structure as it heats up, not just what it looks like at absolute zero.
Lev: And from an engineering standpoint, they are showing that their model can be used for upscaling DFT simulations because of this accuracy with finite temperature data. However, they do flag that the predictive power is highly dependent on the quality of the underlying interatomic potential when you're looking at large systems over longer timescales.
Kai: That makes sense; if you try to run a simulation with this potential on a system much bigger than what they trained it on, you might lose some of that DFT accuracy. What about the specific structural details they found regarding deformation?
Mira: They identified a single energetically favorable shear pathway along the one zero zero one M zero zero one M direction in the stacking fault energy surface, which is a direct consequence of their analysis and gives atomistic insight into deformation mechanisms.
Lev: That specific shear pathway along that direction is what they are pointing toward for future studies on slip systems in this material. It's not just about stability; it's about how the atoms actually move when stressed.
Conclusion: Kai: So, looking at the full picture of this paper, "A high-dimensional neural network potential for finite-temperature phenomena in NiTi martensite," they’ve successfully built a machine learning potential that bridges the gap between DFT accuracy and the efficiency needed for molecular dynamics.
Mira: The implications are that we now have a robust tool to perform atomistic simulations of complex structural and functional behavior in NiTi systems, particularly involving hundreds of thousands of atoms over nanosecond time scales. They've shown this potential by reproducing key properties like equilibrium structures, elastic constants, and stacking fault energies with DFT accuracy.
Lev: What they provide is a way to explore the exceptional properties of the B19' structure better, specifically things like that easy one zero one M zero zero one M slip mentioned in their conclusion. This is important because that slip system is unusual for intermetallic martensitic systems.
Kai: And they’ve shown how temperature drives structural evolution in this material through a shear-like mechanism, predicting the evolution of lattice vectors and atomic positions as the temperature changes, not just static snapshots.
Mira: The paper provides a clear path forward for atomistic modeling where you need to combine high accuracy with the ability to handle large systems and longer time scales, all while being mindful of how much you are extrapolating beyond the training set.
Lev: Ultimately, this HDNNP potential is presented at three seven in a form compatible with both LAMMPS and n2p2 packages, making it ready for implementation in real hardware simulations. It's a solid first step for studying these complex materials further.
Department of Solid State Engineering, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague · Institute of Thermomechanics, Czech Academy of Sciences · Institute of Physical Engineering, Brno University of Technology · CEITEC BUT, Brno University of Technology · Lehrstuhl für Theoretische Chemie II, Ruhr-Universität Bochum · Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr
cond-mat.mtrl-sci
Submitted: 2026-07-22
Updated: 2026-11-10
Comments: Manuscript accepted to Materials \& Design
DOI: 10.1016/j.matdes.2026.117424
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 86/100
The gist: We present a high-dimensional neural network potential (HDNNP) for the martensitic phase of the NiTi shape-memory alloy trained to density functional theory (DFT) data.
Key concepts
- High-Dimensional Neural Network Potential (HDNNP)
- This is a machine learning model trained on DFT data to predict the forces and energies of atoms in NiTi. It acts as a fast, accurate substitute for expensive quantum mechanical calculations, enabling large-scale simulations of atomic movements over time.
- Martensitic Phases (B19′ and B33)
- These are specific crystal structures that NiTi can adopt during its phase transformation. The HDNNP successfully predicts which phase is more stable at different temperatures, helping researchers understand the structural changes that occur during this transformation.
- Stacking-Fault Energy Landscape
- This describes how easily the crystal structure can slip or deform along specific planes. The model shows this landscape is highly directional, pointing to a specific shear direction ([100]M(001)M) where deformation is energetically favored.
- Finite-Temperature Molecular Dynamics (MD)
- This simulation method tracks the movement of individual atoms in NiTi over time at specific temperatures. The HDNNP enables these simulations to show how the structure gradually evolves from one phase (like B33) to another (like B19′) as heat is added.
Terminology
Summary
We present a high-dimensional neural network potential (HDNNP) for the martensitic phase of the NiTi shape-memory alloy trained to density functional theory (DFT) data. p. 0 A high-dimensional neural network potential for
finite-temperature phenomena in NiTi martensite
Petr Jaroš1
, Petr Sedlák∗2
, Petr Šesták2,3,4, Miroslav Černý3,4, Jörg Behler5,6, and Hanuš
Seiner2
1Department of Solid State Engineering, Faculty of Nuclear Science and Physical Engineering, Czech
Technical University in Prague, Prague, Czech Republic
2
Institute of Thermomechanics, Czech Academy of Sciences, Dolejškova 5, Prague 182 00, Czech Republic
3
Institute of Physical Engineering, Brno University of Technology, 616 69 Brno, Czech Republic
4CEITEC BUT, Brno University of Technology, 612 00 Brno, Czech Republic
5Lehrstuhl für Theoretische Chemie II, Ruhr-Universität Bochum, 44780 Bochum, Germany
6Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, 44780 Bochum, Germany
Abstract
We present a high-dimensional neural network potential (HDNNP) for the martensitic
phase of the NiTi shape-memory alloy trained to density functional theory (DFT) data. A central aspect of this work is the systematic validation of the potential with respect to
the underlying DFT reference method for key properties governing structural evolution,
including equilibrium crystal structures, elastic constants, generalized-stacking fault energies, and vibrational spectra. A central aspect of this work is the systematic validation of the potential with respect to
the underlying DFT reference method for key properties governing structural evolution,
including equilibrium crystal structures, elastic constants, generalized-stacking fault energies, and vibrational spectra. The HDNNP accurately describes the relative stability of the
B19′ and B33 phases, including subtle energy differences on the order of meV/atom. The HDNNP accurately describes the relative stability of the
B19′ and B33 phases, including subtle energy differences on the order of meV/atom. The
predicted stacking-fault energy landscape is strongly anisotropic and reveals a preferential
shear pathway, providing atomistic insight into deformation and twinning mechanisms. The
predicted stacking-fault energy landscape is strongly anisotropic and reveals a preferential
shear pathway, providing atomistic insight into deformation and twinning mechanisms. Finite-temperature molecular dynamics simulations further enable the investigation of unconstrained structural evolution as a function of temperature. Finite-temperature molecular dynamics simulations further enable the investigation of unconstrained structural evolution as a function of temperature. Overall, the developed
HDNNP provides a robust basis for atomistic simulations of the complex structural and
functional behavior of martensitic NiTi systems containing hundreds of thousands of atoms
on nanosecond time scales.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Validation Against Reference Data
The central aspect of this work is the systematic validation of the potential with respect to the underlying DFT reference method for key properties governing structural evolution, including equilibrium crystal structures, elastic constants, generalized-stacking fault energies, and vibrational spectra. The HDNNP reproduces the relative stability of these phases as well as the pronounced elastic anisotropy characteristic of B19′ at zero temperature. p. 2 The HDNNP is focused on the martensitic B19′ phase and its relation to the B33 structure. At
zero temperature, we demonstrate that the HDNNP reproduces the relative stability of these
phases as well as the pronounced elastic anisotropy characteristic of B19′
. At
zero temperature, we demonstrate that the HDNNP reproduces the relative stability of these
phases as well as the pronounced elastic anisotropy characteristic of B19′
. The analysis of
the generalized stacking-fault energy surface reveals a strong crystallographic anisotropy and
identifies a single energetically favorable shear pathway along the [100]M(001)M direction.Where the paper says it
Where the paper says it
Dataset Construction Methodology
The dataset is a key component of any machine learning model, constructed with a focus on two characteristic features of the martensitic phase of NiTi. The first is the discrepancy between the predicted and experimentally observed groundstate structures. p. 3 The dataset used in this
work was constructed with a focus on two characteristic features of the martensitic phase of
NiTi. The first is the discrepancy between the predicted and experimentally observed groundstate structures. However, this phase has never been observed experimentally; instead, experiments consistently report the B19′
structure [29, 30, 31]. The second
motivation arises from experimental observations indicating a higher concentration of defects
along a specific direction, namely in the (001) plane [29]. The presence of a preferred slip
system and strong anisotropy may play a crucial role in the plastic deformation process and
has been further investigated in [6]. Motivated by these findings, we constructed a DFT-based dataset comprising 7300 structures, ranging from 4-atom primitive cells to 180-atom supercells. Using this preliminary model, we performed molecular
dynamics (MD) simulations under various thermodynamic ensembles and conditions. During
these simulations, the dataset was iteratively expanded with additional configurations selected
according to two criteria: extrapolation parameters and an active learning strategy. Large discrepancies among the predictions of the committee members indicated that a given
local atomic environment was poorly represented in the current dataset and should therefore be
included in the retraining set. Because atomic forces serve as more sensitive indicators of model
accuracy than total energies, we analyzed the MD trajectories and extracted the configurations
exhibiting the largest force discrepancies, while preserving periodic boundary conditions, and
added them to the dataset. Because atomic forces serve as more sensitive indicators of model
accuracy than total energies, we analyzed the MD trajectories and extracted the configurations
exhibiting the largest force discrepancies, while preserving periodic boundary conditions, and
added them to the dataset. This procedure was repeated iteratively until the standard deviation
of the committee disagreement for each force component was less than 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Structural Property Reconstruction
The HDNNP reproduces the equilibrium structures and relative stability of the B19′ and B33 phases, as well as their elastic response at zero temperature. The HDNNP reproduces the strong elastic anisotropy of B19′, reflected in the low C55 elastic constant, in agreement with DFT results and with values reported in the literature. p. 7 The resulting elastic constants are summarized in
Table 3. The HDNNP reproduces the strong elastic anisotropy of B19′
, reflected in the low C55
elastic constant, in agreement with DFT results and with values reported in the literature [8,
8 This validation establishes a physically consistent reference state for the finite-temperature
simulations discussed in the following sections. Validation of the HDNNP with respect to DFT
calculations demonstrates that the HDNNP methodology is applied consistently, which can
then be considered as a method that enables the upscaling of DFT calculations and can still be
considered as an ab-initio method without any empirical inputs.Where the paper says it
Where the paper says it
Finite Temperature Dynamics
MD simulations performed in the NpT ensemble reveal a gradual structural evolution from configurations close to B33 at lower temperatures toward configurations approaching B19′ with increasing temperature, driven by a shear-like mechanism. For each temperature, a representative mean structure is identified and summarized in Table 4. p. 2 As
introduced in [14], configurations closely related to the B33 structure emerge as local minima
along this pathway, suggesting a shear-mediated structural connection between B19′ and B33. By allowing the system to relax freely at each temperature during MD simulations, we identify
representative structures corresponding to free energy minima. All models were trained using the same fitting procedure, while
differing in the composition of the training and test subsets and in the number of neurons. 3 Validation of HDNNP
In this section, we examine whether the HDNNP model reproduces the key structural characteristics obtained from DFT that are relevant to the temperature-driven evolution of the
martensitic phase. In the first case, the free energy was calculated based on the DFT energy and the
vibrational spectra, however, due to computational complexity, the free energy evolution was
mapped only along the transformation path defined by a single parameter (monoclinic angle
of the B19′
structure). In the latter, ab-initio molecular dynamics was used to simulate the
time evolution of the structure at a finite temperature, but only for a small atomistic model
(144-atom supercells—4×3×3 supercell of the 4 atom unit cells [13]).Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The construction of this potential is an important first step for further atomistic simulations of the exceptional properties of the B19′ structure, such as the presence of the easy [100]M(001)M slip, which is unusual for intermetallic martensitic systems.<ref:2607.20681X,The construction of this potential is an important first step for further atomistic simulations of the exceptional properties of the B19′ structure, such as the presence of the easy [100]M(001)M slip, which is unusual for intermetallic martensitic systems> The presented HDNNP potential for the B19′ martensitic phase is available at [37] in a form compatible with the LAMMPS and n2p2 packages. p. 1 These new findings have motivated a deeper study of the fundamental properties
of the B19′ martensitic structure. At the microstructural level, a mechanism for the unusual
plasticity of monoclinic martensite has recently been proposed [6]: based on a detailed theoretical analysis of experimentally observed plastically deformed martensitic microstructures, it
has been proposed that plastic deformation in martensite proceeds as cooperative twinning and
the formation and propagation of kink bands by a cooperative [100]M(001)M slip mechanism.Where the paper says it
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
-
Improved AI system can accurately predict atomic forces for large systems over longer timescales by using a high-dimensional neural network potential trained on density functional theory data, providing an
ab-initio method without any empirical inputs.
-
Improved AI system can investigate unconstrained structural evolution as a function of temperature by performing molecular dynamics simulations where the free energy is calculated within the harmonic approximation, allowing for a
new possibility of an unconstrained search for stable structures at finite temperatures while maintaining the accuracy of DFT energy calculations.
-
Improved AI system can predict structural and elastic properties with DFT-level accuracy by accurately reconstructing
the ground state structure of B33, the energy evolution along the B33 to B19′ path, the generalized stacking fault energy surface, the elastic properties of both B33 and B19' phases, and the phonon dispersion.
-
Improved AI system can identify a preferred shear pathway for plastic deformation by analyzing
the generalized stacking-fault energy surface
whichreveals a strong crystallographic anisotropy and identifies a single energetically favorable shear pathway along the [100]M(001)M direction.
Abstract
We present a high-dimensional neural network potential (HDNNP) for the martensitic phase of the NiTi shape-memory alloy trained to density functional theory (DFT) data. A central aspect of this work is the systematic validation of the potential with respect to the underlying DFT reference method for key properties governing structural evolution, including equilibrium crystal structures, elastic constants, generalized-stacking fault energies, and vibrational spectra. The HDNNP accurately describes the relative stability of the B19 and B33 phases, including subtle energy differences on the order of meV/atom. The predicted stacking-fault energy landscape is strongly anisotropic and reveals a preferential shear pathway, providing atomistic insight into deformation and twinning mechanisms. Finite-temperature molecular dynamics simulations further enable the investigation of unconstrained structural evolution as a function of temperature. Overall, the developed HDNNP provides a robust basis for atomistic simulations of the complex structural and functional behavior of martensitic NiTi systems containing hundreds of thousands of atoms on nanosecond time scales.
Related papers
- Structural contribution to light-induced gap suppression in Ta 2 NiSe 5
- Anatomy of fast current-induced skyrmion motion in synthetic antiferromagnets
- Magnetic control of an exciton-polariton condensate in a van der Waals magnet
- High-pressure elastic properties of GeO2 polymorphs up to 120 GPa
- Decoupling Strain-Rate Sensitivity and Deformation Length Scale Effects in Neutron-Irradiated Tungsten: A Coupled Nano-Indentation, HR-EBSD and Crystal Plasticity Study
- Magnetism modulation in Co3Sn2S2 by current-assisted domain wall motion