Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band Calculations for Small-Polaron Hopping in Rutile TiO2
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: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band Calculations for Small-Polaron Hopping in Rutile TiO2".
Mira: Here's the gist of the paper. Nudged elastic band (NEB) calculations of small-polaron hopping are demanding because nuclear motion is coupled to a self-consistent localized electronic state Near the charge-transfer region, slow self-consistent field convergence, localization switching, and competing electronic solutions complicate surrogate-assisted optimization The workflow demonstrates electron small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U correction (DFT+U) with SIESTA Three independently DFT-verified 96-atom
001: calculations using 11, 13, and 15 images and different surrogate settings and GP parameterizations yield barriers of 0.1645, 0.1637, and 0.1640 eV, respectively.
Kai: Pruning redundant reasoning tokens The paper develops a Gaussian-process-assisted NEB workflow designed to accelerate the calculation of small-polaron hopping barriers in rutile TiO2 by integrating controlled surrogate motion with targeted first-principles acquisition and explicit DFT acceptance.
Mira: This method addresses the computational challenges arising from slow self-consistent field convergence and competing electronic solutions near the charge-transfer region.
Kai: The workflow combines controlled surrogate extrapolation, targeted charge-transfer region interpolation, and independent DFT verification to obtain polaron migration pathways.
Title and authors: Kai: So, we're looking at this paper today. It’s called "Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band Calculations for Small-Polaron Hopping in Rutile TiO2". It sounds super technical, but what it's doing is trying to speed up calculations for how tiny charges hop around in titanium dioxide.
Mira: Yeah, it’s a tricky problem because the movement of the atoms—the nuclear motion—is tied directly to a localized electronic state. And that coupling creates these huge headaches when you try to use standard methods like Nudged Elastic Band calculations, especially near that charge-transfer region where things get messy.
Lev: From an engineering standpoint, what I hear is that the traditional way of doing this calculation requires too many slow steps for self-consistent field convergence and it can get stuck switching between different localized electronic solutions. That's a real bottleneck for getting reliable results on hardware.
Kai: Exactly, and this paper proposes using a Gaussian process assisted Nudged Elastic Band workflow to handle that coupling better by integrating controlled surrogate motion with targeted first-principles evaluations. It’s trying to find the migration pathways without running an endless cycle of slow convergence checks.
Mira: The core idea is to use a joint energy force Gaussian process with a Matérn covariance model to advance the elastic band, and they are training this process using geometries evaluated by spin-polarized density functional theory with a Hubbard U correction, DFT plus SIESTA.
Lev: They’re training it on ninety-six-atom systems that they verified independently using eleven thirteen and fifteen images under different surrogate settings and Gaussian process parameterizations. The results they got for the hopping barriers were about zero point one six four five eV, then dropping to about zero point one six three seven eV and finally around zero point one six four zero eV depending on the setup they used.
Kai: That's a pretty consistent result across those different setups, which is interesting because it shows the method is robust even when you change how you set up the surrogate model initially.
Mira: It’s also worth noting that their independent DFT-verified calculations showed barrier values of zero point one six four five eV, zero point one six three seven eV, and zero point one six four zero eV for the same system, which gives a nice baseline to compare against what this new workflow finds.
Lev: And what they did next was focus on how to make it faster without losing accuracy by using trust-limited optimization and DFT verification to separate surrogate convergence from true first-principles convergence. That’s where the real efficiency gain seems to come from, right?
Title and authors: Kai: Right, because they replace that slow full surrogate convergence with a sequence of locally trusted path updates, where atomic root mean square displacement and maximum atom displacement measures restrict how far the optimizer moves from the geometries they’ve already seen.
Mira: The controller alternates between doing this bounded surrogate motion and performing targeted DFT evaluations, stopping the optimizer before it even gets to full surrogate force convergence to trigger a reference acquisition.
Lev: That separation is key because a low surrogate Nudged Elastic Band force doesn't mean the underlying first-principles molecular electronic structure has converged yet, which is a big distinction for real computational physics.
Kai: They log staged self-consistent field iterations and repeated central configuration acquisitions to track this process, and then they freeze the band coordinates once the Gaussian process band satisfies both its surrogate force and uncertainty gates.
Mira: After that freezing point, they reevaluate all internal images using full DFT, and for that final full-band verification, they require that the maximum projected Nudged Elastic Band force be less than zero point zero five zero eV per Angstrom.
Lev: So, to put it plainly, this trust-limited optimization and DFT verification is how they ensure the reported barrier is actually a DFT-verified adiabatic barrier on the potential energy surface, not just a guess from the surrogate model.
Kai: And this process allows them to reduce the number of expensive first-principles evaluations substantially; for example, they demonstrated that under matched settings, this GP assisted approach reduced the number of DFT evaluations from two hundred fifty-five down to forty-seven which is an eighty-one point six percent reduction while still reproducing that conventional Nudged Elastic Band migration barrier within about zero point zero eight one meV.
Mira: That eighty-one point six percent reduction in DFT work is significant, but we have to remember that those barriers are calculated as zero K potential energy results using the adopted DFT plus U treatment, so we should be careful not to equate them directly with finite temperature transport activation energies.
Lev: Yeah, that’s a necessary caution. And they also looked at how cell size affects things; increasing the supercell from two by two by four to three by three by six changes both the cost of the underlying DFT calculation and the dimensionality of their joint energy force Gaussian process.
Kai: That change in dimensionality is substantial because dense Gaussian process factorization scales cubically with observations, and storage scales quadratically with the number of observations, which is why they tested a larger system.
Title and authors: Mira: They tested a three by three by six system, which has three hundred twenty-four atoms, to see if this energy force surrogate and active learning controller are still usable when each additional reference geometry introduces almost one thousand correlated observations.
Lev: And they found that the large cell barrier for the zero zero one barrier is about zero point one three zero two eV, and they pointed out that this remains substantially higher than the fifty-eight meV Nudged Elastic Band barrier reported by Morita et al., which gives a more realistic physical reference for transport properties.
Kai: So, to summarize what we’ve seen in the paper "Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band Calculations for Small-Polaron Hopping in Rutile TiO2," this workflow successfully integrates AI to handle the complexities of coupled nuclear motion and electronic states in rutile titanium dioxide.
Mira: The main implication is that this method provides a way to accelerate these computationally demanding calculations by smartly balancing surrogate exploration with targeted first-principles verification, especially near charge-transfer regions where standard methods struggle with convergence issues.
Lev: For those of us building real hardware, the point about separating surrogate convergence from DFT convergence is crucial because it gives us a clear stopping criterion that ensures the underlying first-principles molecular electronic structure has actually settled before we trust the result.
Kai: And they’ve shown that this approach can cut down on DFT evaluations by eighty-one point six percent while maintaining accuracy within about zero point zero eight one meV of the conventional Nudged Elastic Band migration barrier, which is a solid performance metric for this type of calculation.
Mira: But there are still things they flag, like the need for component ablation studies to quantify exactly what benefits come from trust-limited sampling versus local Gaussian process prediction.
Lev: And they also stressed that electronic branch consistency needs direct localization diagnostics alongside nuclear force verification to truly nail down the electronic structure stability of these localized states.
Kai: So, as a final thought on this paper, the results support the feasibility of controlled Gaussian assisted optimization for periodic polaron pathways in rutile TiO2, showing it’s a viable tool for tackling these specific computational challenges.
Mira: We should keep an eye on how they use this trust-limited sampling idea in other systems, because handling coupled ionic and electronic transport is a shared challenge across many solid-oxide electrochemical systems.
Lev: I think the immediate practical value is in showing that we can get reliable, DFT-verified migration barriers much faster than before using this machine learning approach for these complex oxide environments.
The paper's summary: Kai: So, to wrap up what we just heard about that paper on rutile titanium dioxide hopping barriers, they’re essentially saying this new machine learning workflow lets them figure out those migration pathways much faster by smartly mixing trial geometry searches with actual first-principles calculations.
Mira: Right, and the real kicker is how they handle that coupling between the atomic movement and the electronic state near the charge-transfer region. They’re using a joint energy force Gaussian process to guide this search without getting bogged down in those slow self-consistent field convergence issues we talked about earlier.
Lev: From an error correction standpoint, what I find interesting is their method for separating surrogate convergence from true first-principles convergence using that force criterion gate. That means they’re not just trusting the surrogate model when it looks good; they’re forcing the system to meet a physical threshold before they let the result stand.
Kai: And that separation is crucial because a low force prediction doesn't mean the underlying electronic structure has actually settled, which is something we always worry about when building models for real hardware fidelity.
Mira: They’re also being pretty rigorous by running independent DFT verifications on different ninety-six-atom images—eleven, thirteen, and fifteen images—and they found consistent barrier values around zero point one six four five electron volts.
Lev: That consistency across different setups tells us the method is robust for this type of system, but it also means we need to be careful about what those zero point one six four five electron volt numbers actually represent, because they are based on a specific DFT plus U treatment.
Kai: And that’s a big caveat; they explicitly state those results are zero K potential energy results and shouldn't be directly compared to finite temperature transport activation energies, which is something we have to remember when we talk about real device performance.
Mira: They did look at how cell size affects things too, showing that going from a two by two by four system up to a three by three by six system means the dimensionality of their Gaussian process gets much bigger because the number of observations scales cubically with that.
Lev: That scaling is huge for computational cost; dense Gaussian process factorization scales cubically with the number of observations and storage scales quadratically, so they tested that larger system to see if it still made sense when each new reference geometry adds almost a thousand correlated observations.
Kai: And what they found in that large cell test was interesting: the barrier for that zero zero one barrier in the larger cell, which is about zero point one three zero two electron volts, remains substantially higher than the fifty eight millivolt migration barrier reported by Morita et al., which gives a more realistic physical reference for transport properties.
Mira: So, while they show this AI approach can cut down on expensive first-principles evaluations by eighty-one point six percent to get within zero point zero eight one millivolt of the conventional Nudged Elastic Band migration barrier, they also flag that component ablation studies are still needed to quantify exactly what benefit comes from the trust-limited sampling versus just local Gaussian process prediction.
Lev: And for me, the implication is that this method shows we can get reliable, DFT verified migration barriers much faster than before using this machine learning approach for these complex oxide environments, which is a practical win for accelerating error correction simulations.
Kai: It’s definitely promising because it moves us away from just brute-force sampling toward a more intelligent way of navigating those complicated potential energy surfaces in materials science.
The paper's improvements: Kai: So, we’re looking at what they suggest as improvements to this trust-limited workflow for hopping in rutile titanium dioxide. They’re really pushing the idea of using component ablation studies to see exactly what benefit you get from doing the trust-limited sampling versus just relying on that local Gaussian process prediction.
Mira: Right, and they also pointed out that electronic branch consistency needs direct localization diagnostics alongside just checking the nuclear force verification, which is a big point because we need to know if the electronic structure itself is stable during those hops.
Lev: From my side, what I think is important is that this pushes us toward a more holistic approach; it’s not just about speeding up one step, it’s about making sure every part of the calculation—the nuclear motion and the electronic state—is validated against first-principles accuracy.
Kai: And they are implying that we need to be careful when we talk about those zero K potential energy results because they aren't directly comparable to finite temperature transport activation energies, which is a necessary clarification for anyone using this for real device simulation.
Mira: They’re also suggesting that the results should not be equated directly with finite temperature transport activation energies, so even with the speed boost and verification gates, we have to keep that thermal energy aspect in mind when interpreting the final numbers.
Lev: I think what they are pointing toward is a pathway where we can build more robust error correction models because this workflow gives us a way to verify if the electronic structure settled correctly before we even try to run the full path optimization.
Kai: It means that instead of just accepting a barrier from the surrogate model, you have this structured way—this ablation study—to prove that the underlying first-principles molecular electronic structure is stable during those hops.
Mira: So, they’re suggesting we need to be explicit about what we gain from the trust-limited sampling versus just using the local GP prediction, which helps us understand where the real computational savings actually come from in practice.
Lev: That sounds like a necessary step toward making these AI-assisted methods truly reliable for high-fidelity research, moving beyond just getting a fast number to understanding the physics behind that number.
Kai: It shows that this machine learning approach isn't just about finding shortcuts; it’s about building a more controlled and verifiable process for exploring those complex potential energy surfaces in materials science.
Conclusion: Kai: So, to wrap up this discussion on the "Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band Calculations for Small-Polaron Hopping in Rutile TiO2," the main point is that they’ve shown a way to use AI to navigate those tricky coupled electronic and nuclear motions without getting stuck in slow convergence loops.
Mira: Exactly, it’s about integrating controlled surrogate motion with targeted first-principles evaluation to find those migration pathways efficiently, especially near charge-transfer regions where things get complicated.
Lev: What this means for running simulations on real hardware is that we can use a verification gate based on force criteria to stop the optimization before we trust a result that hasn't settled electronically.
Kai: And they did prove that this AI approach cuts down the number of expensive first-principles evaluations by eighty-one point six percent while still getting within zero point zero eight one millivolt of the conventional migration barrier.
Mira: That reduction in DFT work is significant, but we have to remember those barriers are calculated as zero K potential energy results and shouldn't be directly equated with finite temperature transport activation energies.
Lev: Yeah, that’s a necessary caution; we still need component ablation studies to quantify exactly what benefit you get from the trust-limited sampling versus just local Gaussian process prediction.
Kai: They also flagged that electronic branch consistency needs direct localization diagnostics alongside nuclear force verification to truly nail down the electronic structure stability of those localized states.
Mira: So, this paper provides a framework for controlled Gaussian assisted optimization that works well for periodic polaron pathways in rutile titanium dioxide.
Lev: It’s a solid step forward in showing we can accelerate these kinds of calculations by intelligently balancing surrogate exploration with targeted first-principles verification.
Kai: That gives us a better tool to tackle the computational challenges that come up when modeling complex oxide environments like rutile titanium dioxide.
Kai: We've covered the details of how this specific workflow works and what the numbers tell us about the reduction in DFT evaluations for these small-polaron hopping barriers.
Mira: It really highlights how important it is to have those verification gates built into an optimization loop when dealing with systems where nuclear motion is tied to localized electronic states.
Lev: For me, seeing that separation between surrogate convergence and true first-principles convergence is a real win because it gives us a clear stopping criterion for what we can trust in our error correction simulations.
Kai: So the main implication is that this AI workflow offers a way to accelerate these computationally demanding calculations by smartly balancing surrogate exploration with targeted first-principles verification, which is really useful for getting reliable results faster.
Mira: It’s about making sure that when we do run these complex Nudged Elastic Band calculations, we aren't just accepting a guess from the surrogate model, but something that has been verified against the underlying molecular electronic structure.
Lev: We can use this to build more robust error correction models because this workflow gives us a way to verify if the electronic structure settled correctly before we even try to run the full path optimization.
Kai: So, as a final thought on the "Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band Calculations for Small-Polaron Hopping in Rutile TiO2," it supports the feasibility of controlled Gaussian assisted optimization for periodic polaron pathways.
Mira: It’s definitely promising because it moves us away from just brute-force sampling toward a more intelligent way of navigating those complicated potential energy surfaces in materials science.
Lev: I think the immediate practical value is in showing that we can get reliable, DFT verified migration barriers much faster than before using this machine learning approach for these complex oxide environments.
Ashish Kumar Singh, Irina V. Lebedeva
Institute of Atomic Physics and Spectroscopy, University of Latvia · CIC nanoGUNE BRTA · Simune Atomistics
cond-mat.mtrl-sci
Submitted: 2026-10-12
Updated: 2026-11-02
License: http://creativecommons.org/publicdomain/zero/1.0/
Importance score: 80/100
The gist: Nudged elastic band (NEB) calculations of small-polaron hopping are demanding because nuclear motion is coupled to a self-consistent localized electronic state Near the charge-transfer region…
Key concepts
- Nudged Elastic Band (NEB)
- A computational method used to find the minimum energy path between two stable atomic configurations. It models the movement of atoms along a path by treating them as a flexible chain, minimizing the total energy of that chain to locate an activation barrier for processes like hopping.
- Gaussian Process (GP)
- A machine learning tool used here to build an intelligent surrogate model. It learns the relationship between atomic positions and their corresponding energies and forces from initial DFT calculations. This allows the system to predict energies quickly along a path, guiding where expensive DFT calculations are needed most.
- Trust-Limited Optimization
- A strategy where the optimization process stops relying solely on fast surrogate predictions. The system alternates between making small, trusted steps using the surrogate model and performing targeted DFT checks. This ensures that the final calculated barrier is verified by high-accuracy first-principles calculations.
- DFT+U Correction
- A specific type of electronic structure calculation used to accurately model materials like TiO2. The 'U' term corrects for strong electron correlation effects that standard DFT methods miss, which is crucial for correctly describing the localized electronic states and polaron behavior.
Terminology
Summary
Nudged elastic band (NEB) calculations of small-polaron hopping are demanding because nuclear motion is coupled to a self-consistent localized electronic state Near the charge-transfer region, slow self-consistent field convergence, localization switching, and competing electronic solutions complicate surrogate-assisted optimization The workflow demonstrates electron small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U correction (DFT+U) with SIESTA Three independently DFT-verified 96-atom [001] calculations using 11, 13, and 15 images and different surrogate settings and GP parameterizations yield barriers of 0. p. 0 Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Three independently DFT-verified 96-atom [001]
calculations using 11, 13, and 15 images and different surrogate settings and GP parameterizations yield
barriers of 0. 1637, and 0. 1637, and 0. 1640 eV, respectively, spanning 0.Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The paper develops a Gaussian-process-assisted NEB workflow designed to accelerate the calculation of small-polaron hopping barriers in rutile TiO2 by integrating controlled surrogate motion with targeted first-principles acquisition and explicit DFT acceptance. This method addresses the computational challenges arising from slow self-consistent field convergence and competing electronic solutions near the charge-transfer region. p. 0 com
Abstract
Nudged elastic band (NEB) calculations of small-polaron hopping are demanding because nuclear motion is
coupled to a self-consistent localized electronic state. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. The workflow combines controlled surrogate extrapolation, targeted charge-transferregion interpolation, and independent DFT verification to obtain polaron migration pathways, emphasizing
reproducible barriers and full-band force convergence. 1 Introduction
Small polarons arise from the localization of charge
carriers through their coupling to lattice distortions,
resulting in quasiparticles whose migration involves
both charge transfer and local structural reorganization [1–5].Where the paper says it
Where the paper says it
How the training signal is built
The Gaussian process (GP) is formulated as a joint energy–force GP with a Matérn covariance model that advances the elastic band. The training dataset consists of Ng evaluated geometries, each contributing the full-system DFT energy and their Cartesian forces. p. 4 3 Training configurations and Gaussian process prior
Let A contain Na atoms, D = 3Na their Cartesian
degrees of freedom, and q ∈ R
D their coordinates. The
retained dataset contains Ng evaluated geometries,
each contributing the full-system DFT energy and
their Cartesian forces:
D = {(qn, En, Fn,A)}
Ng
n=1 , M = Ng(1 + D) (4)
The joint energy–force GP formulation for the
RMS Cartesian metric and Matérn covariance kernel
is defined as follows [50, 51, 53]. 3 Training configurations and Gaussian process prior
Let A contain Na atoms, D = 3Na their Cartesian
degrees of freedom, and q ∈ R
D their coordinates. The
retained dataset contains Ng evaluated geometries,
each contributing the full-system DFT energy and
their Cartesian forces:
D = {(qn, En, Fn,A)}
Ng
n=1 , M = Ng(1 + D) (4)
The joint energy–force GP formulation for the
RMS Cartesian metric and Matérn covariance kernel
is defined as follows [50, 51, 53].Where the paper says it
Where the paper says it
Where the gains come from
The primary efficiency gain comes from replacing full surrogate convergence with a sequence of locally trusted path updates, where atomic root-mean-square (RMS) and maximum-atom displacement measures restrict the distance from evaluated configurations. The controller alternates bounded surrogate motion and targeted DFT evaluation, stopping the optimizer before surrogate force convergence to trigger a reference acquisition. p. 5 1 Trust-limited acquisition and full-band
verification
The controller alternates bounded surrogate motion
and targeted DFT evaluation. A training-distance or
cycle boundary can stop the optimizer before surrogate force convergence and trigger a reference acquisition. A low surrogate NEB force does not, by itself, establish convergence of the underlying first-principles
MEP. The production controller therefore separates
surrogate convergence from DFT convergence.Where the paper says it
Where the paper says it
Trust-limited optimization and DFT verification
The workflow controller logs staged central-image SCF iterations, repeated central configuration acquisitions, and pre-acquisition prediction audits. The production controller separates surrogate convergence from DFT convergence, once the GP band satisfies the surrogate force and uncertainty gates, the band coordinates are frozen and all internal images are reevaluated using DFT. p. 7 The production controller therefore separates
surrogate convergence from DFT convergence. Once
the GP band satisfies the surrogate force and uncertainty gates, the band coordinates are frozen and all
internal images are reevaluated using DFT. In the 15-image calculation,
repeated central-image sampling activates a separately calibrated local GP during surrogate optimization. Final full-band DFT verification gives a maximum projected NEB force of 0.Where the paper says it
Where the paper says it
Scaling and supercell-size dependence
Increasing the supercell from 2 × 2 × 4 to 3 × 3 × 6 changes not only the cost of the underlying DFT calculation but also the dimensionality of the joint energy-force GP. The increase is substantial because dense GP factorization scales cubically with the number of observations and storage scales quadratically. p. 9 With the initial 11-geometry training band, the corresponding
covariance dimensions are therefore 3179 and 10703. The increase is substantial because dense GP factorization scales cubically with the number of observations and storage scales quadratically. The increase is substantial because dense GP factorization scales cubically with the number of observations and storage scales quadratically. The 324-atom
calculation therefore represents more than a straightforward application of the same algorithm to a more
expensive DFT system: it directly tests whether the
energy-force surrogate and active-learning controller
remain usable when each additional reference geometry introduces almost one thousand correlated observations.Where the paper says it
Where the paper says it
Polaron localization and comparison with previous work
The localized endpoint in the present calculations has a site-resolved spin population of approximately 0.83 in the adopted population analysis. p. 3 point geometry obtained by linear interpolation between the endpoints. The spatial distribution of the
excess electron was assessed using site-resolved Hirshfeld and Voronoi population analyses. We therefore formulate the GP-assisted NEB calculation as a sequence of locally trusted path updates
rather than requiring full surrogate convergence as the
routine inner-loop objective. A global GP advances
the elastic band while atomic root-mean-square (RMS)
and maximum-atom displacement measures restrict
the distance from evaluated configurations. First-principles calculations are widely employed
to investigate polaron localization, structural relaxation, and hopping barriers in oxides and related materials [5, 31–35]. Comparisons across TiO2 polymorphs
have demonstrated that the stability and spatial extent of localized states depend on crystal structure
and the treatment of self-interaction [36]. Optimized NormConserving Vanderbilt (ONCVPSP) fully relativistic
pseudopotentials [65] in the PSML format [66] were
sourced from the PseudoDojo database [67]. The
pseudopotential for Ti explicitly included the 3s and
3p semicore states in the valence pool, corresponding
to a valence configuration of 3s
23p
63d
24s
2
. These contributions, together with recent methodological reviews [59], demonstrate that uncertaintydirected acquisition, adaptive surrogate refinement,
and reference-level validation are established ideas. Their application to polaron migration nevertheless
requires attention to the self-consistent electronic solutions underlying each energy and force observation. 050 eV/Å criterion. The barrier spread of 0. The barrier spread of 0. 793 meV shows agreement
across the tested setups.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
Where the paper says it
Conclusions and Outlook
The results support the feasibility of controlled GPassisted optimization for periodic polaron pathways. The 96-atom, 13-image system shows that the GP-assisted approach reduces the number of DFT evaluations from 255 to 47 (81. p. 8 The two approaches yielded migration barriers differing by only 0. Under otherwise identical
computational settings, the GP-assisted approach reduced the number of DFT evaluations from 255 to
47, corresponding to a reduction of 81. Spinpolarized DFT+U calculations with SIESTA are performed for different NEB discretizations and supercell
sizes to assess the reproducibility and scalability of
the approach. A direct comparison with conventional
NEB under identical computational settings is used to
quantify the reduction in first-principles evaluations
while maintaining DFT-verified migration barriers
and force convergence. Direct
localization diagnostics remain necessary to attribute
individual errors to SCF branch changes. After activation, the controller explicitly reported
the central prediction model as the local GP and resumed production surrogate optimization. The controller subsequently recorded local-GP
predictions during production optimization. Direct
localization diagnostics remain necessary to attribute
individual errors to SCF branch changes. Molecular dynamics simulations probe finite-temperature carrier motion more
directly [44–46], but generally require substantially
greater statistical sampling to characterize infrequent
hopping events. The NEB method is a standard approach for determining minimum-energy paths (MEPs) and saddlepoint barriers between known end states [40, 41, 47].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
Acknowledgements
This research by A. K. S. was conducted during a Secondment at Simune Atomistics and was supported by Horizon Europe MSCA Postdoctoral Fellowship grant no. p. 0 Trust-Limited Machine-Learning-Accelerated Nudged Elastic Band
Calculations for Small-Polaron Hopping in Rutile TiO2
Ashish Kumar Singh1 and Irina V. Lebedeva2, 3
1
Institute of Atomic Physics and Spectroscopy, University of Latvia, Riga, Latvia
2CIC nanoGUNE BRTA, Avenida de Tolosa 76, San Sebastián 20018, Spain
3Simune Atomistics, Avenida de Tolosa 76, San Sebastián 20018, Spain
*Email: kumar24@lu. The illustrated Ti53–Ti55 pathway
follows the crystallographic c axis and represents the
[001] electron-hopping direction. The 324-atom (3 × 3 × 6) GP-assisted NEB calculation was performed on the Hyperion supercomputer
at the DIPC Supercomputing Center, using SIESTA
version 5.Where the paper says it
Where the paper says it
Supporting Information
Additional mathematical details of the joint energy–force Gaussian-process formulation, training-data management and DFT evaluation accounting, together with supplementary atomic-structure, sampling-history, and electronic-density figures (PDF).
Data Availability Statement
The data supporting the findings of this study, including the output logs from the first-principles calculations and NEB optimizations, are available from the corresponding authors upon reasonable request. The source code developed for the GP-assisted NEB calculations will be made publicly available in an open-source repository. p. 1 implemented
an on-the-fly energy–force GP calculator that adapts
reference sampling to predictive uncertainty during
NEB calculations [56], while Simon et al. developed an active-learning NEB algorithm for identifying
transition states in molecular and high-dimensional
benchmark systems [57].Where the paper says it
References
(1) Holstein, T. Ann. Phys. 1959, 8, 325–342
(4) Austin, I. G.; Mott, N. F. Adv. Phys. 1969, 18, 41–102
(3) Emin, D. J. Solid State Chem. p. 1 Charge localization and defect-mediated electronic transport are also relevant to resistive-switching
devices [25–28], while coupled ionic and electronic
transport plays an important role in solid-oxide electrochemical systems [29, 30]. Despite their different
functional roles, these materials share a computational
challenge: determining the energetics and migration
pathways of localized charge carriers.Where the paper says it
(7) Rühle, V.; Lukyanov, A.; May, F.; Schrader, M.; Vehoff, T.; Kirkpatrick, J.; Baumeier, B.; Andrienko, D. J. Chem. Theory Comput. p. 0 Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA. Near the charge-transfer region, slow self-consistentfield convergence, localization switching, and competing electronic solutions complicate surrogate-assisted
optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron
small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U
correction (DFT+U) with SIESTA.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
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
(5) Franchini, C.; Reticcioli, M.; Setvin, M.; Diebold, U. Nat. Rev. Mater. 2021
(9) Sun, L.; Huang, X.; Wang, L.; Janotti, A. Phys. Rev. B 2017
(3) Janotti, A.; Varley, J. B.; Choi, M.; Van de Walle, C. G. Phys. Rev. B 2014
(32) Deskins, N. A.; Dupuis, M. Phys. Rev. B 2007
(47) Henkelman, G.; Uberuaga, B. P.; Jónsson, H. p. 3 The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41]. The calculation was executed on a compute node equipped with Intel Xeon Gold 6342 (Ice
Lake) processors, providing 48 CPU cores per node,
of which 24 were allocated to the calculation. 2 Nudged elastic band formulation
For image i, the physical force is Fi = −∇E(Ri), and
τ i
is the normalized path tangent, evaluated using
the improved tangent prescription of Henkelman and
Jónsson [41].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
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
(53) Koistinen, O.-P.; Ásgeirsson, V.; Vehtari, A.; Jónsson, H. J. Chem. Theory Comput. 2019
(74) Simune Atomistics, S.L. ASAP-2025.1 — Atomistic Simulation Advanced Platform
(73) Marchesin, F.; Koval, P.; Pouillon, Y.; Lebedeva, I.; García, A.; García-Mota, M.; Kimmel, A. In Psi-k Conference 2022
(75) Bitzek, E.; Koskinen, P.; Gähler, F.; Moseler, M.; Gumbsch, P. Phys. Rev. Lett. 2006
(77) Freysoldt, C.; Grabowski, B.; Hickel, T.; Neugebauer, J.; Kresse, G.; Janotti, A.; Van de Walle, C. G. Rev. Mod. Phys. 2014
(79) Baur, W. H. Acta Crystallogr. 1956
(33) Bylaska, E. J.; Rosso, K. J.; Chem. Theory Comput. 2018
(34) De Lile, J. R.; Kang, S. G.; Son, Y.-A.; Lee, S. G. ACS Omega 2019
(35) Himmetoglu, B.; Janotti, A.; Bjaalie, L.; Van de Walle, C. G. Phys. Rev. B 2014
(39) Carey, J. J.; Quirk, J. A.; McKenna, K. P. J. Phys Chem C 2021
(47) Henkelman, G.; Uberuaga, B. P.; Jónsson, H. J. Chem. Phys. 2000
(53) Koistinen, O.-P.; Ásgeirsson, V.; Vehtari, A.; Jónsson, H. J. Chem. Theory Comput. 2019
(74) Simune Atomistics, S.L. ASAP-2025.1 — Atomistic Simulation Advanced Platform
(73) Marchesin, F.; Koval, P.; Pouillon, Y.; Lebedeva, I.; García, A.; García-Mota, M.; Kimmel, A. In Psi-k Conference 2022
(75) Bitzek, E.; Koskinen, P.; Gähler, F.; Moseler, M.; Gumbsch, P. Phys. Rev. Lett. 2006
(77) Freysoldt, C.; Grabowski, B.; Hickel, T.; Neugebauer, J.; Kresse, G.; Janotti, A.; Van de Walle, C. G. Rev. Mod. Phys. 2014
(79) Baur, W. H. Acta Crystallogr. 1956
(33) Bylaska, E. J.; Rosso, K. J.; Chem. Theory Comput. 2018
(47) Henkelman, G.; Uberuaga, B. P.; Jónsson, H. J. Chem. Phys. 2000
(53) Koistinen, O.-P.; Ásgeirsson, V.; Vehtari, A.; Jónsson, H. J. Chem. Theory Comput. 2019
(74) Simune Atomistics, S.L. ASAP-2025.1 — Atomistic Simulation Advanced Platform
(73) Marchesin, F.; Koval, P.; Pouillon, Y.; Lebedeva, I.; García, A.; García-Mota, M.; Kimmel, A. In Psi-k Conference 2022
(75) Bitzek, E.; Koskinen, P.; Gähler, F.; Moseler, M.; Gumbsch, P. Phys. Rev. Lett. 2006
(77) Freysoldt, C.; Grabowski, B.; Hickel, T.; Neugebauer, J.; Kresse, G.; Janotti, A.; Van de Walle, C. G. Rev. Mod. Phys. 2014
(79) Baur, W. H. Acta Crystallogr. 1956
(33) Bylaska, E. J.; Rosso, K. J.; Chem. Theory Comput. 2018
(47) Henkelman, G.; Uberuaga, B. P.; Jónsson, H. J. Chem. Phys. 2000
(53) Koistinen, O.-P.; Ásgeirsson, V.; Vehtari, A.; Jónsson, H. J. Chem. Theory Comput. 2019
(74) Simune Atomistics, S.L. ASAP-2025.1 — Atomistic Simulation Advanced Platform
(73) Marchesin, F.; Koval, P.; Pouillon, Y.; Lebedeva, I.; García, A.; García-Mota, M.; Kimmel, A. In Psi-k Conference 2022
(75) Bitzek, E.; Koskinen, P.; Gähler, F.; Moseler, M.; Gumbsch, P. Phys. Rev. Lett. 2006
(77) Freysoldt, C.; Grabowski, B.; Hickel, T.; Neugebauer, J.; Kresse, G.; Janotti, A.; Van de Walle, C. G. Rev. Mod. Phys. 2014
(79) Baur, W. H. Acta Crystallogr. 1956
(33) Bylaska, E. J.; Rosso, K. J.; Chem. Theory Comput. 2018
(47) Henkelman, G.; Uberuaga, B. P.; Jónsson, H. J. Chem. Phys. 2000
(53) Koistinen, O.-P.; Ásgeirsson, V.; Vehtari, A.; Jónsson, H. J. Chem. Theory Comput. 2019
(74) Simune Atomistics, S.L. ASAP-2025.1 — Atomistic Simulation Advanced Platform
(73) Marchesin, F.; Koval, P.; Pouillon, Y.; Lebedeva, I.; García, A.; García-Mota, M.; Kimmel, A. In Psi-k Conference 2022
(75) Bitzek, E.; Koskinen, P.; Gähler, F.; Moseler, M.; Gumbsch, P. Phys. Rev. Lett. 2006
(77) Freysoldt, C.; Grabowski, B.; Hickel, T.; Neugebauer, J.; Kresse, G.; Janotti, A.; Van de Walle, C. G. Rev. Mod. Phys. 2014
(79) Baur, W. H. Acta Crystallogr. 1956
(33) Bylaska, E. J.; Rosso, K. J.; Chem. Theory Comput. 2018
(47) Henkelman, G.; Uberuaga, B. P.; Jónsson, H. J. Chem. Phys. 2000
(53) Koistinen, O.-P.; Ásgeirsson, V.; Vehtari, A.; Jónsson, H. J. Chem. Theory Comput. 2019
(74) Simune Atomistics, S.L. ASAP-2025.1 — Atomistic Simulation Advanced Platform
(73) Marchesin, F.; Koval, P.; Pouillon, Y.; Lebedeva, I.; García, A.; García-Mota, M.; Kimmel, A. In Psi-k Conference 2022
(75) Bitzek, E.; Koskinen, P.; Gähler, F.; Moseler, M.; Gumbsch, P. Phys. Rev. Lett. 2006
(77) Freysoldt, C.; Grabowski, B.; Hickel, T.; Neugebauer, J.; Kresse, G.; Janotti, A.; Van de Walle, C. G. Rev. Mod. Phys. 2014
(79) Baur, W. H.
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
-
Trust-limited Gaussian-processassisted NEB workflow for small-polaron hopping allows AI systems to explore complex potential energy surfaces by balancing surrogate motion with first-principles evaluations, enabling the determination of migration pathways while controlling computational cost.
-
The system can perform automated DFT verification where
Final full-band DFT verification gives a maximum projected NEB force of 0.0346 eV/Å,
ensuring that reported barriers are not merely surrogate predictions but are insteadDFT-verified adiabatic barrier[s] associated with the chosen PBE+U/LCAO potential energy surface.
-
The AI can implement a hybrid global and local Gaussian process model where
The controller alternates bounded surrogate motion and targeted DFT evaluation,
allowing for efficient exploration of the entire band while specifically activating a local GP around theproblematic central region
to address electronic-state sensitivity. -
The system can distinguish between apparent convergence in energy profiles and true first-principles convergence by implementing a verification gate based on force criteria, as shown by the observation that
A low surrogate NEB force does not, by itself, establish convergence of the underlying first-principles MEP.
-
The AI can reduce the number of expensive first-principles evaluations substantially; for example, it demonstrated that under matched settings,
the GP-assisted approach reduced the number of DFT evaluations from 255 to 47 (81.6%) while reproducing the conventional NEB migration barrier within 0.081 meV.
-
The system can quantify cell-size sensitivity by comparing results across different supercell sizes, showing that
The large-cell [001] barrier of 0.1302 eV remains substantially higher than the 58 meV NEB barrier reported by Morita et al.,
providing a more realistic physical reference for transport properties.
Abstract
Nudged elastic band (NEB) calculations of small-polaron hopping are demanding because nuclear motion is coupled to a self-consistent localized electronic state. Near the charge-transfer region, slow self-consistent-field convergence, localization switching, and competing electronic solutions complicate surrogate-assisted optimization. We develop a Gaussian-process-assisted NEB workflow for this regime and demonstrate electron small-polaron hopping in rutile TiO2 using spin-polarized density functional theory with a Hubbard U correction (DFT+U) with SIESTA. A joint energy-force Gaussian process (GP) with a Matérn covariance model advances the elastic band, while active-atom root-mean-square and maximum-atom displacement limits constrain surrogate motion relative to evaluated configurations. Reaching these limits triggers first-principles sampling without requiring full surrogate convergence. Three independently DFT-verified 96-atom [001] calculations using 11, 13, and 15 images and different surrogate settings and GP parameterizations yield barriers of 0.1645, 0.1637, and 0.1640 eV, respectively, spanning 0.793 meV. In the 15-image calculation, repeated central-image sampling activates a separately calibrated local GP during surrogate optimization. Final full-band DFT verification gives a maximum projected NEB force of 0.0346 eV/Å. A 324-atom calculation yields a lower barrier of 0.1302 eV, indicating supercell sensitivity and demonstrating feasibility at larger GP observation dimension. The workflow combines controlled surrogate extrapolation, targeted charge-transfer-region interpolation, and independent DFT verification to obtain polaron migration pathways, emphasizing reproducible barriers and full-band force convergence.
Sources
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