Pith. sign in

REVIEW 1 major objections 5 minor 147 references

Machine learning for materials must learn free energy, not static energy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:14 UTC pith:O6AXBJYV

load-bearing objection A useful, competent Perspective that makes the case for free-energy-aware ML; the direct-learning roadmap has a real smoothness gap at first-order transitions, but the broader thesis holds. the 1 major comments →

arxiv 2607.26296 v1 pith:O6AXBJYV submitted 2026-07-28 cond-mat.mtrl-sci

Thermodynamics-Informed Machine Learning for Energy Materials Discovery

Pol Ben\'itez , Cibr\'an L\'opez , Claudio Cazorla This is my paper
classification cond-mat.mtrl-sci
keywords free energymachine learningmaterials discoverythermodynamicsentropyanharmonicityphase stabilityenergy materials
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that current machine learning models for materials are fundamentally limited because they learn zero-temperature internal energies from static structures, and so have no access to entropy and cannot reproduce the free energy that governs real materials at operating temperatures. The authors propose shifting the target quantity from internal energy U to free energy F, and making temperature an explicit input and training data that encode thermal physics—molecular dynamics configurations, anharmonic vibrations, and entropy contributions. They outline a roadmap centered on direct free-energy learning, entropy-aware representations, and active learning across temperature, and they argue this is essential for predicting phase stability and functionality in energy materials such as photovoltaic absorbers, catalysts, and solid electrolytes. A sympathetic reader would care because many technologically important materials are only stable at finite temperature, and zero-temperature ML models can qualitatively mispredict which phases exist and which properties they have.

Core claim

The paper's central claim is that the zero-temperature approximation is a fundamental, not merely technical, limitation of ML models for materials: a model trained on internal energy U has no access to entropy S and therefore cannot reproduce the Gibbs free energy G(T). It cannot distinguish a phase stable only at 0 K from one stabilized by entropy at finite temperature, cannot predict thermally induced phase transformations, and can mislead convex-hull screening. The authors propose reorienting the field toward thermodynamics-informed ML, where temperature is an explicit input, the free energy F is the primary learning target, and thermodynamic consistency—such as S = -∂F/∂T and positive he

What carries the argument

The central object is the free energy F (or Gibbs free energy G = F + PV) as the thermodynamic potential that determines phase stability. The paper uses the identities G = U - TS + PV and S = S_vib + S_conf + S_elec + S_mag to show why energy-only models fail, and it proposes a roadmap built on direct learning of F(T) with smooth activation functions, multi-task training on F, S, and C_v, and active-learning loops that sample high-uncertainty points in composition–temperature space.

Load-bearing premise

The roadmap assumes that finite-temperature training labels can be produced in sufficient quantity and quality, and that a learned free-energy function can extrapolate across phase transitions; the paper itself notes that training sets are small and that models fail when extrapolated to new temperature ranges, especially near phase transitions.

What would settle it

A direct test: take a strongly anharmonic material with a well-known temperature-driven phase transition, train a free-energy-learning model on finite-temperature molecular dynamics labels, and check whether it predicts the experimental transition temperature and entropy jump. If a model trained only on zero-temperature static energies also reproduces the transition, or if the learned free energy violates thermodynamic consistency (for instance, predicting negative heat capacity near the transition), the central claim loses its force.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, ML screening pipelines should shift from zero-temperature convex hulls to finite-temperature free-energy hulls, changing which phases are predicted to be stable.
  • Strongly anharmonic materials, such as halide perovskites whose room-temperature phase is dynamically unstable at 0 K, become tractable for ML prediction.
  • Thermodynamically consistent ML models could predict temperature-dependent band gaps, ionic conductivities, and catalytic free-energy barriers in regimes where harmonic approximations fail.
  • Training data generation must include molecular dynamics trajectories and entropy-related labels, not just relaxed static structures.
  • A foundational model trained on both static and finite-temperature data could enable transferable prediction of phase stability across chemical space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A near-term testable extension: train a graph neural network on (structure, temperature) to predict free energies for a small set of polymorphs with known experimental transition temperatures, then compare predicted phase diagrams to experiment.
  • If direct free-energy learning proves unable to handle discontinuities at first-order transitions, the hybrid route of machine-learned interatomic potentials plus thermodynamic integration may remain the practical workhorse, even though the paper emphasizes direct learning.
  • Active learning on finite-temperature convex hulls could be benchmarked by recovering the experimentally known temperature-dependent stability of superionic solid electrolytes, where zero-temperature hulls are known to misjudge the stable phase.
  • Entropy estimators based on mutual information could be combined with learned structural representations to estimate entropic contributions without running long molecular dynamics, an implicit but natural extension of the paper's entropy-aware agenda.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 5 minor

Summary. This Perspective argues that conventional machine-learning models for materials, trained on zero-temperature DFT energies, neglect entropy and finite-temperature effects, and that the field should shift its target from internal energy U to free energy F, with training data encoding thermal physics. It reviews current strategies (machine-learned interatomic potentials, thermodynamic workflows, direct learning of temperature-dependent properties), identifies bottlenecks (data scarcity, poor extrapolation near phase transitions, lack of thermodynamic consistency constraints), and proposes a roadmap centered on direct free-energy learning, entropy-aware representations, and active sampling in temperature space. The examples span halide perovskites, heterogeneous catalysts, and solid-state electrolytes.

Significance. The paper addresses a real and important gap in computational materials science: most ML models indeed learn static potential-energy surfaces and cannot by construction predict finite-temperature stability. The thermodynamic identities in Eqs. (1)–(3) are correct, and the surveyed application literature (e.g., anharmonic stabilization of cubic perovskites, temperature-dependent band-gap renormalization, finite-temperature catalytic barriers, superionic conductivity) is representative and well documented. The paper also gives credit to concrete recent advances, including MLIPs for long-timescale MD, phonon-informed GNNs, and entropy-estimation methods, which strengthens its credibility as a field assessment. However, its value as a roadmap depends on the viability of the direct free-energy learning proposal, and that proposal currently has a representational gap for first-order phase transitions, which are central to several of the motivating examples. The paper is a Perspective rather than a methods paper, so this gap is not disqualifying, but it needs to be addressed for the roadmap to be defensible.

major comments (1)
  1. [§VI.B, §VI.C] The catalytic and solid-electrolyte examples are well chosen, but the connection to the proposed direct free-energy learning framework is loose: the cited works use MLIPs with thermodynamic integration, umbrella sampling, or NEB, not direct F(T) regression. This is fine for a review, but the paper should state explicitly which of the proposed roadmap elements (direct learning, entropy-aware representation, active temperature sampling) would have helped in each example, otherwise the roadmap appears disconnected from the application literature it surveys.
minor comments (5)
  1. [Fig. 1 caption] The caption contains a formatting artifact: 'T emperature' with a stray space. Please correct.
  2. [§V.A] The sentence 'A direct approach consists in learning the free energy itself as a function of temperature' could benefit from a reference to at least one concrete implementation beyond the authors' own work (e.g., thermodynamic integration with neural-network free-energy parametrizations), to show the idea is not purely programmatic.
  3. [References] Some references are incomplete or inconsistent: [55] uses initials in a nonstandard order, [90] lacks a volume/article number, and [94] is a preprint with no year of journal acceptance. Please check the bibliography against the journal's reference style.
  4. [§II.A, Eq. (3)] In Eq. (3), the zero-point term in the harmonic free-energy expression is spelled out (ℏω/2), which is clear, but the text could note that this expression assumes the quasi-harmonic volume dependence, since the QHA volume scan is referenced later.
  5. [§VII] The Outlook section lists five research directions. The first, 'Direct free-energy learning with thermodynamic constraints,' is the paper's flagship proposal, and the reader would benefit from a sentence linking it back to the phase-transition limitation raised in §III.C and §IV, so that the roadmap reads as internally coherent.

Circularity Check

0 steps flagged

No significant circularity: the paper's central claim follows from standard thermodynamic identities and its proposal is a roadmap, not a fitted prediction; author self-citations are illustrative rather than load-bearing.

full rationale

The paper's central argument is definitional: G(T,P)=U−TS+PV (Eq. 1) directly entails 'A zero-temperature energy model has no access to S and therefore cannot reproduce G(T)' (Sec. II.B). This is standard thermodynamics, not a quantity derived from the same data it predicts. Sections III–V review existing work and propose a future roadmap; there is no fitted parameter renamed as a prediction and no equation whose output is its input by construction. The authors' own works (e.g., refs. [70,71,110–113]) appear as examples of ML applications; the only place the roadmap cites an author preprint as evidence ([71] in V.A, alongside external [72]) is a literature-review statement, not the load-bearing derivation. The paper self-acknowledges its practical limitations ('training sets are typically small and narrow', 'models trained on them extrapolate unreliably ... particularly across phase transitions where the property may change discontinuously', Sec. III.C/IV). These are feasibility and representational caveats (including the smooth-function issue for first-order transitions), not circularity: they do not make the proposed outputs equivalent to the inputs. Accordingly, no circular step is identified.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper's physics backbone is standard equilibrium thermodynamics. The main nonstandard load-bearing assumptions are the additive entropy decomposition (Eq. 2), the tractability of finite-temperature training data, and the feasibility of learning smooth free-energy surfaces with thermodynamic constraints. No free parameters or invented entities appear because the paper is a review.

axioms (5)
  • standard math Stability is governed by the Gibbs free energy G=U−TS+PV, not internal energy U (Eq. 1).
    Invoked throughout; standard equilibrium thermodynamics.
  • domain assumption Entropy can be decomposed into additive vibrational, configurational, electronic, and magnetic contributions (Eq. 2).
    Used to justify learning individual free-energy contributions; additivity is an approximation when contributions are strongly coupled, which the paper acknowledges for high-entropy alloys and perovskites.
  • standard math Harmonic phonon free energy (Eq. 3) with temperature-renormalized frequencies provides a valid starting point for vibrational free energies.
    Standard formula; the paper notes harmonic approximation fails for strongly anharmonic systems.
  • domain assumption Finite-temperature training data of sufficient scope can be generated (via MLIP MD, thermodynamic integration, SCHA/TDEP) to train direct free-energy models.
    Central to the roadmap; the paper itself lists data scarcity as a key bottleneck, so this is an asserted feasibility rather than established fact.
  • ad hoc to paper Free-energy surfaces learned from structure–temperature pairs can be made smooth and thermodynamically consistent (S=−∂F/∂T, Cv≥0) and will generalize across temperatures.
    Proposed in Section V.A; no demonstration beyond cited works [71,72,88]. Distinct from the claim that zero-T models are limited.

pith-pipeline@v1.3.0-alltime-deepseek · 24771 in / 13175 out tokens · 116161 ms · 2026-08-01T00:14:12.847287+00:00 · methodology

0 comments
read the original abstract

Machine learning (ML) is transforming materials discovery by enabling rapid prediction of properties that previously required computationally expensive first-principles calculations. Yet most current ML models remain fundamentally limited to zero-temperature descriptions, learning static lattice energies while neglecting the thermodynamic effects that govern materials behaviour at finite temperature. Because phase stability, functional response, and performance are governed by free-energy landscapes rather than static energies alone, this limitation represents a major barrier to predictive materials design under realistic operating conditions. In this Perspective, we argue that developing thermodynamics-informed ML constitutes one of the most important and least explored frontiers in materials discovery. We examine the fundamental shortcomings of energy-based models, highlighting the essential roles of entropy and anharmonicity in determining free energies and materials functionality. We review emerging strategies, including machine-learned interatomic potentials and hybrid ML-statistical mechanics frameworks, while identifying key challenges related to data availability, transferability, and thermodynamic consistency. Building on these advances, we outline a roadmap for thermodynamics-informed ML centred on direct free-energy learning, entropy-aware representations, and adaptive sampling across temperature. We highlight the transformative opportunities this paradigm offers for energy materials and argue that the next generation of ML models must move beyond static energy predictions towards a thermodynamic description of materials behaviour under realistic operating conditions.

Figures

Figures reproduced from arXiv: 2607.26296 by Cibr\'an L\'opez, Claudio Cazorla, Pol Ben\'itez.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

147 extracted references · 1 linked inside Pith

  1. [1]

    Jonathan Schmidt, M´ ario R. G. Marques, Silvana Botti, and Miguel A. L. Marques. Recent advances and ap- plications of machine learning in solid-state materials science.npj Comput. Mater., 5:83, 2019

    2019

  2. [2]

    MD samplingfinite-T trajectories, large supercells, anharmonicity

  3. [3]

    Free-energy evaluationevaluate F and other quantities + consistency constraints

  4. [4]

    3.F ramework for thermodynamics-informed ML.Closed loop in which temperature is an explicit input

    Active learningselect high-uncertainty points in composition and T inputs 𝑆=−𝜕𝐹/𝜕𝑇𝐶𝑣=−𝑇𝜕2𝐹/𝜕𝑇2≥0 Thermodynamic consistency prediction Fc FIG. 3.F ramework for thermodynamics-informed ML.Closed loop in which temperature is an explicit input. (1) A physics-informed model predictsFfrom structure and temperature; (2) MD sampling generates anharmonic finite-te...

  5. [5]

    Lawrence Zitnick, and Zachary Ulissi

    Lowik Chanussot, Abhishek Das, Siddharth Goyal, Thibaut Lavril, Muhammed Shuaibi, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Aini Palizhati, Anuroop Sriram, Brandon Wood, Junwoong Yoon, Devi Parikh, C. Lawrence Zitnick, and Zachary Ulissi. Open catalyst 2020 (OC20) dataset and community challenges.ACS Catal., 11:6059–6072, 2021

    2020

  6. [6]

    The NOMAD laboratory: from data sharing to artificial intelligence

    Claudia Draxl and Matthias Scheffler. The NOMAD laboratory: from data sharing to artificial intelligence. J. Phys. Mater., 2:036001, 2019

    2019

  7. [7]

    Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk

    Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624:80–85, 2023

    2023

  8. [8]

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The Ma- terials Project: A materials genome approach to accel- erating materials innovation.APL Mater., 1:011002, 2013

    2013

  9. [9]

    Saal, Bryce Meredig, Alex Thompson, Jeff W

    Scott Kirklin, James E. Saal, Bryce Meredig, Alex Thompson, Jeff W. Doak, Muratahan Aykol, Stephan R¨ uhl, and Chris Wolverton. The open quantum mate- rials database (OQMD): assessing the accuracy of DFT formation energies.npj Comput. Mater., 1:15010, 2015

    2015

  10. [10]

    Ruban, Sergii Khmelevskyi, Peter Mohn, and B¨ orje Johansson

    Andrei V. Ruban, Sergii Khmelevskyi, Peter Mohn, and B¨ orje Johansson. Temperature-induced longitu- dinal spin fluctuations in Fe and Ni.Phys. Rev. B, 75:054402, 2007

    2007

  11. [11]

    Multiple structural transitions driven by spin-phonon couplings in a perovskite oxide.Sci

    Claudio Cazorla, Oswaldo Di´ eguez, and Jorge ´I˜ niguez. Multiple structural transitions driven by spin-phonon couplings in a perovskite oxide.Sci. Adv., 3:e1700288, 2017

    2017

  12. [12]

    The effect of lattice vibrations on substitutional alloy thermodynam- ics.Rev

    Axel van de Walle and Gerbrand Ceder. The effect of lattice vibrations on substitutional alloy thermodynam- ics.Rev. Mod. Phys., 74:11–45, 2002

    2002

  13. [13]

    Liaw, Michael C

    Rui Feng, Peter K. Liaw, Michael C. Gao, and Michael Widom. First-principles prediction of high-entropy- alloy stability.npj Comput. Mater., 3:50, 2017

    2017

  14. [14]

    David Mermin

    N. David Mermin. Thermal properties of the inhomoge- neous electron gas.Phys. Rev., 137:A1441–A1443, 1965

    1965

  15. [15]

    Unified graph neural network force-field for the peri- odic table: solid state applications.Digital Discovery, 2(2):346–355, 2023

    Kamal Choudhary, Brian DeCost, Lily Major, Keith Butler, Jeyan Thiyagalingam, and Francesca Tavazza. Unified graph neural network force-field for the peri- odic table: solid state applications.Digital Discovery, 2(2):346–355, 2023

    2023

  16. [16]

    Mace: Higher order equiv- ariant message passing neural networks for fast and ac- curate force fields.Advances in neural information pro- cessing systems, 35:11423–11436, 2022

    Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and G´ abor Cs´ anyi. Mace: Higher order equiv- ariant message passing neural networks for fast and ac- curate force fields.Advances in neural information pro- cessing systems, 35:11423–11436, 2022

    2022

  17. [17]

    Dynamical tun- ing of the thermal conductivity via magnetophononic effects.Phys

    Claudio Cazorla and Riccardo Rurali. Dynamical tun- ing of the thermal conductivity via magnetophononic effects.Phys. Rev. B, 105:104401, 2022. 15

    2022

  18. [18]

    Generalized neural- network representation of high-dimensional potential- energy surfaces.Phys

    J¨ org Behler and Michele Parrinello. Generalized neural- network representation of high-dimensional potential- energy surfaces.Phys. Rev. Lett., 98:146401, 2007

    2007

  19. [19]

    Bart´ ok, Mike C

    Albert P. Bart´ ok, Mike C. Payne, Risi Kondor, and G´ abor Cs´ anyi. Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons. Phys. Rev. Lett., 104:136403, 2010

    2010

  20. [20]

    The hiphive package for the extraction of high-order force constants by machine learning.Adv

    Fredrik Eriksson, Erik Fransson, Oliver Hellman, and Paul Erhart. The hiphive package for the extraction of high-order force constants by machine learning.Adv. Theory Simul., 2:1800184, 2019

    2019

  21. [21]

    Lattice vibrations and molecular conforma- tions in a plastic crystal: A path to solid-state cooling materials.PRX Energy, 4:033019, 2025

    Ares Sanuy, Carlos Escorihuela-Sayalero, Pol Lloveras, Josep-Llu ´ ıs Tamarit, Luis Carlos Pardo, and Claudio Cazorla. Lattice vibrations and molecular conforma- tions in a plastic crystal: A path to solid-state cooling materials.PRX Energy, 4:033019, 2025

    2025

  22. [22]

    Claudio Cazorla, Dario Alf` e, and Michael J. Gillan. Constraints on the phase diagram of molybdenum from first-principles free-energy calculations.Phys. Rev. B, 85:064113, 2012

    2012

  23. [23]

    Gillan, and Geoffrey D

    Dario Alf` e, Michael J. Gillan, and Geoffrey D. Price. Composition and temperature of the Earth’s core con- strained by combining ab initio calculations and seismic data.Earth Planet. Sci. Lett., 195:91–98, 2002

    2002

  24. [24]

    Gillan, Ian G

    Lidunka Vocˇ adlo, Dario Alf` e, Michael J. Gillan, Ian G. Wood, John P. Brodholt, and Geoffrey D. Price. Possi- ble thermal and chemical stabilization of body-centred- cubic iron in the Earth’s core.Nature, 424:536–539, 2003

    2003

  25. [25]

    Sagotra, Dewei Chu, and Claudio Cazorla

    Arun K. Sagotra, Dewei Chu, and Claudio Cazorla. Room-temperature mechanocaloric effects in lithium- based superionic materials.Nature Communications, 9:3337, 2018

    2018

  26. [26]

    Sagotra, and Claudio Cazorla

    Jie Min, Arun K. Sagotra, and Claudio Cazorla. Large barocaloric effects in thermoelectric superionic materi- als.Phys. Rev. Mater., 4:015403, 2020

    2020

  27. [27]

    Prediction and understanding of barocaloric effects in orientationally disordered materials from molecular dynamics simulations.npj Computational Materials, 10:13, 2024

    Carlos Escorihuela-Sayalero, Luis Carlos Pardo, Michela Romanini, Nicolas Obrecht, Sophie Loehl´ e, Pol Lloveras, Josep-Llu ´ ıs Tamarit, and Claudio Cazorla. Prediction and understanding of barocaloric effects in orientationally disordered materials from molecular dynamics simulations.npj Computational Materials, 10:13, 2024

    2024

  28. [28]

    Carlos Escorihuela-Sayalero, Ares Sanuy, Luis Carlos Pardo, and Claudio Cazorla. Orientational disorder and molecular correlations in hybrid organic–inorganic perovskites: From fundamental insights to technologi- cal applications.ACS Applied Materials & Interfaces, 17(1):1428–1440, 2025

    2025

  29. [29]

    Sagotra, Daniel Errandonea, and Claudio Ca- zorla

    Arun K. Sagotra, Daniel Errandonea, and Claudio Ca- zorla. Mechanocaloric effects in superionic thin films from atomistic simulations.Nature Communications, 8:963, 2017

    2017

  30. [30]

    First-principles phonon calculations of thermal expansion in Ti 3SiC2, Ti 3AlC2, and Ti 3GeC2.Phys

    Atsushi Togo, Laurent Chaput, Isao Tanaka, and Gilles Hug. First-principles phonon calculations of thermal expansion in Ti 3SiC2, Ti 3AlC2, and Ti 3GeC2.Phys. Rev. B, 81:174301, 2010

    2010

  31. [31]

    Cazorla and J

    C. Cazorla and J. Boronat. Simulation and understand- ing of atomic and molecular quantum crystals.Rev. Mod. Phys., 89:035003, 2017

    2017

  32. [32]

    Sagotra, Meng Li, Michael B

    Sheik Md Kazi Nazrul Islam, Prince Mayank, Yu- lou Ouyang, Jie Chen, Arun.K. Sagotra, Meng Li, Michael B. Cortie, Richard Mole, Claudio Cazorla, Dehong Yu, Xiaolin Wang, Robert A. Robinson, and David Laurence Cortie. Copper diffusion rates and hop- ping pathways in superionic cu2se.Acta Materialia, 215:117026, 2021

    2021

  33. [33]

    Phonons and related crystal properties from density-functional perturbation theory.Rev

    Stefano Baroni, Stefano de Gironcoli, Andrea Dal Corso, and Paolo Giannozzi. Phonons and related crystal properties from density-functional perturbation theory.Rev. Mod. Phys., 73:515–562, 2001

    2001

  34. [34]

    Parlinski, Z

    K. Parlinski, Z. Q. Li, and Y. Kawazoe. First-principles determination of the soft mode in cubic ZrO 2.Phys. Rev. Lett., 78:4063–4066, 1997

    1997

  35. [35]

    Popov, Donald M

    Florian Mayer, Maxim N. Popov, Donald M. Evans, Stephan Krohns, Marco Deluca, and J¨ urgen Spitaler. Improved description of the potential energy surface in batio3 by anharmonic phonon coupling.Phys. Rev. B, 106:064108, 2022

    2022

  36. [36]

    Giant photocaloric effects across a vast temperature range in ferroelectric perovskites

    Riccardo Rurali, Carlos Escorihuela-Sayalero, Josep Llu ´ ıs Tamarit, Jorge ´I˜ niguez Gonz´ alez, and Claudio Cazorla. Giant photocaloric effects across a vast temperature range in ferroelectric perovskites. Phys. Rev. Lett., 133:116401, 2024

    2024

  37. [37]

    Ben ´ ıtez, R

    P. Ben ´ ıtez, R. Jiang, S. Chen, C. L´ opez, J. L. Tamarit, E. Saucedo, B. Monserrat, and C. Cazorla. Band-gap tunability in anharmonic perovskite-like semiconduc- tors driven by polar electron-phonon coupling.J. Am. Chem. Soc., 147:37506–37520, 2025

    2025

  38. [38]

    Crystal structure prediction and phase stability in highly anharmonic silver-based chalcohalide antiper- ovskites.PRX Energy, 4:023002, 2025

    Pol Ben ´ ıtez, Cibr´ an L´ opez, Cong Liu, Ivan Ca˜ no, Josep- Llu ´ ıs Tamarit, Edgardo Saucedo, and Claudio Ca- zorla. Crystal structure prediction and phase stability in highly anharmonic silver-based chalcohalide antiper- ovskites.PRX Energy, 4:023002, 2025

    2025

  39. [39]

    Struc- tural phase transitions and dielectric properties of BaTio3 from a second-principles method.Phys

    Jingtong Zhang, Louis Bastogne, Xu He, Gang Tang, Yajun Zhang, Philippe Ghosez, and Jie Wang. Struc- tural phase transitions and dielectric properties of BaTio3 from a second-principles method.Phys. Rev. B, 108:134117, 2023

    2023

  40. [40]

    First-principles calculations of phonon frequencies, lifetimes, and spectral functions from weak to strong anharmonicity: the example of palladium hy- 16 drides.Phys

    Lorenzo Paulatto, Ion Errea, Matteo Calandra, and Francesco Mauri. First-principles calculations of phonon frequencies, lifetimes, and spectral functions from weak to strong anharmonicity: the example of palladium hy- 16 drides.Phys. Rev. B, 91:054304, 2015

    2015

  41. [41]

    Raffaello Bianco, Ion Errea, Lorenzo Paulatto, Mat- teo Calandra, and Francesco Mauri. Second-order structural phase transitions, free energy curvature, and temperature-dependent anharmonic phonons in the self-consistent harmonic approximation: theory and stochastic implementation.Phys. Rev. B, 96:014111, 2017

    2017

  42. [42]

    V. Tuli, P. Burr, A. Claisse, and C. Cazorla. Thermody- namic stability ofβ-phases in Zr-Nb alloys.Phys. Rev. Mater., 7:113607, 2023

    2023

  43. [43]

    Dy- naPhoPy: A code for extracting phonon quasiparticles from molecular dynamics simulations.Comput

    Abel Carreras, Atsushi Togo, and Isao Tanaka. Dy- naPhoPy: A code for extracting phonon quasiparticles from molecular dynamics simulations.Comput. Phys. Commun., 221:221–234, 2017

    2017

  44. [44]

    An- harmonic free energies and phonon dispersions from the stochastic self-consistent harmonic approximation: ap- plication to platinum and palladium hydrides.Phys

    Ion Errea, Matteo Calandra, and Francesco Mauri. An- harmonic free energies and phonon dispersions from the stochastic self-consistent harmonic approximation: ap- plication to platinum and palladium hydrides.Phys. Rev. B, 89:064302, 2014

    2014

  45. [45]

    Grossman

    Tian Xie and Jeffrey C. Grossman. Crystal graph con- volutional neural networks for an accurate and inter- pretable prediction of material properties.Phys. Rev. Lett., 120:145301, 2018

    2018

  46. [46]

    Graph networks as a universal machine learning framework for molecules and crystals.Chem

    Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals.Chem. Mater., 31:3564–3572, 2019

    2019

  47. [47]

    The stochastic self-consistent harmonic approximation: cal- culating vibrational properties of materials with full quantum and anharmonic effects.J

    Lorenzo Monacelli, Raffaello Bianco, Marco Cherubini, Matteo Calandra, Ion Errea, and Francesco Mauri. The stochastic self-consistent harmonic approximation: cal- culating vibrational properties of materials with full quantum and anharmonic effects.J. Phys.: Condens. Matter, 33:363001, 2021

    2021

  48. [48]

    Christopher J. Bartel. Review of computational ap- proaches to predict the thermodynamic stability of in- organic solids.Journal of Materials Science, 57:10475– 10498, 2022

    2022

  49. [49]

    Machine-learning-assisted construction of ternary convex hull diagrams.Journal of Chemical Information and Modeling, 64:1828–1840, 2024

    Hugo Rossignol, Michail Minotakis, Matteo Cobelli, and Stefano Sanvito. Machine-learning-assisted construction of ternary convex hull diagrams.Journal of Chemical Information and Modeling, 64:1828–1840, 2024

    2024

  50. [50]

    A. P. Thompson, L. P. Swiler, C. R. Trott, S. M. Foiles, and G. J. Tucker. Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials.Journal of Computational Physics, 285:316– 330, 2015

    2015

  51. [51]

    Alexander V. Shapeev. Moment tensor potentials: A class of systematically improvable interatomic poten- tials.Multiscale Modeling & Simulation, 14(3):1153– 1173, 2016

    2016

  52. [52]

    L´ opez, I

    C. L´ opez, I. Ca˜ no, D. Rovira, P. Ben ´ ıtez, J. M. Asensi, Z. Jehl, J.-L. Tamarit, E. Saucedo, and C. Ca- zorla. Machine-learning aided first-principles prediction of earth-abundant pnictogen chalcohalide solid solutions for solar-cell devices.Adv. Funct. Mater., 34:2406678, 2024

    2024

  53. [53]

    Frost, Jonathan M

    Federico Brivio, Jarvist M. Frost, Jonathan M. Skelton, Adam J. Jackson, Oliver J. Weber, Mark T. Weller, Alejandro R. Go˜ ni, Aur´ elien M. A. Leguy, Piers R. F. Barnes, and Aron Walsh. Lattice dynamics and vibra- tional spectra of the orthorhombic, tetragonal, and cu- bic phases of methylammonium lead iodide.Phys. Rev. B, 92:144308, 2015

    2015

  54. [54]

    Patrick, Karsten W

    Christopher E. Patrick, Karsten W. Jacobsen, and Kris- tian S. Thygesen. Anharmonic stabilization and band gap renormalization in the perovskite CsSnI 3.Phys. Rev. B, 92:201205, 2015

    2015

  55. [55]

    A. P. A. Subramanyam and D. Perez. Information- entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials.npj Com- putational Materials, 11:218, 2025

    2025

  56. [56]

    K. Li, K. Choudhary, B. DeCost, M. Greenwood, and J. Hattrick-Simpers. Efficient first principles based mod- eling via machine learning: from simple representa- tions to high entropy materials.J. Mater. Chem. A, 12:12412–12422, 2024

    2024

  57. [57]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, Kornbluth. M., N. Molinari, T. E. Smidt, and B. Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nat Commun., 4:2453, 2022

    2022

  58. [58]

    A foundation model for atomistic materials chemistry.arXiv, 2401.00096, 2023

    Ilyes Batatia et al. A foundation model for atomistic materials chemistry.arXiv, 2401.00096, 2023

    Pith/arXiv arXiv 2023

  59. [59]

    Bartel, and Ger- brand Ceder

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Ger- brand Ceder. CHGNet as a pretrained universal neural network potential for charge-informed atomistic mod- elling.Nat. Mach. Intell., 5:1031–1041, 2023

    2023

  60. [60]

    learn on the fly

    G. Cs´ anyi, T. Albaret, M. C. Payne, and A. De Vita. “learn on the fly”: A hybrid classical and quantum- mechanical molecular dynamics simulation.Phys. Rev. Lett., 93:175503, 2004

    2004

  61. [61]

    V. Botu, R. Batra, J. Chapman, and R. Ramprasad. Machine learning force fields: Construction, validation, and outlook.The Journal of Physical Chemistry C, 121:511–522, 2017

    2017

  62. [62]

    Assessment and appli- cation of universal machine learning interatomic poten- tials in solid-state electrolyte research.ACS Materials Letters, 7(10):3403–3412, 2025

    Hongwei Du, Xiang Huang, Jian Hui, Lanting Zhang, Yuanxun Zhou, and Hong Wang. Assessment and appli- cation of universal machine learning interatomic poten- tials in solid-state electrolyte research.ACS Materials Letters, 7(10):3403–3412, 2025

    2025

  63. [63]

    E. V. Podryabinkin and A. V. Shapeev. Active learning of linearly parametrized interatomic potentials.Com- putational Materials Science, 140:171–180, 2017

    2017

  64. [64]

    S. B. Torrisi, S. Batzner, Y. Xie, L. Sun, A. M. Kolpak, and B. Kozinsky. On-the-fly active learning of inter- pretable bayesian force fields for atomistic rare events. npj Computational Materials, 6:20, 2020

    2020

  65. [65]

    Al-Fahdi, C

    M. Al-Fahdi, C. Lin, C. Shen, H. Zhang, and M. Hu. Rapid prediction of phonon density of states by crys- tal attention graph neural network and high-throughput screening of candidate substrates for wide bandgap elec- tronic cooling.Materials Today Physics, 50:101632, 2025

    2025

  66. [66]

    J. Ojih, M. Al-Fahdi, Y. Yao, J. Hu, and M. Hu. Graph theory and graph neural network assisted high- throughput crystal structure prediction and screening for energy conversion and storage.J. Mater. Chem. A, 12:8502–8515, 2024

    2024

  67. [67]

    Okabe, A

    R. Okabe, A. Chotrattanapituk, A. Boonkird, N. An- drejevic, X. Fu, T. S. Jaakkola, Q. Song, T. Nguyen, N. Drucker, S. Mu, Y. Wang, B. Liao, Y. Cheng, and Mingda Li. Virtual node graph neural network for full phonon prediction.Nature Computational Science, 4:522–531, 2024

    2024

  68. [68]

    H. Lee, V. I. Hegde, C. Wolverton, and Y. Xia. Acceler- ating high-throughput phonon calculations via machine learning universal potentials.Materials Today Physics, 53:101688, 2025

    2025

  69. [69]

    Abrikosov

    Olle Hellman and Igor A. Abrikosov. Temperature- dependent effective third-order interatomic force con- stants from first principles.Phys. Rev. B, 88:144301, 2013

    2013

  70. [70]

    Ben ´ ıtez, C

    P. Ben ´ ıtez, C. L´ opez, E. Saucedo, T. Mizoguchi, and C. Cazorla. Why physics still matters: improving ma- chine learning prediction of material properties with phonon-informed datasets.Advanced Intelligent Discov- ery, 2:e202500229, 2025

    2025

  71. [71]

    L´ opez, J

    C. L´ opez, J. Ojih, M. Hu, J. L. Tamarit, E. Saucedo, and C. Cazorla. Machine learning-guided discovery of temperature-induced solid-solid phase transitions in in- organic materials.arXiv, 2025

    2025

  72. [72]

    M. Hu. Unleashing the power of artificial intelligence in phonon thermal transport: Current challenges and prospects.Journal of Applied Physics, 135:170904, 2024. 17

    2024

  73. [73]

    Accelerating the calculation of electron–phonon coupling strength with machine learn- ing.Nat

    Yang Zhong, Shixu Liu, Binhua Zhang, Zhiguo Tao, Yuting Sun, Weibin Chu, Xin-Gao Gong, Ji-Hui Yang, and Hongjun Xiang. Accelerating the calculation of electron–phonon coupling strength with machine learn- ing.Nat. Comput. Sci., 4:615–625, 2024

    2024

  74. [74]

    Machine learning approach for vibronically renormalized electronic band structures.Phys

    Niraj Aryal, Sheng Zhang, Weiguo Yin, and Gia-Wei Chern. Machine learning approach for vibronically renormalized electronic band structures.Phys. Rev. B, 113:085127, 2026

    2026

  75. [75]

    Kaufmann, D

    K. Kaufmann, D. Maryanovsky, W. M. Mellor, C. Zhu, A. S. Rosengarten, T. J. Harrington, C. Oses, C. Toher, S. Curtarolo, and K. S. Vecchio. Discovery of high- entropy ceramics via machine learning.npj Computa- tional Materials, 6:42, 2020

    2020

  76. [76]

    Frueh, C

    C. Frueh, C. Aras, ¨O. B¨ uy¨ ukuslu, and M. to Baben. aimp and aioq databases in factsage: Materials infor- matics relying on ab initio, machine learning and cal- phad data.Calphad, 90:102838, 2025

    2025

  77. [77]

    Forslund, J

    A. Forslund, J. H. Jung, Y. Ikeda, and B. Grabowski. Free-energy perturbation in the exchange-correlation space accelerated by machine learning: application to silica polymorphs.npj Computational Materials, 12:14, 2025

    2025

  78. [78]

    Herzog, A

    B. Herzog, A. Gallo, F. Hummel, M. Badawi, T. Buˇ cko, S. Leb` egue, A. Gr¨ uneis, and D. Rocca. Coupled cluster finite temperature simulations of periodic materials via machine learning.npj Computational Materials, 10:68, 2024

    2024

  79. [79]

    Chang, Y.-X

    R. Chang, Y.-X. Wang, and E. Ertekin. Towards over- coming data scarcity in materials science: unifying mod- els and datasets with a mixture of experts framework. npj Computational Materials, 8:242, 2022

    2022

  80. [80]

    A. M. Deml, R. O’Hayre, C. Wolverton, and V. Ste- vanovi´ c. Predicting density functional theory total ener- gies and enthalpies of formation of metal-nonmetal com- pounds by linear regression.Phys. Rev. B, 93:085142, 2016

    2016

Showing first 80 references.