Machine Learning-Guided Discovery of Temperature-Induced Solid-Solid Phase Transitions in Inorganic Materials

Cibr\'an L\'opez; Claudio Cazorla; Edgardo Saucedo; Josep Lluis Tamarit; Joshua Ojih; Ming Hu

REVIEW 3 major objections 4 minor 66 references

Machine Learning-Guided Discovery of Temperature-Induced Solid-Solid Phase Transitions in Inorganic Materials

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A machine-learning screen of roughly 50,000 inorganic crystals predicts 2,079 solid-solid phase transitions, many near room temperature with entropy changes large enough for cooling applications.

desk verdict Useful candidate pool, but the screen never checks that the predicted polymorphs are dynamically stable, so treat the 2,079 transitions as hypotheses, not predictions. read the letter →

arxiv 2506.01449 v1 pith:5DWWFLAP submitted 2025-06-02 cond-mat.mtrl-sci

Cibrán López , Joshua Ojih , Ming Hu , Josep Lluis Tamarit , Edgardo Saucedo , Claudio Cazorla This is my paper

classification cond-mat.mtrl-sci PACS 64.70.Kb63.20.-e71.15.Mb

keywords solid-solidphasetransitionsmachinelearningvibrationalfreeenergyquasi-harmonicapproximationgraphneuralnetworkshigh-throughputscreeningsolid-statecoolingthermalswitching

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a machine-learning model trained on 6,674 first-principles phonon spectra can estimate vibrational free energies accurately enough to screen roughly 50,000 inorganic crystals and predict where temperature alone drives one crystal structure into another. Applying this screen, the authors report 2,079 polymorphic phase transitions with transition temperatures between 300 and 600 K, of which 615 are stable and 1,464 metastable. The practical payoff is a large, chemically diverse menu of candidate phase-change materials: transitions with entropy changes above 300 joules per kelvin and per kilogram near room temperature, promising for solid-state cooling, and 21 transitions with 20–70% changes in lattice thermal conductivity, promising for thermal switches. Agreement with three experimentally known transitions and with full DFT checks on a subset is offered as evidence that the predictions are physically meaningful.

What carries the argument

The load-bearing object is the machine-learned vibrational free energy $F_v^{\mathrm{ML}}(T) \approx \alpha + \beta T^2 + \gamma T^4$, produced by a graph convolutional neural network — a message-passing network on the crystal graph, with atoms as nodes and bonds as edges — trained exclusively on 6,674 DFT phonon spectra whose frequencies are all strictly positive. Quasi-harmonic theory supplies the physical scaffold: $F(V,T) = E_0(V) + F_v(V,T)$ with $F_v$ built from the phonon mode sum, and a transition occurs where the free energies of two polymorphs cross at equal pressure. The neural network makes the phonon sum unnecessary at screening scale; the polynomial smoothing keeps the free energy differentiable so entropy $S_v = -\partial F_v/\partial T$ and transition temperatures can be extracted; and a latent-space distance (the Euclidean distance between the pooled graph embedding of a new structure and the training-set embeddings) provides the uncertainty estimate used to discard unreliable predictions. A second surrogate model trained on about 4,700 DFT lattice thermal conductivity values supplies the thermal-switching estimates.

What would settle it

Compute the zero-temperature phonon spectrum of the polymorphs involved in the predicted transitions, starting with the 50 largest-entropy cases: if any of them has an imaginary (negative-frequency) vibration, its quasi-harmonic free energy does not exist and that transition prediction collapses. The same check on a random sample of the full set of screened crystals would settle whether the count of 2,079 is trustworthy.

Watch

Extended reading notes

Core claim

The central claim is that temperature-induced solid-solid phase transitions can be discovered in bulk by replacing costly quasi-harmonic phonon calculations with a graph neural network: given a crystal structure, the network predicts its vibrational free energy $F_v^{\mathrm{ML}}(T)$ across 200–700 K, and adding the database's static DFT energy gives $F(V,T) \approx E_{\mathrm{DFT}} + F_v^{\mathrm{ML}}$. Free-energy crossings between polymorphs of the same compound, $F_A(V_A,T_t) = F_B(V_B,T_t)$ at zero pressure, then define the transition temperature $T_t$. On a curated pool of about 50,000 nonmetallic, nonmagnetic compounds of earth-abundant elements, the screen yields 2,079 transitions in the 300–600 K window, 615 stable and 1,464 metastable, each with an assigned $T_t$ and entropy change $\Delta S_t$, after discarding predictions whose uncertainty exceeds 50% of the vibrational free-energy difference. The paper further claims that hundreds of these transitions involve polar (non-centrosymmetric) phases near room temperature; that several entropy changes exceed 300 joules per kelvin and per kilogram, up to 360.7 for PNF2; and that 21 stable polar-nonpolar transitions show relative lattice thermal conductivity changes of 20–70%, identifying them as switchable heat conductors.

Load-bearing premise

The entire screen assumes that every crystal it evaluates is vibrationally stable at absolute zero, because the free-energy formula and the trained model are only defined for such crystals — yet the dynamic stability of the roughly 50,000 screened structures is never checked, only that of the 6,674 training spectra.

Editorial extensions

If this is right

If the screen holds, roughly 2,000 inorganic compounds become concrete experimental targets for phase-change behavior at 300–600 K, most of them chemically complex, low-symmetry oxides that routine quasi-harmonic DFT would be too expensive to survey.
The near-room-temperature entropy changes predicted for Li3MnF7, PNCl2, Fe4OF7, VOF3, and Mg2TiO4 (all above 200 joules per kelvin and per kilogram) make these five compounds specific first candidates for caloric cooling tests.
The polar-nonpolar transitions with predicted thermal-conductivity ratios of about 1.7–1.9 (ZrSeO, Bi2W2O9, and CoO2) identify specific compounds to test as electric-field-controlled heat switches.
Because the screen also reports 1,464 metastable transitions, it implies a large supply of transformations inducible by external fields — pressure or electric bias — in addition to temperature alone.
The excess of transitions involving polar phases over purely nonpolar ones suggests that polarity is a common driver of near-room-temperature polymorphic change, not just an occasional feature.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the same trained free-energy model would apply to any future structure added to the open crystal databases, so the method's reach should grow as the databases grow; the bottleneck would shift to checking that new candidate structures are dynamically stable at 0 K, a check this screen omits.
Beyond the paper: the 50%-of-free-energy-difference uncertainty cut is a heuristic; the paper does not report how the count of 2,079 varies as the threshold is raised or lowered, so the list's precision remains untested against that knob.
Beyond the paper: several predicted transitions run from a nonpolar low-temperature phase to a polar high-temperature phase, the reverse of typical ferroelectrics; if confirmed, this would define a class of electrically switchable materials whose polar state is the hot one, changing how caloric cycles would be driven.
Beyond the paper: combining the two surrogate models on the same transitions would allow ranking candidates by entropy change times conductivity contrast — a dual-function cooling-and-switching metric the paper does not compute.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a serious screening paper, not a proof. The new thing is a graph CNN trained on 6,674 DFT phonon spectra to predict vibrational free energies, then applied to roughly 50,000 Materials Project polymorphs to find free-energy crossings in the 300–600 K window. That yields 2,079 candidate transitions with Tt and ΔSt, plus 21 thermal-switching candidates. As a curated candidate pool for caloric cooling and thermal switching, it's useful and new. The held-out MgS and NaH tests look credible, and the authors were honest about several limitations (no thermal expansion, well-defined crystals only, dataset scope). Data are promised on GitHub.

The soft spots are real and need attention. The biggest: the quasi-harmonic free energy is only defined for dynamically stable crystals with positive phonon frequencies, and the ML model was trained only on such spectra. Yet the screen of about 50,000 polymorphs never applies a stability filter. For any screened structure with imaginary modes at 0 K, F_ML_v is an extrapolation, not a physical free energy. The latent-space UQ measures distance to training data, not phonon stability. So the 2,079 count very likely includes a chunk of transitions where one or both sides are dynamically unstable. That makes the abstract's "robustness and predictive power" wording an overstatement.

Second, the experimental validation is thin. KNO3: predicted 539 K vs observed 350–400 K; KNO2: predicted 529 K vs 313 K; CaCO3: 370 vs 336 K. Two of three are off by roughly 40–70%. The authors call these "reasonably close" — I'd call them order-of-magnitude okay but not strong validation. The 10 DFT validation cases are selected to exclude materials with imaginary phonons, so they don't test the risky part of the screen.

Third, the reported entropy changes come from derivatives of a polynomial fitted to the surrogate's discrete outputs, with no uncertainty propagation. The ΔSt values, some above 300 J K−1 kg−1, likely carry large error bars that aren't shown.

None of this kills the paper. The pipeline is clearly described and the candidate list is a legitimate resource. But the claims need to be scaled back until a stability filter (even a cheap ML one) is applied, or at least until the authors report the fraction of predicted transitions with stable polymorphs. I'd send it to review, with majors: add the stability check, soften the abstract, and report uncertainties.

Who's it for: people doing high-throughput screening for caloric materials or thermal switches will want the list. I wouldn't rely on any single predicted transition without checking it. For peer review: yes, as a screening paper.

Referee Report

3 major / 4 minor

Summary. The manuscript presents a machine-learning-guided high-throughput framework for predicting temperature-induced solid-solid phase transitions in inorganic crystals. The authors combine static DFT energies from the Materials Project with a graph neural network (GCNN) trained on 6,674 phonon spectra to estimate vibrational free energies, then screen roughly 50,000 nonmetallic, nonmagnetic, earth-abundant compounds. They report 2,079 polymorphic transformations in the 300–600 K interval, classify them as stable or metastable, and further identify 21 compounds with large relative changes in lattice thermal conductivity, proposing them as thermal switching materials. Validation includes held-out MgS and NaH phonon tests, DFT checks on ten transitions, and comparisons with experimental transition temperatures for KNO3, KNO2, and CaCO3.

Significance. If the central screening results are correct, the paper provides a large, chemically diverse candidate pool for caloric cooling and thermal switching, which would be a valuable resource for experimental materials discovery. The pipeline is clearly specified, the ML model is tested on held-out polymorphs, and the LTC surrogate is validated against DFT for two representative compounds. These strengths make the framework potentially useful even if the specific list of 2,079 transitions requires qualification. However, the main claim is load-bearing on an unverified assumption about dynamical stability of the screened polymorphs, and the experimental validations show large quantitative errors, so the current version does not justify the word 'confidently' attached to the 2,079 predictions.

major comments (3)

[Sec. II A and Sec. II B / III A] The quasi-harmonic free energy in Eq. (2) is defined only for crystals with strictly positive phonon frequencies, as stated in Sec. II A. The ML model was trained only on 6,674 spectra without imaginary frequencies (Sec. II C), yet the screen in Sec. II B applies it to about 50,000 polymorphs without a dynamical-stability filter. For dynamically unstable structures, F_ML_v is an extrapolation and the subsequent Tt and ΔSt are not physically meaningful. The Sec. IV note that DFT validation was limited to materials 'that do not exhibit imaginary phonon frequencies at zero temperature' confirms that this condition was not verified for the full set. This is a load-bearing issue for the 2,079-transition claim; the authors should add a stability filter or explicitly recast the predictions as conditional on dynamical stability and quantify the expected stability rate in the screened pool.
[Sec. IV] The experimental comparison is less supportive than the text suggests. For KNO3, the predicted Tt of 539 K deviates from the experimentally reported 350–400 K range by roughly 35–54%; for KNO2, the predicted 529 K is about 69% above the reported 313 K; only CaCO3 (370 K vs 336 K, about 10% error) is close. Calling these agreements 'reasonably close' and 'very good' overstates the accuracy. Given that the paper uses these three systems to 'support the reliability' of the entire 2,079-entry database, the authors should report quantitative error metrics, discuss the likely sources of systematic error (neglected thermal expansion, functional mismatch, ML extrapolation), and temper the claim of confidence accordingly.
[Sec. II C, Eq. (5), and Tables 1–5] The reported entropy changes ΔSt are obtained as derivatives of a polynomial of the form F_ML_v(T) ≈ α + βT² + γT⁴ fitted to ML-predicted vibrational free energies. The ML model has a test RMSE of 12.31 meV/atom, but no uncertainty is propagated to ΔSt, and the latent-distance UQ of Sec. II D measures a distance to training data rather than derivative accuracy. Since ΔSt values up to 342 J K⁻¹ kg⁻¹ are used to shortlist materials for caloric applications, the sensitivity of ΔSt to the polynomial fit and to the ML error should be quantified; otherwise the reported entropy changes carry unknown error bars.

minor comments (4)

[Sec. III A] The filtering criterion that discards transitions where vibrational free-energy uncertainties exceed 50% of |ΔF_ML_v| = |F_ML_v(600 K) − F_ML_v(300 K)| is defined on the temperature variation of each polymorph's F_v, not on the difference of free energies between the two polymorphs. The connection between this threshold and the error in the derived transition temperature should be clarified.
[Sec. II B and Methods] The static energies E_MP_DFT are taken from the Materials Project, whose DFT calculations typically use the PBE functional, while the validation DFT calculations in Methods use PBEsol. The functional consistency between the static energies and the phonon/training dataset should be stated explicitly, as combining different XC functionals can introduce systematic errors in free-energy differences.
[Sec. II C, Eq. (5)] The polynomial form F_ML_v(T) ≈ α + βT² + γT⁴ is described as physically motivated, but the authors do not state whether α, β, γ are constrained to reproduce the correct low-temperature behavior (e.g., the linear-in-T heat capacity limit and the vanishing derivative of F_v at T = 0). Please provide the constraints or a note on why the unconstrained form is adequate over 200–700 K.
[Tables and abstract] The abstract says 'over 2,000' while Sec. III A states exactly 2,079; please use a consistent number. Also, Table 6 lists 'Bi2W2O9' but the Conclusions refer to 'Bi4W2O9'; check the stoichiometry for consistency.

Circularity Check

1 steps flagged · score 1.0 of 10

No circular reduction: the 2,079 predicted transitions are out-of-sample outputs of a surrogate trained on intermediate vibrational free energies, benchmarked against fresh DFT and experiments; self-citations are data provenance, and the imaginary-mode gap is a validity risk, not a tautology.

other [Sec. III C (Application 2: Thermal switching) and Methods, "LTC subsidiary ML model"]
"by employing a surrogate ML model capable of predicting a material's lattice thermal conductivity (LTC) based solely on its composition and atomic structure [42–45] ... In total, ≈ 4,700 LTC data calculated by full DFT calculations were used as end property for ALIGNN model training, which were accumulated in our recent works [42–45]."

Flagged self-citation: the 21 thermal-switch candidates and all Δλt/λ1 values in Table 6 are produced by an ALIGNN surrogate whose training data were accumulated in the authors' own prior works (refs. [43,44] share co-authors Ojih and Hu with this paper; [42,45] share Hu). This is the closest the paper comes to a load-bearing self-citation chain, because the section's headline numbers inherit the surrogate's accuracy. On inspection it is not circular: the surrogate's training labels are first-principles Boltzmann-transport conductivities, not the reported Δλ ratios; the predictions are out-of-sample extrapolations; and two fresh DFT checks (Li4TiS4, NaNO3) are provided. Hence minor non-load-bearing self-citation, not reduction-by-construction.

full rationale

The claimed 2,079 transitions are emergent outputs, not fitting targets. The pipeline is F(V,T) ≈ E_MP_DFT(V) + F_ML_v(V,T) (Eq. 4); Tt is the crossing FA = FB (Eq. 3); ΔSt is obtained from the temperature derivative of the polynomial smoothing (Eq. 5) fitted to F_ML_v output. F_ML_v is a GCNN trained on DFT vibrational free energies (Eq. 2) for 6,674 phonon spectra, so the training labels are intermediate quantities (Fv curves); neither the 2,079 Tt values nor any reported ΔSt appears in the training set. The predictions therefore do not reduce to the inputs by construction: there is no self-definitional equality and no parameter fitted to the target quantity. This is machine-learned prediction rather than a purely first-principles derivation, but that is not circularity. External benchmarks keep the derivation self-contained: experimental checks (KNO3, KNO2, CaCO3 in Sec. IV), fresh in-paper DFT checks (MgS and NaH polymorphs, Fig. 3; ten DFT QHA transition validations; two BTE LTC checks with ϵDFT = 2.23 vs 1.43 for Li4TiS4 and 1.40 vs 1.47 for NaNO3). Self-citations exist — the phonon dataset cites ref. [15] (co-author Hu) and the LTC surrogate relies on refs. [42–45] (co-authors Ojih and/or Hu) — but they supply first-principles ground-truth data and are independently re-validated inside the paper, so they are provenance rather than load-bearing circular argument. The acknowledged limitations (Sec. II A: "the QH approach is applicable only to vibrationally stable systems with strictly positive phonon frequencies"; Sec. II C: training set contains only "well-behaved full phonon spectra (i.e., without imaginary frequencies)"; Sec. IV: validation restricted to "materials that do not exhibit imaginary phonon frequencies") expose a genuine validity risk — the ~50,000 screened polymorphs are never filtered for 0 K dynamical stability, so some predicted crossings may rest on ill-defined free energies. That is an accuracy/validity error, not a tautology, and per the review rules belongs under correctness risk rather than circularity. Likewise, the loose agreement for KNO3 (predicted 539 K vs observed 350–400 K) and KNO2 (predicted 529 K vs observed 313 K) is a calibration concern, not evidence of circular reasoning. No uniqueness theorem or ansatz is imported from self-citations. Verdict: no significant circularity; score 1 reflects the minor non-load-bearing self-citation in the thermal-switching surrogate chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The predictions rest on a fitted ML free-energy surrogate and on several approximations that are stated but not fully verified: quasi-harmonic validity, no thermal expansion, no electronic or magnetic entropy, consistent Materials Project static energies, and a heuristic latent-distance uncertainty metric. The ML weights and per-material polynomial coefficients are the main fitted quantities; no new physical entities are introduced.

free parameters (3)

GCNN vibrational free-energy model weights = not disclosed (trained on 6,674 DFT phonon spectra)
The central Fv predictions, and hence all transition temperatures and entropy changes, depend on these fitted weights. Only architecture and hyperparameters are reported.
Polynomial smoothing coefficients alpha, beta, gamma per material = not reported
Eq. 5 fits predicted Fv(T) points to alpha + beta T^2 + gamma T^4; the reported Delta-St values are derivatives of these fitted curves.
ALIGNN lattice thermal conductivity model weights = not disclosed (trained on about 4,700 DFT LTC values)
The 21 thermal-switching candidates are selected from ALIGNN predictions, and no per-candidate uncertainty or test-set error is reported.

assumptions (5)

domain assumption The quasi-harmonic approximation is accurate for nonmetallic, nonmagnetic crystals with positive phonon frequencies.
Used in Eqs. (1)-(3); the paper restricts the screened set to nonmetallic, nonmagnetic compounds but does not verify dynamical stability of every target polymorph.
domain assumption Materials Project static DFT energies are accurate and mutually consistent across polymorphs of the same compound.
Eq. 4 uses E_MP_DFT directly; any errors or protocol differences between polymorph entries would shift the free-energy crossings and Tt values.
domain assumption Thermal expansion and anharmonicity can be neglected; each polymorph stays at its 0 K relaxed volume at all temperatures.
Stated in Sec. II A as neglecting thermal expansion effects, and acknowledged as a limitation in Sec. IV.
domain assumption Electronic and magnetic entropy contributions are negligible.
The screening excludes metallic and magnetic materials, stated in Sec. II B.
ad hoc to paper Latent-space Euclidean distance in the GCNN hyper-representation is a valid uncertainty proxy.
Defined in Sec. II D; no calibration against true model errors is provided, yet it is used to filter out low-confidence transitions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine Learning-Guided Discovery of Temperature-Induced Solid-Solid Phase Transitions in Inorganic Materials." pith.science (2026). https://pith.science/paper/5DWWFLAP

@misc{pith2026250601449,
  author       = {Pith},
  title        = {Pith review of: Machine Learning-Guided Discovery of Temperature-Induced Solid-Solid Phase Transitions in Inorganic Materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DWWFLAP}},
  note         = {Machine review of arXiv:2506.01449}
}

abstract

Predicting solid-solid phase transitions remains a long-standing challenge in materials science. Solid-solid transformations underpin a wide range of functional properties critical to energy conversion, information storage, and thermal management technologies. However, their prediction is computationally intensive due to the need to account for finite-temperature effects. Here, we present an uncertainty-aware machine-learning-guided framework for high-throughput prediction of temperature-induced polymorphic phase transitions in inorganic crystals. By combining density functional theory calculations with graph-based neural networks trained to estimate vibrational free energies, we screened a curated dataset of approximately 50,000 inorganic compounds and identified over 2,000 potential solid-solid transitions within the technologically relevant temperature interval 300-600 K. Among our key findings, we uncover numerous phase transitions exhibiting large entropy changes (> 300 J K$^{-1}$ kg$^{-1}$), many of which occur near room temperature hence offering strong potential for solid-state cooling applications. We also identify $21$ compounds that exhibit substantial relative changes in lattice thermal conductivity (20-70%) across a phase transition, highlighting them as promising thermal switching materials. Validation against experimental observations and first-principles calculations supports the robustness and predictive power of our approach. Overall, this work establishes a scalable route to discover functional phase-change materials under realistic thermal conditions, and lays the foundation for future high-throughput studies leveraging generative models and expanding open-access materials databases.

Figures

Figures reproduced from arXiv: 2506.01449 by the authors.

**Figure 2.** FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: a displays the total number of predicted stable phase transformations, NP T , classified according to their transition temperature and entropy change. Regardless of temperature, higher values of NP T are generally correlated with smaller phase-transition entropy chang…

**Figure 7.** Figure 7: a summarizes the outcomes of our ML-guided high-throughput screening of LTC, showing the number of phase transitions, NP T , as a function of transition temperature and relative LTC change (in absolute value), |∆λt|/λ1 (with λ1 and λ2 denoting the LTC of the low- and…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 65 canonical work pages

[1]

The theory of phase transitions

Landau, L. The theory of phase transitions. Nature 138, 860 (1936)

work page 1936
[2]

and Hemley, R

Mao, H.-K. and Hemley, R. J. Ultrahigh-pressure transi- tions in solid hydrogen. Rev. Mod. Phys. 66, 671 (1994)

work page 1994
[3]

Composition-induced phase transitions and enhanced electrical properties in bismuth sodium titanate ceramics

Yin J., Zhao C., Zhang Y., Wu J. Composition-induced phase transitions and enhanced electrical properties in bismuth sodium titanate ceramics. J. Am. Ceram. Soc. 100, 5601 (2017)

work page 2017
[4]

and Cazorla, C

Rurali, R, Escorihuela-Sayalero, C., Tamarit, J.-Ll., ´I˜ niguez-Gonz´ alez, J. and Cazorla, C. Giant photocaloric effects across a vast temperature range in ferroelectric perovskites. Phys. Rev. Lett. 113, 116401 (2024)

work page 2024
[5]

A., Kresse, G

Wijs, G. A., Kresse, G. and Gillan, M. J. First-order phase transitions by first-principles free-energy calcula- tions: The melting of Al. Phys. Rev. B 57, 8223 (1998)

work page 1998
[6]

The Monte Carlo method for the study of phase transitions: A review of some recent progress

Binder, K. The Monte Carlo method for the study of phase transitions: A review of some recent progress. J. Comput. Phys. 59, 1 (1985)

work page 1985
[7]

and Gillan, M

Cazorla, C., Alf´ e, D. and Gillan, M. J. Constraints on the phase diagram of molybdenum from first-principles free-energy calculations. Phys. Rev. B 85, 064113 (2012)

work page 2012
[8]

B., LeBacq, O., Ahuja, R

Belonoshko, A. B., LeBacq, O., Ahuja, R. and Johansson, B. Molecular dynamics study of phase transitions in Xe. J. Chem. Phys. 117, 7233 (2002)

work page 2002

Show all 66 references

[9]

and Boronat, J., Simulation and understand- ing of atomic and molecular quantum crystals, Rev

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

2017
[10]

and Isaev, E., Density- functional perturbation theory for quasi-harmonic calcu- lations, Rev

Baroni, S., Giannozzi, P. and Isaev, E., Density- functional perturbation theory for quasi-harmonic calcu- lations, Rev. Mineral. Geochem. 71, 39 (2010)

2010
[11]

and Cazorla, C

Tuli, V., Burr, P., Claisse, A. and Cazorla, C. Thermo- dynamic stability of β-phases in Zr-Nb alloys. Phys. Rev. Mater. 7, 113607 (2023)

2023
[12]

P., Hautier, G

Jain, A., Ong, S. P., Hautier, G. et al. Commentary: The materials project: A materials genome approach to accelerating materials innovation. Apl. Mater. 1, 011002 (2013)

2013
[13]

Reiser, P., Neubert, M., Eberhard, A. et al. Graph neural networks for materials science and chemistry. Commun. Mater. 3, 93 (2022)

2022
[14]

and Tanaka, I

Togo, A., Chaput, L., Tadano, T. and Tanaka, I. Imple- mentation strategies in phonopy and phono3py. J. Phys. Condens. Matter 35, 353001 (2023)

2023
[15]

and Hu, M

Al-Fahdi, M., Lin, C., Shen, C., Zhang, H. and Hu, M. 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. Mater. Today Phys. 50, 101632 (2025)

2025
[16]

and Waegeman, W

H¨ ullermeier, E. and Waegeman, W. Aleatoric and epis- temic uncertainty in machine learning: an introduction to concepts and methods. Mach. Learn. 110, 457 (2021)

2021
[17]

P., Yang, X., Gautam, A., Gupta, G., Li, Z., Jacobs, R., and Mor- gan, D

Palmer, G., Du, S., Politowicz, A., Emory, J. P., Yang, X., Gautam, A., Gupta, G., Li, Z., Jacobs, R., and Mor- gan, D. Calibration after bootstrap for accurate uncer- tainty quantification in regression models. npj Comput. Mater. 8, 115 (2022)

2022
[18]

R., Urata, S., Goldman, S., Dietschreit, J

Tan, A. R., Urata, S., Goldman, S., Dietschreit, J. C. B. and G´ omez-Bombarelli, R. Single-model uncer- tainty quantification in neural network potentials does not consistently outperform model ensembles. npj Com- put. Mater. 9, 225 (2023)

2023
[19]

and Kitchin, J

Musielewicz, J., Lan, J., Uyttendaele, M. and Kitchin, J. R. Improved uncertainty estimation of graph neural net- work potentials using engineered latent space distances. J. Phys. Chem. C 128, 20799 (2024)

2024
[20]

Varivoda, D., Dong, R., Omee, S. S. and Hu, J. Materials property prediction with uncertainty quantification: A benchmark study. Appl. Phys. Rev. 10, 021409 (2023)

2023
[21]

Chen, L. Y. and Li, Y. P. Uncertainty quantification with graph neural networks for efficient molecular design. Nat. Commun. 16, 3262 (2025)

2025
[22]

and Setter, N

Davis, M., Damjanovic, D. and Setter, N. Electric-field- , temperature-, and stress-induced phase transitions in relaxor ferroelectric single crystals. Phys. Rev. B 73, 014115 (2006)

2006
[23]

and ´I˜ niguez, J

Cazorla, C. and ´I˜ niguez, J. Giant direct and inverse elec- trocaloric effects in multiferroic thin films. Phys. Rev. B 98, 174105 (2018)

2018
[24]

and Fu, D

Moriwake, H., Konishi, A., Ogawa, T., Fisher, C.A.J., Kuwabara, A. and Fu, D. The electric field induced fer- roelectric phase transition of AgNbO 3. J. Appl. Phys. 119, 064102 (2016)

2016
[25]

and Planes, A

Ma˜ nosa, Ll. and Planes, A. Solid-state cooling by stress: A perspective. Appl. Phys. Lett. 116, 050501 (2020)

2020
[26]

and Tamarit, J.-Ll

Lloveras, P. and Tamarit, J.-Ll. Advances and obsta- cles in pressure-driven solid-state cooling: A review of barocaloric materials. MRS Energy Sustain. 8, 3 (2021)

2021
[27]

Novel mechanocaloric materials for solid- state cooling applications

Cazorla, C. Novel mechanocaloric materials for solid- state cooling applications. Appl. Phys. Rev. 6, 041316 (2019)

2019
[28]

and Defay, E

Torell´ o, A. and Defay, E. Electrocaloric Coolers: A Re- view. Adv. Electron. Mater. 8, 2101031 (2022)

2022
[29]

and Sandeman, K

Takeuchi, I. and Sandeman, K. Solid-state cooling with caloric materials. Phys. Today 68, 48 (2015)

2015
[30]

C., Romanini, M., Obrecht, N., Loehl´ e, S., Lloveras, P., Tamarit, J.-Ll

Escorihuela-Sayalero, C., Pardo, L. C., Romanini, M., Obrecht, N., Loehl´ e, S., Lloveras, P., Tamarit, J.-Ll. and 20 Cazorla, C. npj Comput. Mater. 10, 13 (2024)

2024
[31]

E., Cazorla, C., Lu, T., Zhang, L., Tay, Y

Shirsath, S. E., Cazorla, C., Lu, T., Zhang, L., Tay, Y. Y., Lou, X., Liu, Y., Li, S. and Wang, D. Interface-charge induced giant electrocaloric effect in lead free ferroelectric thin-film bilayers. Nano Lett. 20, 1262 (2020)

2020
[32]

and Ma˜ nosa, Ll

Aznar, A., Lloveras, P., Romanini, M., Barrio, M., Tamarit, J.-Ll., Cazorla, C., Errandonea, D., Mathur, N.D., Planes, A., Moya, X. and Ma˜ nosa, Ll. Giant barocaloric effects over a wide temperature range in su- perionic conductor AgI. Nat. Commun. 8, 1851 (2017)

2017
[33]

and Sau, K

Zeng, M., Escorihuela-Sayalero, C., Ikeshoji, T., Takagi, S., Kim, S., Orimo, S.-I., Barrio, M., Tamarit, J.-Ll., Lloveras, P., Cazorla, C. and Sau, K. Colossal reversible barocaloric effects in a plastic crystal mediated by lat- tice vibrations and ion diffusion. Adv. Sci. 11...

2024
[34]

and Cazorla, C

Men´ endez, C., Rurali, R. and Cazorla, C. Colossal room- temperature electrocaloric strength aided by hydrostatic pressure in lead-free multiferroic solid solutions. Phys. Chem. Chem. Phys. 25, 17450 (2023)

2023
[35]

and Tsuneyuki, S

Tadano, T. and Tsuneyuki, S. Quartic anharmonicity of rattlers and its effect on lattice thermal conductivity of clathrates from first principles. Phys. Rev. Lett. 120, 105901 (2018)

2018
[36]

P., Di, C., Yan, X.J

Zang, Y. P., Di, C., Yan, X.J. et al. Giant thermal transport tuning at a metal/ferroelectric interface. Adv. Mater. 33, 2105778 (2021)

2021
[37]

Lead-free room-temperature ferroelectric thermal con- ductivity switch using anisotropies in thermal conductiv- ities

F´ eger, L., Escorihuela-Sayalero, C., Rampnoux, J.-M.et al. Lead-free room-temperature ferroelectric thermal con- ductivity switch using anisotropies in thermal conductiv- ities. Phys. Rev. Mater. 8, 094403 (2024)

2024
[38]

and Rurali, R

Cazorla, C., Bichelmaier, S., Escorihuela-Sayalero, C., ´I˜ niguez, J., Carrete, J. and Rurali, R. Light-driven dy- namical tuning of the thermal conductivity in ferro- electrics. Nanoscale 16, 8335 (2024)

2024
[39]

and Rurali, R

Cazorla, C. and Rurali, R. Dynamical tuning of the thermal conductivity via magnetophononic effects. Phys. Rev. B 105, 104401 (2022)

2022
[40]

and Rurali, R

Cazorla, C., Stengel, M., ´I˜ niguez, J. and Rurali, R. Giant multiphononic effects in a perovskite oxide. npj Comput. Mater. 9, 97 (2023)

2023
[41]

and Li, B

Wang, L. and Li, B. Thermal logic gates: Computation with phonons. Phys. Rev. Lett. 99, 177208 (2007)

2007
[42]

Guangzhao, Q., Yi, W., Linfeng, Y. et al. Predicting lattice thermal conductivity from fundamental material properties using machine learning techniques. J. Mater. Chem. A 11, 5801 (2023)

2023
[43]

and Hu, M

Ojih, J., Al-Fahdi, M., Yao, Y., Hu, J. and Hu, M. 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 (2024)

2024
[44]

and Hu, M

Ojih, J., Shen, C., Rodriguez, A., Zhang, H., Choudhary, K. and Hu, M. High-throughput computational discov- ery of 3218 ultralow thermal conductivity and dynami- cally stable materials by dual machine learning models. J. Mater. Chem. A 11, 24169 (2023)

2023
[45]

and Hu, M

Rodriguez, A., Lin, C., Shen, C., Yuan, K., Al-Fahdi, M., Zhang, X., Zhang, H. and Hu, M. Unlocking phonon properties of a Large and diverse set of cubic crystals by indirect bottom-up machine learning approach. Com- mun. Mater. 4, 61 (2023)

2023
[46]

Enhanced out-of-plane electromechanical re- sponse of Janus ZrSeO

Pham, T.H., Ullah, H., Shafique, A., Kim, H.J., and Shin, Y.-H. Enhanced out-of-plane electromechanical re- sponse of Janus ZrSeO. Phys. Chem. Chem. Phys. 23, 16289 (2021)

2021
[47]

and Tateyama, Y

Zhou, Z., Cazorla, C., Gao, B., Luong, H.D., Momma, T. and Tateyama, Y. First-principles study on the inter- play of strain and state-of-charge with Li-ion diffusion in the battery cathode material LiCoO 2. ACS Appl. Mater. Interfaces 15, 53614 (2023)

2023
[48]

Christensen, A., Norby, P., Hanson, J. C. and Shimada, S. Phase transition of KNO 3 monitored by synchrotron x-ray powder diffraction. J. Appl. Cryst. 29, 265 (1996)

1996
[49]

F., Egorova, V

Markova, Y. F., Egorova, V. M., Roginskiia, E. M. and Stukova, E. V. Phase transition in potassium nitrate fer- roelectric in a nanoporous matrix. Tech. Phys. Lett. 46, 905 (2020)

2020
[50]

Rao, K. J. and Rao, C. N. R. Ferroelectric phase trans- formation in KNO 2. Br. J. Appl. Phys. 17, 1653 (1966)

1966
[51]

and Bassett, W

Merrill, L. and Bassett, W. A. The crystal structure of CaCO3(II), a high-pressure metastable phase of calcium carbonate. Acta Cryst. B31, 343 (1975)

1975
[52]

Universal ion-transport descriptors and classes of inorganic solid-state electrolytes

L´ opez, C., Emperador, A., Saucedo, E., Rurali, R., and Cazorla, C. Universal ion-transport descriptors and classes of inorganic solid-state electrolytes. Mater. Horiz. 10, 1757 (2023)

2023
[53]

How concerted are ionic hops in inorganic solid-state electrolytes? J

L´ opez, C., Rurali, R., and Cazorla, C. How concerted are ionic hops in inorganic solid-state electrolytes? J. Am. Chem. Soc. 146, 8269 (2024)

2024
[54]

and Cazorla, C

Ben ´ ıtez, P., L´ opez, C., Liu, C., Ca˜ no, I., Tamarit, J.-Ll., Saucedo, E. and Cazorla, C. Crystal structure prediction and phase stability in highly anharmonic silver-based chalcohalide anti-perovskites. PRX Energy 4, 023002 (2025)

2025
[55]

R., and Klarbring, J., Tolborg, K

Mosquera-Lois, I., Kavanagh, S. R., and Klarbring, J., Tolborg, K. and Walsh, A. Imperfections are not 0 K: free energy of point defects in crystals. Chem. Soc. Rev. 52, 5812 (2023)

2023
[56]

A generative model for inorganic materials design

Zeni, C., Pinsler, R., Z¨ ugner, D.et al. A generative model for inorganic materials design. Nature 639, 624 (2025)

2025
[57]

and Boronat, J

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

2017
[58]

and Furthm¨ uller, J

Kresse, G. and Furthm¨ uller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 54, 11169 (1996)

1996
[59]

P., Ruzsinszky, A., Csonka, G

Perdew, J. P., Ruzsinszky, A., Csonka, G. I., Vydrov, O. A., Scuseria, G. E., Constantin, L. A., Zhou, X. and Burke, K. Restoring the density-gradient expansion for exchange in solids and surfaces. Phys. Rev. Lett. 100, 136406 (2008)

2008
[60]

Bl¨ ochl, P. E. Projector augmented-wave method. Phys. Rev. B 50, 17953 (1994)

1994
[61]

and Tanaka, I

Togo, A. and Tanaka, I. First principles phonon calcula- tions in materials science. Scr. Mater 108, 1 (2015)

2015
[62]

Phase 1” and the high- T polymorph to “Phase 2

on small training sets of 80 epochs with the fol- lowing grid values: graph convolutional layers with [64, 128, 256, 512] number of neurons, linear convolu- tional layers with [16 , 32, 64, 128, 256] number of neu- rons, dropouts of [0 .1, 0.2, 0.3, 0.4, 0.5] and learning rate...
[63]

and Bengio, Y

Bergstra, J. and Bengio, Y. Random search for hyper- parameter optimization. J. Mach. Learn. Res. 13, 281 (2012)

2012
[64]

Li, W., Carrete, J., Katcho, N. A. and Mingo, N. Sheng- BTE: A solver of the Boltzmann transport equation for phonons. Comput. Phys. Commun. 185, 1747 (2014)

2014
[65]

and DeCost, B

Choudhary, K. and DeCost, B. Atomistic line graph neu- ral network for improved materials property predictions. npj Comput. Mater. 7, 185 (2021)

2021
[66]

https://github.com/IonRepo/ml-phasetransitions

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

The theory of phase transitions

Landau, L. The theory of phase transitions. Nature 138, 860 (1936)

work page 1936

[2] [2]

and Hemley, R

Mao, H.-K. and Hemley, R. J. Ultrahigh-pressure transi- tions in solid hydrogen. Rev. Mod. Phys. 66, 671 (1994)

work page 1994

[3] [3]

Composition-induced phase transitions and enhanced electrical properties in bismuth sodium titanate ceramics

Yin J., Zhao C., Zhang Y., Wu J. Composition-induced phase transitions and enhanced electrical properties in bismuth sodium titanate ceramics. J. Am. Ceram. Soc. 100, 5601 (2017)

work page 2017

[4] [4]

and Cazorla, C

Rurali, R, Escorihuela-Sayalero, C., Tamarit, J.-Ll., ´I˜ niguez-Gonz´ alez, J. and Cazorla, C. Giant photocaloric effects across a vast temperature range in ferroelectric perovskites. Phys. Rev. Lett. 113, 116401 (2024)

work page 2024

[5] [5]

A., Kresse, G

Wijs, G. A., Kresse, G. and Gillan, M. J. First-order phase transitions by first-principles free-energy calcula- tions: The melting of Al. Phys. Rev. B 57, 8223 (1998)

work page 1998

[6] [6]

The Monte Carlo method for the study of phase transitions: A review of some recent progress

Binder, K. The Monte Carlo method for the study of phase transitions: A review of some recent progress. J. Comput. Phys. 59, 1 (1985)

work page 1985

[7] [7]

and Gillan, M

Cazorla, C., Alf´ e, D. and Gillan, M. J. Constraints on the phase diagram of molybdenum from first-principles free-energy calculations. Phys. Rev. B 85, 064113 (2012)

work page 2012

[8] [8]

B., LeBacq, O., Ahuja, R

Belonoshko, A. B., LeBacq, O., Ahuja, R. and Johansson, B. Molecular dynamics study of phase transitions in Xe. J. Chem. Phys. 117, 7233 (2002)

work page 2002