{"id":"479c85d5-a4da-4710-b679-bdb5a263edc8","arxiv_id":"2506.01449","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A machine-learning screen of about 50,000 inorganic crystals predicts 2,079 solid-solid phase transitions between 300 and 600 K and names candidate materials for cooling and thermal switching.","lead":"This paper uses a machine-learning model trained on computed vibrational free energies to screen about 50,000 inorganic crystals and predict over 2,000 solid-solid phase transitions between 300 and 600 K. It also highlights dozens of candidate materials for solid-state cooling and thermal switching.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The screening pipeline never checks whether screened polymorphs are dynamically stable, so many predicted free-energy crossings may rest on ill-defined quasi-harmonic free energies.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the quasi-harmonic free energy in Eqs. (1) and (2) is only defined for crystals with strictly positive phonon frequencies, but the high-throughput screening never verifies 0 K dynamical stability for the screened polymorphs. My independent reading confirms this is the most consequential issue. It is not a minor correction because the entire discovery pipeline -- the 2,079 predicted transitions, the tabulated transition temperatures, and the reported entropy changes -- is built on F_ML_v values that are physically meaningful only for vibrationally stable crystals. The paper's own limitation statement in Sec. IV acknowledges thermal expansion and superionic/melting pathways, but it does not address the absence of a stability filter. The proposed test is concrete and feasible: running phonon calculations on a sample of the screened polymorphs would settle whether the gap is real and how large its impact is. Given the uncertainty-aware framing and the promising validation on a few known systems, the reader's CONDITIONAL verdict remains appropriate: the method is plausible, but the central list cannot be accepted as 'confidently predicted' until the dynamical-stability condition is checked at scale.","tokens_in":22771,"tokens_out":3227,"duration_ms":38245,"concrete_test":"Select a stratified random sample of 100 transitions from the 2,079 (covering stable/metastable, polar/nonpolar, and a range of Delta St values). For each of the 200 polymorphs, compute the 0 K phonon spectrum with the same PBEsol/VASP/PhonoPy settings used in the paper, using supercells comparable to the stated 4x4x4, and count the fraction of polymorphs with imaginary phonon modes. If more than 20% of the sampled transitions contain at least one dynamically unstable polymorph, the central candidate list must be re-derived after applying an explicit positive-frequency stability filter, and the number of valid predicted transitions should be recomputed and reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that 2,079 polymorphic transformations in the 300–600 K range are confidently predicted, with transition temperatures and entropy changes assigned to each. This claim depends on the quasi-harmonic free energy F(V,T) = E0(V) + Fv(V,T) being physically meaningful for every screened polymorph. As the paper itself states in Sec. II A, the quasi-harmonic formalism is applicable only to vibrationally stable systems with strictly positive phonon frequencies; imaginary frequencies render the logarithm in Eq. (2) ill-defined. The ML model was trained exclusively on 6,674 'well-behaved full phonon spectra (i.e., without imaginary frequencies)' (Sec. II C), yet the screening in Sec. II B is applied to roughly 50,000 Materials Project polymorphs without any explicit 0 K dynamical-stability filter. A large fraction of high-throughput crystal-structure databases, especially low-symmetry theoretical polymorphs, are known to exhibit imaginary phonon modes at 0 K. For such structures, the ML model's F_ML_v output is an extrapolation from dynamically stable training data, not a valid quasi-harmonic free energy. The uncertainty quantification in Sec. II D measures latent-space distance to training data, not phonon stability, and the 50% |Delta F_ML_v| filter does not detect systematic errors from imaginary modes. If even a moderate fraction of the 2,079 reported transitions involve one dynamically unstable polymorph, their Tt and Delta St values are not physical predictions, and the abstract's 'robustness and predictive power' claim is substantially overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":23082,"tokens_out":4368,"duration_ms":47249,"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":[{"comment":"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.","section":"Sec. II A and Sec. II B / III A"},{"comment":"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.","section":"Sec. IV"},{"comment":"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.","section":"Sec. II C, Eq. (5), and Tables 1–5"}],"minor_comments":[{"comment":"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.","section":"Sec. III A"},{"comment":"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.","section":"Sec. II B and Methods"},{"comment":"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.","section":"Sec. II C, Eq. (5)"},{"comment":"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.","section":"Tables and abstract"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about missing dynamical-stability filtering is well founded and should be treated as the primary obstacle. The paper's own statements in Sec. II A and Sec. IV acknowledge the conditionality, but the headline claim of 2,079 confident transitions does not. The experimental errors for KNO3 and KNO2 further weaken the calibration argument. These issues are fixable within the manuscript's scope by adding a stability filter or re-benchmarking the claims, so rejection is not warranted, but the current version considerably overstates its support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nSecond, 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.\n\nThird, 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.\n\nNone 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.\n\nWho'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.","headline":"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.","tokens_in":23628,"tokens_out":2242,"would_cite":true,"duration_ms":23126,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["64.70.Kb","63.20.-e","71.15.Mb"],"model":"deepseek-v4-flash","headline":"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.","keywords":["solid-solid phase transitions","machine learning","vibrational free energy","quasi-harmonic approximation","graph neural networks","high-throughput screening","solid-state cooling","thermal switching"],"falsifier":"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.","tokens_in":22574,"feed_emoji":"🧊","tokens_out":14967,"duration_ms":129440,"temperature":0.7,"pith_summary":"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.","feed_headline":"Machine learning finds 2,079 solid-solid transitions","feed_subtitle":"Hundreds of the predicted transitions sit near room temperature and could drive solid-state cooling and thermal switches.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the roughly 50,000 screened crystal structures and their static DFT energies, the pool from which all predicted transitions are drawn.","marker":"[12]"},{"why":"Provides the phonon-calculation methodology behind the DFT vibrational spectra used to train and validate the free-energy model.","marker":"[14]"},{"why":"Source of the 6,674 full phonon spectra, all without imaginary frequencies, on which the machine-learning model is trained.","marker":"[15]"},{"why":"Establishes the typical numerical uncertainty of quasi-harmonic DFT free energies, the benchmark against which the model's training and test errors are judged.","marker":"[11]"},{"why":"Accumulated DFT lattice thermal conductivity data, about 4,700 values, used to train the surrogate model behind the thermal-switching predictions.","marker":"[42–45]"},{"why":"The graph neural network architecture adopted as the surrogate for lattice thermal conductivity prediction.","marker":"[64]"},{"why":"The solver used to obtain the first-principles lattice thermal conductivity ratios (2.23 for Li4TiS4 and 1.40 for NaNO3) that validate the machine-learning estimates.","marker":"[63]"},{"why":"Experimental KNO3 transition-temperature observations, one of the three experimental benchmarks used for validation.","marker":"[48, 49]"},{"why":"Experimental CaCO3 transition data, the validation case with closest agreement to the screen's prediction.","marker":"[51]"}],"fun_headline_variants":["ML predicts 2,079 solid-solid transitions in one sweep","AI finds 2,079 hidden phase changes in crystals","2,079 solid-solid transitions predicted by neural net","Machine learning reveals 2,079 thermal phase switches","Neural net predicts 2,079 transitions for cooling and switching"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["ML predicts 2,079 solid-solid transitions in one sweep","AI finds 2,079 hidden phase changes in crystals","2,079 solid-solid transitions predicted by neural net","Machine learning reveals 2,079 thermal phase switches","Neural net predicts 2,079 transitions for cooling and switching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":2086,"prompt_tokens":1082,"completion_tokens":1004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":932}},"tokens_in":698,"tokens_out":1004,"duration_ms":11530,"temperature":1.0,"reasoning_tokens":932,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:42:40.237517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"P., Hautier, G","cited_arxiv_id":null,"evidence_quote":"Supplies the roughly 50,000 screened crystal structures and their static DFT energies, the pool from which all predicted transitions are drawn."},{"cited_title":"and Tanaka, I","cited_arxiv_id":null,"evidence_quote":"Provides the phonon-calculation methodology behind the DFT vibrational spectra used to train and validate the free-energy model."},{"cited_title":"and Hu, M","cited_arxiv_id":null,"evidence_quote":"Source of the 6,674 full phonon spectra, all without imaginary frequencies, on which the machine-learning model is trained."},{"cited_title":"and Cazorla, C","cited_arxiv_id":null,"evidence_quote":"Establishes the typical numerical uncertainty of quasi-harmonic DFT free energies, the benchmark against which the model's training and test errors are judged."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The graph neural network architecture adopted as the surrogate for lattice thermal conductivity prediction."},{"cited_title":"and Bengio, Y","cited_arxiv_id":null,"evidence_quote":"The solver used to obtain the first-principles lattice thermal conductivity ratios (2.23 for Li4TiS4 and 1.40 for NaNO3) that validate the machine-learning estimates."},{"cited_title":"and Bassett, W","cited_arxiv_id":null,"evidence_quote":"Experimental CaCO3 transition data, the validation case with closest agreement to the screen's prediction."}],"review_version":1}