betaFeedback

WorksPhysics and Astronomy

Thermal Conductivity Predictions with Foundation Atomistic Models

Abstract

Advances in machine learning have led to the development of foundation models for atomistic materials chemistry, enabling quantum-accurate descriptions of interatomic forces across chemically diverse compounds at reduced computational cost. Hitherto, the accuracy and utility of these models have been assessed relying on descriptors based on formation energies or idealized harmonic atomic vibrations. Yet, the rigorous and physically interpretable quantification of their capability to describe both realistic anharmonic atomic dynamics and technologically relevant observables remains a pressing problem. Here, we address this problem, leveraging the Wigner formulation of heat transport and the Grüneisen approach to thermal expansion to connect the atomic-physics awareness of foundation models to their utility in predicting experimentally observable thermomechanical properties, presenting standards and fine-tuning protocols needed to achieve first-principles accuracy. We apply our framework to a database of 103 solids with diverse compositions and structures, demonstrating that it overcomes the major bottlenecks of current methods for designing heat-management materials – high cost, limited transferability, or lack of physics awareness – and its potential to discover materials for next-gen technologies ranging from thermal insulation to neuromorphic computing.

Over the past decades, intense research efforts have been dedicated to tackle the challenging task of fitting the Born-Oppenheimer potential energy surface (PES) as a function of atomic coordinates[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. These resulted in the development of machine-learning potentials (MLPs), which allow us to reproduce atoms’ interaction energies and forces with nearly quantum accuracy and computational cost orders-of-magnitude lower than first-principles methods. These advances have enabled scientists to make progress in understanding how atomistic structure and composition influence macroscopic observables — an example is the thermal conductivity, a property determined by atomic structure and forces, and relevant for technological innovations ranging from energy harvesting[11] to efficient neuromorphic computing[12]. Major drawbacks of methods based on MLPs are the significant work required to generate a first-principles database needed to retrain and validate MLPs, as well as the applicability limited to specific materials’ compositions or structural phases. Past works attempted to bypass these limitations by employing end-to-end machine-learning methods, which predict microscopic[13, 14] or macroscopic[15, 16, 17] materials’ properties very efficiently, but without necessarily accounting for the fundamental physics that determines them. Fittingly, recent work[18] has formally demonstrated the possibility to obtain a physics-aware and mathematically complete description of atomic environments (and thus of forces) by employing Message Passing Networks[19] with many-body messages[20] (with the MACE architecture[20] this can be achieved, e.g., using 4-body terms and one message pass). This breakthrough has enabled the development of foundation machine-learning potentials (fMLPs)[21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33], which are trained across nearly all chemical elements and can be combined to describe almost arbitrary compounds. Recent works have assessed the accuracy of fMLPs in predicting, e.g., atoms’ interaction energies (thermodynamic stability[34]) or harmonic vibrations[35, 36, 37]. However, the rigorous and physically interpretable quantification of the fMLP’s capability to describe atomic-level features of the PES (i.e., "their physics awareness") and their overall utility to determine macroscopic observables are pressing questions that are largely unexplored, due to the complexity of the relation between many-body interatomic forces and observables[38, 39].

Here we introduce a framework to rigorously address such a question, quantifying how the accuracy of fMLPs in describing interatomic forces and PES derivatives connects to their accuracy in predicting experimentally observable thermomechanical properties. We elucidate how fundamental insights into the physics awareness, accuracy, and overall utility of fMLPs can be derived from microscopic analysis of the linear-response solution of the Wigner Transport Equation (WTE) for heat conduction in crystals and glasses[38, 39], and the quasi-harmonic Grüneisen approach to describe thermal expansion[40, 41, 42, 43]. We leverage these insights to define physically interpretable, robust, and stringent benchmark metrics that expose subtle artifacts or inaccuracies in the smoothness (second and third derivatives) of the PES encoded by fMLPs. As these derivatives expose even subtle discontinuities in the PES, compared to established energy-based metrics, our metrics provide information that is physically richer, more robust, and stricter for measuring both the utility of fMLPs and the physical accuracy (smoothness) of the encoded PES[44]. We showcase the capabilities of this combined fMLP-WTE-Grüneisen framework in state-of-the-art (SOTA) non-proprietary fMLPs trained on the Materials Project database[45] (hereafter collectively referred to as ‘mp-fMLPs’) — M3GNet[21], CHGNet[22], MACE-MP-0[23], SevenNET[24], and ORB-v1-MPtraj[46] — assessing their performance over a database of first-principles reference data for 103 chemically and structurally diverse compounds[47, 48]. We discuss how these metrics can be used to assess and improve the expressivity, training, and fine-tuning of fMLPs, enabling first-principles-level predictions of anharmonic vibrational properties and heat conductivity in solids with arbitrary compositions and structures. We conclude by showing how these developments enable the theory-driven design of heat-management materials with microscopic heat-transport physics that violates the semiclassical Boltzmann regime, which are critical for energy and information-management technologies.

Atomic vibrations & thermomechanical observables

The recently developed WTE[38, 39] offers a comprehensive approach to predict from atomistic vibrational properties the thermal conductivity of anharmonic crystals[49], disordered glasses[50], as well as the intermediate case of anharmonic-and-disordered ‘complex crystals’[51]. To assess how the accuracy in the conductivity prediction is affected by the precision with which fMLPs describe atomic vibrations, it is sufficient to consider the directionally averaged trace of the conductivity tensor obtained from the WTE solution in the relaxation-time approximation[39]

κ(T)=κP(T)+κC(T)=1VNc∑q,s[C(q)s∥v(q)s,s∥23[Γ(q)s]−1+∑s′≠sC(q)sC(q)s′C(q)s+C(q)s′ω(q)s+ω(q)s′2(C(q)sω(q)s+C(q)s′ω(q)s′)×∥v(q)s,s′∥2312[Γ(q)s+Γ(q)s′][ω(q)s′−ω(q)s]2+14[Γ(q)s+Γ(q)s′]2].(1)\begin{aligned} \kappa(T) &= \kappa_{\mathrm{P}}(T)+\kappa_{\mathrm{C}}(T) \\ &= \frac{1}{\mathcal{V}N_{\mathrm{c}}}\sum_{\mathbf{q},s}\left[ \frac{C(\mathbf{q})_{s}\lVert\mathbf{v}(\mathbf{q})_{s,s}\rVert^{2}}{3}\left[\Gamma(\mathbf{q})_{s}\right]^{-1} \right.\\ &\quad+\sum_{s'\ne s}\frac{C(\mathbf{q})_{s}C(\mathbf{q})_{s'}}{C(\mathbf{q})_{s}+C(\mathbf{q})_{s'}}\frac{\omega(\mathbf{q})_{s}+\omega(\mathbf{q})_{s'}}{2} \left(\frac{C(\mathbf{q})_{s}}{\omega(\mathbf{q})_{s}}+\frac{C(\mathbf{q})_{s'}}{\omega(\mathbf{q})_{s'}}\right)\\ &\quad\left.\times\frac{\lVert\mathbf{v}(\mathbf{q})_{s,s'}\rVert^{2}}{3} \frac{\frac{1}{2}\left[\Gamma(\mathbf{q})_{s}+\Gamma(\mathbf{q})_{s'}\right]} {\left[\omega(\mathbf{q})_{s'}-\omega(\mathbf{q})_{s}\right]^{2}+\frac{1}{4}\left[\Gamma(\mathbf{q})_{s}+\Gamma(\mathbf{q})_{s'}\right]^{2}} \right]. \tag*{(1)} \end{aligned}

Here, C(q)s=ℏ2ω(q)s2kBT2Nˉ(q)s[Nˉ(q)s+1]C(\mathbf{q})_{s}=\frac{\hbar^{2}\omega(\mathbf{q})_{s}^{2}}{k_{\mathrm{B}}T^{2}}\bar{N}(\mathbf{q})_{s}\left[\bar{N}(\mathbf{q})_{s}+1\right] is the specific heat at temperature TT of the vibration having: wavevector q\mathbf{q}; mode ss; energy ω(q)s\omega(\mathbf{q})_{s}; population given by Bose-Einstein distribution Nˉ(q)s=[exp⁡(ℏω(q)s/kBT)−1]−1\bar{N}(\mathbf{q})_{s}=[\exp(\hbar\omega(\mathbf{q})_{s}/k_{\mathrm{B}}T)-1]^{-1}; total linewidth Γ(q)s=Γa(q)s+Γi(q)s\Gamma(\mathbf{q})_{s}=\Gamma_{\mathrm{a}}(\mathbf{q})_{s}+\Gamma_{\mathrm{i}}(\mathbf{q})_{s}, where Γa(q)s\Gamma_{\mathrm{a}}(\mathbf{q})_{s} depends on TT and is determined by anharmonic third-order derivatives of the interatomic potential, while Γi(q)s\Gamma_{\mathrm{i}}(\mathbf{q})_{s} is TT-independent and regulated by concentration of isotopes (see Methods for details). v(q)s,s′\mathbf{v}(\mathbf{q})_{s,s'} is the velocity operator obtained from wavevector differentiation of the solid’s dynamical matrix[39], its diagonal elements s=s′s=s' are the phonon group velocities and its off-diagonal elements describe couplings between modes ss and s′s' at the same q\mathbf{q}; NcN_{\mathrm{c}} is the number of q\mathbf{q}-points sampling the Brillouin zone and V\mathcal{V} is the crystal’s unit-cell volume.

(1) shows in the first line that the total macroscopic conductivity is determined by two contributions:

the particle-like conductivity (κP\kappa_{\mathrm{P}}), and the wave-like ‘coherence’ conductivity (κC\kappa_{\mathrm{C}}). In such an equation, the term inside the square brackets is the single-phonon contribution to the total conductivity, K(q)s\mathcal{K}(\mathbf{q})_{s}, which, as we will show later, is useful for benchmarking fMLPs. The second line shows that κP\kappa_{\mathrm{P}} originates from phonons that carry heat C(q)sC(\mathbf{q})_{s} by propagating particle-like with velocity v(q)s,s\mathbf{v}(\mathbf{q})_{s,s} over lifetime [Γ(q)s]−1[\Gamma(\mathbf{q})_{s}]^{-1}; it can be shown[38] that κP\kappa_{\mathrm{P}} is equivalent to the conductivity described by the phonon Boltzmann Transport Equation (BTE). The term on the third and fourth lines, instead, shows that κC\kappa_{\mathrm{C}} describes conduction through wave-like tunneling between pairs of phonons s,s′s,s' at wavevector q\mathbf{q}. It has been shown[38, 39] that in simple crystals particle-like propagation dominates over wave-like tunneling (κP≫κC\kappa_{\mathrm{P}}\gg\kappa_{\mathrm{C}}), in complex crystals both these mechanisms co-exist (κP∼κC\kappa_{\mathrm{P}}\sim\kappa_{\mathrm{C}}), and in strongly disordered glasses tunneling dominates (κP≪κC\kappa_{\mathrm{P}}\ll\kappa_{\mathrm{C}})[50].

In addition to the conductivity, the thermal-expansion coefficient is also a property of technological interest, e.g., for aerospace applications[52]. In general, there is no simple correlation between thermal conductivity and thermal expansion coefficients — materials with similar thermal conductivity can exhibit different coefficients of thermal expansion, and vice-versa[53, 54]. Therefore, it is important to investigate to which extent accuracy in predicting thermal conductivity correlates with accuracy in predicting thermal expansion. Thermal expansion, the change in volume VV driven by a change in temperature TT at constant entropy SS, is quantitatively described by the Grüneisen parameter, γ(T)=−∂ln⁡T∂ln⁡V∣S=βBTVC\gamma(T)=-\left.\frac{\partial\ln T}{\partial\ln V}\right|_{S}=\frac{\beta B_{\mathrm{T}}V}{C} — here β\beta is the coefficient of volumetric thermal expansion, BTB_{\mathrm{T}} is the isothermal bulk modulus, and C=1VNc∑q,sC(q)sC=\frac{1}{\mathcal{V}N_{\mathrm{c}}}\sum_{\mathbf{q},s}C(\mathbf{q})_{s} is the specific heat at constant volume[40, 41, 42, 43]. Within the QHA one can resolve the Grüneisen parameter γ\gamma in terms of microscopic phonon contributions (γ(q)s\gamma(\mathbf{q})_{s}),

γ=1VNc∑qs[γ(q)sC(q)sC],withγ(q)s=−Vω(q)s∂ω(q)s∂V.(2)\gamma=\frac{1}{\mathcal{V}N_{\mathrm{c}}}\sum_{\mathbf{q}s}\left[\gamma(\mathbf{q})_{s}\frac{C(\mathbf{q})_{s}}{C}\right], \quad\text{with}\quad \gamma(\mathbf{q})_{s}=-\frac{V}{\omega(\mathbf{q})_{s}}\frac{\partial\omega(\mathbf{q})_{s}}{\partial V}. \tag*{(2)}

This microscopic decomposition shows that, within the QHA, the microscopic phonon-resolved Grüneisen contributions γ(q)s\gamma(\mathbf{q})_{s} are constant over temperature, and the temperature dependence of the macroscopic Grüneisen parameter γ\gamma entirely derives from the phonon specific heat term C(q)sC\frac{C(\mathbf{q})_{s}}{C}. In the following, we will refer to the temperature-dependent microscopic contribution to the macroscopic Grüneisen parameter as the quantity in the square brackets in (2), i.e., G(q)s=γ(q)sC(q)sC\mathcal{G}(\mathbf{q})_{s}=\gamma(\mathbf{q})_{s}\frac{C(\mathbf{q})_{s}}{C}.

In the Methods, we provide a quantitative explanation about the different information about the Potential-Energy Surface (PES) provided by constant-volume properties such as the conductivity, and volume-deforming properties such as the Grüneisen parameter. Specifically, within perturbation theory, the mode

Grüneisen parameter is linearly related to the third-order derivatives of the (PES)[43], while the thermal conductivity is related to the inverse of the square of the third-order PES derivatives[39]. Therefore, macroscopic conductivity and Grüneisen parameter have a very different dependence on microscopic interatomic forces; this implies that conductivity and Grüneisen parameter practically provide complementary information on how accuracy in predicting PES translates into accuracy in describing macroscopic thermomechanical properties.

Vibrational and thermal properties from fMLPs

Since κ(T)\kappa(T) and γ(T)\gamma(T) are observables that can be resolved in terms of microscopic harmonic and anharmonic vibrational properties[39, 55, 56, 57], it is natural to use them to assess the precision of fMLPs in reproducing reference microscopic and macroscopic vibrational properties obtained from Density-Functional Theory (DFT) or from experiments. In particular, we rely on the DFT interatomic force data of Refs.[47, 48], which include 103 diverse binary compounds, involving 34 different chemical species, and rocksalt, zincblende, or wurtzite structures. In the following we will refer to this dataset as "phononDB-PBE" to highlight that it was computed using the Perdew, Burke, and Ernzerhof (PBE) functional[58] used also to generate the training data for the fMLPs[45]. Following the expressions detailed in the Methods, we use these data to calculate the harmonic (ℏω(q)s\hbar\omega(\mathbf{q})_s, vβ(q)s,s′\mathbf{v}^{\beta}(\mathbf{q})_{s,s'}, and Γi(q)s\Gamma_i(\mathbf{q})_s) and third-order anharmonic (Γa(q)s\Gamma_a(\mathbf{q})_s) vibrational properties that determine κ(T)\kappa(T) (Eq. (1)). To evaluate the Grüneisen parameter (Eq. (2)), we extend this dataset to include second-order force constants calculations at expanded and contracted volumes, as detailed in the Methods.

In Figure 1a and Figure 1b we compare the conductivity and Grüneisen parameter, respectively, predicted from first principles (DFT-PBE) or from the mp-MFLP MACE-MP-0 large[23] for the 103 materials in the phononDB-PBE dataset. We highlight that predictions from DFT or fMLPs are generally compatible within a factor of 2. Importantly, in about 20% of the compounds studied, the semiclassical particle-like BTE fails to fully describe heat transport and it is crucial to employ the more general WTE to perform accurate predictions. These cases have very low conductivity (κTOT≲2\kappa_{\mathrm{TOT}} \lesssim2 W/mK), which receives significant contributions from both wave-like tunneling (described by the WTE but missing from the BTE) and particle-like propagation (described by both the WTE and BTE) heat-transport mechanisms, see Methods Figure 6. We highlight how having accurate conductivity predictions does not necessarily imply having accuracy in predictions for the Grüneisen parameter, and vice-versa (e.g., in wurtzite BeO κ\kappa is predicted more accurately than γ\gamma, while in LiBr γ\gamma is more accurate than κ\kappa). We further observe that the distributions of κ\kappa and γ\gamma overlap for zincblende and wurtzite structures, while rocksalt structures generally exhibit lower κ\kappa and higher γ\gamma. Moreover, while fMLPs generally tend to systematically underestimate the conductivity, there is no such bias for the Grüneisen parameter. Overall, our findings highlight that the macroscopic κ\kappa and γ\gamma have different sensitivity to atomistic vibrational properties, and in the following we show that they can provide complementary information on the accuracy with which fMLPs describe reference DFT data.

Conductivity and Grüneisen parameter predictions from DFT-PBE and MACE-MP-0

Figure 1. Conductivity (left) and Grüneisen parameter (right) at 300 K from DFT-PBE and MACE-MP-0 large, for 103 compounds taken from the phononDB-PBE database, which have rocksalt (green), zincblende (orange), or wurtzite (blue) structure. The solid line indicates perfect agreement and the dashed lines discrepancies of a factor of 2. The arrows highlight materials discussed in detail later. Inset, distribution of relative deviations between the conductivities predicted by DFT-PBE and MACE-MP-0. The color gradient in the left panel shows the transition from the WTE regime where heat transport occurs through both propagation and tunneling mechanisms (red) to the semiclassical propagation-dominated BTE regime (white).

Figure 6. Failures of semiclassical Boltzmann transport in binary crystals with low conductivity and having rocksalt (green), zincblende (orange), and wurtzite (blue) structure. These occur when the conductivity described by the BTE in terms of particle-like propagation of phonons (κP\kappa_P) does not dominate over the coherences conductivity due to phonons’ wave-like tunneling (κC\kappa_C). The WTE[39] accounts for both κP\kappa_P and κC\kappa_C, and failures from semiclassical Boltzmann transport are highlighted by a relative strength of particle-wave transport (κP−κC)/(κP+κC)(\kappa_P-\kappa_C)/(\kappa_P+\kappa_C) appreciably smaller than one (more intense red background, as in Figure 1). These conductivities were obtained from the force constants in the phononDB-PBE dataset[47, 48].

Relating macroscopic and microscopic properties

To understand the microscopic origin of the discrepancies between the DFT-PBE and MACE-MP-0 conductivities and Grüneisen parameters, we select three representative materials — wurtzite BeO, zincblende BeTe, and rocksalt LiBr — and show in Figure 2 their phonon band structures, the macroscopic conductivity and Grüneisen resolved in terms of contributions of microscopic phonon modes. Examining the phonon band structures in the first column, we observe that MACE-MP-0 tends to underestimate the high-frequency optical modes compared to DFT-PBE. In particular, we plot the DFT-PBE phonons considering or not the long-range non-analytical correction term[59] (NAC). The NAC term is responsible for the energy splitting between the longitudinal-optical and transverse-optical modes in polar dielectrics[59], and is not fully considered in fMLPs trained using a radial cutoff (e.g., for MACE-MP-0[23] such cutoff is 6 Å, while it is 5 Å for SevenNet[24], CHGNet[22], and M3GNet[21]). This explains why the DFT-PBE phonons without NAC are in closer agreement with the MACE-MP-0 phonons. Importantly, even without NAC, DFT-PBE energies for the optical phonon modes tend to be higher than MACE-MP-0 energies, confirming the general tendency of MACE-MP-0 to underestimate phonon energies[36]. We show in the Methods that such an underestimation of the optical phonon energies, or considering NAC or not, has negligible effect on the specific heat at constant volume — this because in these materials the specific heat is dominated by acoustic phonons. Figure 2 also shows that considering or not the NAC term has a negligible impact on κ(T)\kappa(T) in BeO, BeTe, and LiBr. This can be understood from the third column of Figure 2, where we show that in these materials a significant amount of heat is carried by low-energy acoustic phonons that are negligibly affected by NAC, as well as by the systematic phonon softening. Specifically, we report the energy-linewidth distributions, resolving the anharmonic and isotopic parts of the linewidths (Γa(q)s\Gamma_a(\mathbf{q})_s and Γi(q)s\Gamma_i(\mathbf{q})_s, respectively), and quantifying how much each phonon contributes to the conductivity with the single-mode conductivity contribution K(q)s\mathcal{K}(\mathbf{q})_s defined in Eq. (1). We note, in passing, that while for the 103 materials considered here the NAC term has negligible influence on thermal transport, it could be important in materials where significant heat is carried by optical modes. SOTA fMLPs do not account for NAC, its inclusion will be subject of future work. However, we show in the Methods that, in crystals, this limitation can be bypassed using interatomic force constants calculated using supercells in real space (i.e., accessible by finite-difference supercell calculations done with DFT or fMLPs) and the dipole–dipole interactions calculated using density-functional perturbation theory. As these supercells are smaller than the receptive field of the fMLPs, we use the DFT reference without the NAC contribution to calculate SRME and SRE.

Figure 2. Phonon dispersion, linewidth & mode-Grüneisen distributions, along with macroscopic conductivity & Grüneisen parameter for wurtzite BeO (panel a), zincblende BeTe (b), and rocksalt LiBr (c). DFT-PBE values are computed using the data from Refs.[47, 48], while ‘MACE’ values are computed using the MACE-MP-0 fMLP[23]. For the DFT-PBE phonons, we show the effect of considering (red) or not (green) the long-range non-analytical contribution (NAC); MACE-MP-0 does not fully account for NAC (see text). In the frequency-linewidths distributions at 300 K, we resolve the anharmonic (Γa\Gamma_{\mathrm{a}}) and isotopic (Γi\Gamma_{\mathrm{i}}) parts of the linewidths; the markers’ areas are proportional to the contribution of the given phonon mode to the conductivity. Similarly, the mode Grüneisen parameter distribution against phonon energy is evaluated at 300 K with marker size describing the contribution of the modes to the macroscopic Grüneisen parameter.

The most accurate zero-shot predictions for the energy-linewidth distributions (for BeO in Figure 2a) correspond to a discrepancy of ∼7%\sim7\% between κ(T)\kappa(T) predicted from fMLP (MACE-MP0) and DFT. We stress that fMLPs allow to obtain this accuracy at a computational cost that is more than 3 orders of magnitude lower than DFT (see Methods Figure 10). Figure 2b shows that compatibility between κ(T)\kappa(T) and γ(T)\gamma(T) predicted from DFT or fMLP can also result from cancellation of microscopic errors. E.g., in BeTe there are visible differences in the energy-linewidth distributions, and these largely cancel out when integrated to determine the conductivity. Finally, Figure 2c illustrates that discrepancies in the energy-linewidth distributions do not always compensate, and can translate into significant differences (a factor of 2) on κ(T)\kappa(T). We also note that, depending on the chemical composition, the anharmonic linewidths at room temperature can dominate over the isotopic linewidth (e.g., in BeO) or not (e.g., in BeTe). Similar cancellation of microscopic errors can emerge also for the Grüneisen parameter, as shown in Figure 2b.

Figure 10. Computational cost: DFT vs fMLP MACE-MP-0-L2 for wurtzite BeO. Orange and red show the cost for computing harmonic and anharmonic force constants, respectively; green is the cost to obtain the conductivity from the force constants through the solution of the Wigner transport equation. The force constants were computed on CPU for DFT, and on GPU for MACE (see text for details); these costs were compared using the conversion 1 GPUh = 55 CPUh[121].

The results above motivate us to quantitatively investigate when agreement between DFT and fMLP conductivities and Grüneisen parameters is obtained as a result of accurately described microscopic harmonic and anharmonic vibrational properties, or because of compensating errors. Starting from conductivity, we quantify the macroscopic discrepancies using the Symmetric Relative Error (SRE),

SRE⁡[κ]=2∣κfMLP−κDFT∣κfMLP+κDFT.(3)\operatorname{SRE}[\kappa] = 2 \frac{\left|\kappa_{\mathrm{fMLP}}-\kappa_{\mathrm{DFT}}\right|}{\kappa_{\mathrm{fMLP}}+\kappa_{\mathrm{DFT}}}. \tag*{(3)}

Next, to quantify the error on the microscopic phonon properties, we introduce the Symmetric Relative Mean Error (SRME) on the single-phonon contribution to κ\kappa:

SRME⁡[{K(q)s}]=2Ncν∑qs∣KfMLP(q)s−KDFT(q)s∣κfMLP+κDFT(4)\operatorname{SRME}\left[\{\mathcal{K}(\mathbf{q})_{s}\}\right] = \frac{2}{N_{\mathrm{c}}\nu}\frac{\sum_{\mathbf{q}s}\left|\mathcal{K}_{\mathrm{fMLP}}(\mathbf{q})_{s}-\mathcal{K}_{\mathrm{DFT}}(\mathbf{q})_{s}\right|}{\kappa_{\mathrm{fMLP}}+\kappa_{\mathrm{DFT}}} \tag*{(4)}

where KDFT(q)s\mathcal{K}_{\mathrm{DFT}}(\mathbf{q})_{s} refers to the single-phonon conductivity (term inside square brackets in Eq. (1)) evaluated using DFT, and KfMLP(q)s\mathcal{K}_{\mathrm{fMLP}}(\mathbf{q})_{s} refers to the same expression evaluated using fMLP. Figure 3a illustrates that a large SRME generally implies large SRE, as SRE gives a lower bound on the SRME. Importantly, knowing both SRE and SRME enables us to identify when microscopic error compensation occurs — this is indicated by a large SRME but small SRE. We note that the SRME⁡[K(q)s]\operatorname{SRME}[\mathcal{K}(\mathbf{q})_{s}] error can stem from discrepancies in the harmonic (second-order) or anharmonic (third-order) force constants. In the Methods, Figure 7 we resolve how errors on harmonic and anharmonic vibrational properties contribute to SRME.

Figure 3. SRE & SRME errors on: conductivity (a), Grüneisen parameter (b), and their correlation (c) in 103 chemically and structurally diverse compounds. For conductivity (panel a), large discrepancies between DFT-PBE and mp-fMLP (MACE-MP-0) on microscopic, single-phonon conductivity contributions are described by large conductivity SRME (4). The microscopic SRME metric is more informative than the macroscopic SRE metric, since it captures the cases where microscopic (in)accuracy translates into macroscopic (in)accuracy — e.g., in BeO (LiBr) a low (high) SRME implies a low (high) SRE — and, because SRME is a sum of non-negative contributions, it detects also cases in which compensation of microscopic errors occurs (e.g., BeTe exhibits high SRME but low SRE). For Grüneisen parameter (panel b), the behavior of macroscopic (SRE, Eq. (5)) and microscopic (SRME, Eq. (6)) errors is more complex due to the possibility of having both positive and negative microscopic contributions, which can lead to both cancellation of single-phonon errors and amplification of macroscopic relative errors when large negative and positive contributions coexist, as discussed in the text. However, most of the materials studied here (b) display large SRME and positive microscopic contributions to the Grüneisen parameters, implying that cancellation of errors can be detected, as for the conductivity, by large SRME and low SRE. The insets in panels a and b show materials with very low SRE, while panel c displays the comparison of SRME in thermal conductivity and Grüneisen parameter, demonstrating no clear correlation.

Relation between harmonic and anharmonic errors, and thermal conductivity

Figure 7. Relation between harmonic and anharmonic errors, and thermal conductivity, calculated for 103 compounds using Eq. (16) and Eq. (17) considering DFT and MACE-MP-0[23]. The area of the marker is proportional to SRE[κ]\mathrm{SRE}[\kappa], Eq. (3). We highlight how SRME[har]\mathrm{SRME}[\mathrm{har}] (Eq. (16)), SRME[anh]\mathrm{SRME}[\mathrm{anh}] (Eq. (17)), and SRE[κ]\mathrm{SRE}[\kappa] are all low (large) in BeO (LiBr); in contrast, BeTe displays large SRME[har]\mathrm{SRME}[\mathrm{har}], large SRME[anh]\mathrm{SRME}[\mathrm{anh}], but low SRE[κ]\mathrm{SRE}[\kappa] due to compensation of errors.

Analyzing macroscopic and microscopic errors on the Grüneisen parameter requires additional considerations. In fact, while the positive-definite κ\kappa can be resolved in terms of microscopic non-negative single-phonon contributions, the single-phonon contributions to the Grüneisen parameter can assume positive or negative sign, and yield a macroscopic Grüneisen with positive or negative sign depending on their relative strength. To take this into account, we define the SRE for the Grüneisen parameter as

SRE⁡[γ]=2∣γfMLP−γDFT∣∣γfMLP∣+∣γDFT∣(5)\operatorname{SRE}[\gamma] = 2\frac{\left|\gamma_{\mathrm{fMLP}}-\gamma_{\mathrm{DFT}}\right|}{\left|\gamma_{\mathrm{fMLP}}\right|+\left|\gamma_{\mathrm{DFT}}\right|} \tag*{(5)}

and the corresponding SRME as

SRME⁡[{G(q)s}]=2∑qs∣GfMLP(q)s−GDFT(q)s∣∑qs∣GfMLP(q)s∣+∣GDFT(q)s∣(6)\operatorname{SRME}\left[\{\mathcal{G}(\mathbf{q})_{s}\}\right] = 2\frac{\sum_{\mathbf{q}s}\left|\mathcal{G}_{\mathrm{fMLP}}(\mathbf{q})_{s}-\mathcal{G}_{\mathrm{DFT}}(\mathbf{q})_{s}\right|}{\sum_{\mathbf{q}s}\left|\mathcal{G}_{\mathrm{fMLP}}(\mathbf{q})_{s}\right|+\left|\mathcal{G}_{\mathrm{DFT}}(\mathbf{q})_{s}\right|} \tag*{(6)}

where GDFT(q)s\mathcal{G}_{\mathrm{DFT}}(\mathbf{q})_{s} refers to the single-phonon contribution to the Grüneisen parameter (term inside square brackets in Eq. (2)) evaluated using DFT, and GfMLP(q)s\mathcal{G}_{\mathrm{fMLP}}(\mathbf{q})_{s} refers to the same expression evaluated using fMLP. These definitions yield a non-negative Grüneisen SRME (6), with maximal value 2 and minimal value 0, this in analogy to the conductivity SRME (4).

We have seen that because the microscopic conductivity contributions are non-negative, the SRE provides a lower bound on the SRME. In contrast, microscopic Grüneisen parameter contributions can range from positive to negative, implying four distinct possible scenarios for SRME and SRE: (i) when both SRE and SRME are low, the fMLP’s prediction is accurate (e.g., zincblende AlN in Figure 3b); (ii) when both are high, large microscopic errors add into a large error for the macroscopic Grüneisen parameter, and the prediction is clearly inaccurate (e.g., wurtzite BP in Figure 3b); (iii) when SRME is high but SRE is low, predictions for the macroscopic Grüneisen parameter appear accurate, but are not reliable because they originate from a fortuitous cancellation of microscopic errors in the underlying phonon physics (e.g., wurtzite GaP in Figure 3b); and (iv) when SRE is high and SRME is low, small deviations can induce significant relative errors in cases where the macroscopic Grüneisen parameter vanishes due to large-magnitude and compensating positive and negative microscopic single-phonon contributions (e.g., zincblende AgI in Figure 3b). While cases (i), (ii), and (iii) are totally analogous to those discussed for the positive-definite conductivity, case (iv) emerges only for signed thermomechanical quantities such as the Grüneisen parameter. For most of the 103 materials in the phononDB-PBE database, the macroscopic Grüneisen parameter at 300 K is positive and SRME exceeds SRE, indicating that case (iv) occurs rarely and that for these materials SRME is overall more informative than SRE in assessing the overall accuracy with which the Grüneisen parameter is described.

Importantly, we note that the SRME in the Grüneisen parameter is not clearly correlated with the SRME in the conductivity, see Figure 3c. This demonstrates the importance of analyzing both conductivity and Grüneisen to assess the accuracy and physics-awareness of the fMLPs.

Accuracy of various SOTA foundation models

Here we utilize SRE and SRME for κ\kappa and γ\gamma to study the accuracy of non-proprietary mp-fMLPs on the structures contained in the phononDB-PBE database [47, 48], comparing DFT predictions against M3GNet[21], CHGNet[22], MACE-MP-0[23], SevenNet[24], and ORB-v1-MPTrj[46] in Figure 4. To determine if a particular mp-fMLP tends to systematically overestimate or underestimate the conductivity and Grüneisen compared to DFT-PBE, we rely on the Symmetric Relative Difference (SRD). For the conductivity this is defined as:

Figure 4. Performance of mp-fMLPs in predicting thermal conductivity and Grüneisen parameter. a and c violin plots for the Symmetric Relative Difference (SRD) in the total Wigner conductivity (7) and in the total Grüneisen parameter, respectively, for the materials in phononDB-PBE simulated with ORB-v1-MPtraj[46], SevenNet[24], MACE-MP-[23], CHGNet[22], M3GNet[21]. b and d are SRME[{κ(q)s}\{\kappa(q)_s\}] and SRME[{G(q)s}\{\mathcal{G}(q)_s\}], respectively, in the same materials database. In all panels, the medians are marked with white scatter points, the widths of the boxes represent the interquartile range, and the whiskers show the range of data points without outliers. The pie charts in e display the number of compounds with non-negative frequencies that retained or not the correct crystal symmetry after unconstrained relaxation (green and blue, respectively); we also show whether structural instabilities (imaginary phonon frequencies) were observed in relaxations performed with (orange) or without (red) enforcing symmetries. Unstable structures with imaginary phonon frequencies were included considering κfMLP=0\kappa_{\mathrm{fMLP}}=0 and γfMLP=0\gamma_{\mathrm{fMLP}}=0, i.e., SRD =−2=-2, SRME =2=2.

SRD⁡[κ]=2κfMLP−κDFTκfMLP+κDFT(7)\operatorname{SRD}[\kappa] = 2\frac{\kappa_{\mathrm{fMLP}}-\kappa_{\mathrm{DFT}}}{\kappa_{\mathrm{fMLP}}+\kappa_{\mathrm{DFT}}} \tag*{(7)}

and similarly for the Grüneisen parameter,

SRD⁡[γ]=2γfMLP−γDFT∣γfMLP∣+∣γDFT∣(8)\operatorname{SRD}[\gamma] = 2\frac{\gamma_{\mathrm{fMLP}}-\gamma_{\mathrm{DFT}}}{\left|\gamma_{\mathrm{fMLP}}\right|+\left|\gamma_{\mathrm{DFT}}\right|} \tag*{(8)}

In both cases the SRD ranges from -2 to +2, corresponding to fMLP overestimating and underestimating the reference DFT values, respectively.

Figures 4a,c summarize in violin plots the SRD for the compounds in the phononDB-PBE database. Unstable structures with imaginary phonon frequencies were included considering κfMLP=0\kappa_{\mathrm{fMLP}}=0 and γfMLP=0\gamma_{\mathrm{fMLP}}=0, i.e., SRD=−2-2, SRME=2. This analysis reveals that even the most accurate mp-fMLP tends to underestimate the thermal conductivity, as a consequence of underestimating vibrational frequencies[36], and in some cases overestimating anharmonic linewidths. In contrast, such a systematic error is not observed for the Grüneisen parameter, confirming that the analyses of κ\kappa and γ\gamma provide complementary information on the accuracy of fMLPs.

ORB-v17NetMACECHGN.M3GN.
mean SRE[κ\kappa]1.5550.6050.5281.6951.340
mean SRME[{K(q)s}\{\mathcal{K}(q)_s\}]1.7860.7610.6691.7171.409
mean SRE[γ\gamma]0.6670.4990.2820.9760.694
mean SRME[{G(q)s}\{\mathcal{G}(q)_s\}]0.8990.5570.3741.3450.941
F1 (11/9/2024)[34]0.7630.7240.6690.6130.569

Table 1. Benchmark metrics for fMLPs. The mean of SRE[κ\kappa], SRME[{K(q)s}\{\mathcal{K}(q)_s\}], SRE[γ\gamma] and SRME[{G(q)s}\{\mathcal{G}(q)_s\}] introduced here resolve the accuracy of fMLPs in describing interatomic forces and derived structural and thermal properties. The Matbench F1[34] descriptor is based on atomic energies and resolves thermodynamic stability. Different (similar) values for SRE and SRME indicate presence (absence) of cancellation of errors on microscopic vibrational properties. The lower the SRE is, the higher the accuracy is in predicting macroscopic κ\kappa(300K) and γ\gamma(300K). The lower the SRME is, the higher the accuracy is in predicting both microscopic vibrational properties, and macroscopic κ\kappa(300K) and γ\gamma(300K). The higher the F1 score is, the higher the accuracy is in describing thermodynamic stability.

To examine how the macroscopic SRD in Figure 4a,Figure 4c are influenced by microscopic errors, we show in Figure 4b,Figure 4d violin plots for SRME in the mode contribution to conductivity and Grüneisen parameter. For the conductivity, this directly relates to microscopic error compensation, whereas the relation between SRD (or SRE) and SRME in the Grüneisen parameter is more complex due to the possibility of negative values and amplification of the relative error.

Among the mp-fMLP tested, the one that shows the highest accuracy is MACE-MP-[23], which produces zero-shot conductivities compatible within a factor of two of the DFT-PBE values for 69% of the materials in the phononDB-PBE database[47, 48]. The zero-shot Grüneisen parameter is compatible within a factor of two for 67% of the materials.

To quantify the overall accuracy of a fMLP in predicting the macroscopic conductivity and Grüneisen parameter over a materials’ database, without resolving the possible compensation of microscopic errors, it is informative to consider the mean of the modulus of the deviations, i.e., the mean of the distribution of SRE (3) — we prefer this over the mean of the SRD distribution, as the latter can be close to zero in the presence of very broad but sign-symmetric distribution.

Starting from the conductivity, the mean for SRE[κ\kappa] and mean for SRME[{K(q)s}\{\mathcal{K}(q)_s\}] (Table I) are comparable in the absence of cancellation of microscopic errors, while the mean SRE[κ\kappa] is significantly lower than mean SRME[K(qs)\mathcal{K}(q_s)] when cancellation of microscopic errors occurs. As anticipated, this relation is not necessarily respected for SRE[γ\gamma] and SRME[G(qs)\mathcal{G}(q_s)], however the materials in focus here have non-negligible Grüneisen parameter with dominating positive single-phonon contributions, and therefore mostly display a relation between SRE[γ\gamma] and SRME[G(qs)\mathcal{G}(q_s)] analogous to the one for the conductivity.

Importantly, we note that our benchmarks on SRME[K(qs)\mathcal{K}(q_s)], SRME[G(qs)\mathcal{G}(q_s)] and SRE[κ\kappa], SRE[γ\gamma] evaluate the accuracy of fMLPs in predicting experimentally observable conductivity and Grüneisen parameter, which are macroscopic properties determined not only by single (local-energy-minima) point of the potential energy surface (PES), but also by its second- and third-order derivatives (i.e., harmonic and anharmonic interatomic forces[60]). Because these derivatives expose even subtle discontinuities in the PES, our metrics provide information that is physically richer and more robust compared to established energy-based metrics, and overall stricter in measuring both the utility of fMLPs and the physics encoded in the smoothness of the PES[44].

Table I highlights how fMLPs that are highly accurate according to tests based on interaction energies can be inaccurate in predicting crystal structures, interatomic forces, and thermomechanical properties. Conversely, for the materials studied so far, fMLPs that are highly accurate according to the SRME[K(qs)\mathcal{K}(q_s)] and SRME[G(qs)\mathcal{G}(q_s)] metrics are also highly ranked by energy-based tests, and Figure 4c also shows that low SRME[K(qs)\mathcal{K}(q_s)] is correlated with correctly describing crystal symmetries.

Importantly, the MPtrj dataset, used to train the fMLPs discussed in this work, includes multiple DFT simulations for the same structure computed with DFT settings of different accuracy (see, e.g., BeTe mp-252, which contains tasks mp-657299 and mp-1057046 that were computed with different k-point grids); this may introduce noise into the training data. In this study, to ensure consistency, we adopted the same DFT parameters as those used for the phononDB-PBE database[48, 61] in both conductivity and Grüneisen calculations. Therefore, even if the MPtrj and phononDB-PBE datasets were computed with the same DFT functional, the DFT thresholds in these two datasets are not always exactly the same, and the MPtrj dataset may contain noise related to variations in the stringency of the DFT thresholds. The strong improvement in the accuracy of fMLPs upon fine-tuning shows that these discrepancies are unimportant for assessing the performance of the fMLPs discussed in this work. However, as fMLPs continue to improve in accuracy, discrepancies between DFT and fMLP predictions—especially the residual few-percent differences after fine-tuning—may become comparable to the noise introduced by variations in the stringency of the DFT thresholds in the different datasets. Future work should therefore focus on generating more accurate training and reference DFT data for fMLPs, with higher accuracy, comprehensiveness, and quality[62].

Achieving first-principles accuracy via fine-tuning

In the previous section we discussed LiBr as paradigmatic example where the zero-shot conductivity predictions from MACE-MP-0 are in strong disagreement — about a factor of two — with DFT-PBE reference data. Predictions for the Grüneisen parameter also show a 20% discrepancy. These discrepancies might be due to lack of relevant reference data in the MP dataset[45] used to train MACE-MP-0, and in this section we introduce a fine-tuning protocol to correct them and achieve first-principles accuracy.

Employing the same DFT settings used to generate reference DFT data (see Methods for details), we prepared three distinct datasets for fine-tuning. Each dataset contains a set of different DFT-PBE frames — with ’frame’ we mean a set comprising forces, total energy, and stress computed from DFT for an atomic configuration perturbed from equilibrium. Perturbations consist of rattling (random atomic displacements drawn from Gaussian distribution with standard deviation of 0.03Å)[64], and in some cases also a few-percent isotropic rescaling of volume. DFT frames have the same size as those used to compute the reference phononDB-PBE phonons, 4×4×44 \times4 \times4 supercell of LiBr (512 atoms).

We generated three training datasets for fine-tuning, which contain: (i) one single frame at equilibrium volume and with atomic rattling; (ii) 3 frames, out of which 2 feature ±1%\pm1\% isotropic volume rescaling and rattling, plus the frame in (i) at equilibrium volume and with rattling; (iii) 20 frames, out of which 10 are at equilibrium volume with different rattling, and 10 frames with rattling and volume variations of ±0.25%\pm0.25\%, ±0.50%\pm0.50\%, ±1%\pm1\%, ±2%\pm2\%, and ±5%\pm5\%. Using these datasets, we performed fine-tuning training with number of epochs ranging from 1 to 1000. Training root-mean-square errors (RMSE) on energy, forces, and stresses, are discussed in the Methods (Figure 11). Validation was performed relying on reference data for LiBr in the phononDB-PBE database, in particular Figure 5a and Figure 5b show the validation error for SRME[K(qs)\mathcal{K}(q_s)] and SRME[G(qs)\mathcal{G}(q_s)] respectively. Figure 5a highlights how fine-tuning relying on the 3-frame dataset (ii) and for 200 epochs yields a validation error on SRME[K(qs)\mathcal{K}(q_s)] <0.1<0.1, which corresponds to compatibility within 2% with the reference DFT-PBE conductivity κ(300K)\kappa(300\mathrm{K}). In the case of the Grüneisen parameter, in Figure 5b, the SRME[G(qs)\mathcal{G}(q_s)] is <0.03<0.03, which corresponds to compatibility within 0.5% with the reference DFT-PBE Grüneisen parameter γ(300K)\gamma(300\mathrm{K}). We note that low SRME⁡[K(qs)]\operatorname{SRME}[\mathcal{K}(q_s)] and SRME⁡[G(qs)]\operatorname{SRME}[\mathcal{G}(q_s)] imply also remarkable compatibility with reference phononDB-PBE data for specific heat (Figure 5d), phonons (Figure 5d), microscopic energy-linewidth distribution (Figure 5f) and mode Grüneisen parameter (Figure 5h). The accuracy of the fine-tuned fMLP in predicting the macroscopic, experimentally measurable conductivity was evaluated over a test set including: ab-initio molecular dynamics (aiMD)[56] performed at 300 K using the same PBE functional used in our work; experimental data[63] in the temperature range 100-800 K. Figure 5e highlights how our predictions agree, within error bars, both in trend and magnitude with experiments (error bars show the compatibility threshold of 20% discussed in Ref.[63]), as well as with independent aiMD predictions[56]. We also note that using the larger fine-tuning dataset (iii) does not show significant improvements in accuracy, and the single-frame dataset (i) allows to reduce conductivity discrepancies from 47% to 7% in 100 fine-tuning epochs. Figure 5g shows that the obtained Grüneisen parameter is compatible with the reference DFT-PBE data (the maximum deviation in the temperature range 50-800 K is smaller than 0.7%); comparisons against experiments on thermal expansion or Grüneisen parameter are not reported because, to the best of our knowledge, these are not available in the literature.

Eight-panel comparison of fine-tuned MACE and DFT-PBE predictions for LiBr vibrational and thermal properties

Figure 5. Achieving first-principles accuracy on LiBr’s vibrational & thermal properties through fine-tuning. Top of panel a (b), conductivity (Grüneisen parameter) as a function of number of DFT-PBE supercell frames (see text) in the fine-tuning dataset, and number of fine-tuning epochs. The MACE zero-shot prediction for conductivity (Grüneisen parameter) is purple; performing fine-tuning on one single DFT frame and 100 fine-tuning epochs yields compatibility within 7% for conductivity (5.5% for Grüneisen) with the reference DFT-PBE value (green), while with datasets containing 3 frames with rattling and volume variations yields compatibility within 2% for conductivity (0.5% for Grüneisen) at 200 epochs. The bottom of panel a (b) show that SRME⁡[K(qs)]\operatorname{SRME}[\mathcal{K}(q_s)] (SRME⁡[G(qs)]\operatorname{SRME}[\mathcal{G}(q_s)]) converges to a minimum upon increasing the number of epochs, and larger datasets yield faster convergence. After fine-tuning on 3 frames for 200 epochs, remarkable agreement between fine-tuned MACE and DFT-PBE is observed for: specific heat (c), phonons (d), microscopic phonon energy-linewidth (f) and energy-mode Grüneisen parameter distributions (h), macroscopic κ(T)\kappa(T) (e) and γ(T)\gamma(T) (g). We note that after fine-tuning, our WTE κ(T)\kappa(T) predictions agree with experiments[63], as well as with predictions from ab-initio (PBE) molecular-dynamics (aiMD) at 300 K[56].

Figure 11. Improving fMLP accuracy in LiBr via fine-tuning. Upon increasing the number of fine-tuning epochs, the root mean square training errors on energies (panel a), forces (b and c), and stresses (d) decrease. The green regions highlight the errors at the end of the 3-frame fine-tuning for MACE-MP-0, which was then used to produce Figure 5b-Figure 5d. These errors approximately correspond to a validation error at 300 K of 2% on SRE (3), and of 0.1 on SRME (4).

Finally, we emphasize that employing the WTE is essential for accurately describing heat transport in LiBr over a broad temperature range. In fact, with increasing temperature the tunneling conductivity κC\kappa_C becomes more important and non-negligible compared to the propagation conductivity κP\kappa_P, and the total WTE conductivity κT(T)=κP(T)+κC(T)\kappa_T(T)=\kappa_P(T)+\kappa_C(T) is in better agreement with the experimental trend compared to the BTE (propagation-only) conductivity κP(T)\kappa_P(T).

We highlight how with one single fine-tuning procedure we obtained quantitatively accurate predictions for both thermal conductivity and Grüneisen parameter (evaluated with the volume-derivative method, see Methods). More generally, additional properties like temperature-dependent bulk modulus and thermal-expansion coefficient can be directly obtained from fine-tuned fMLPs by direct free energy minimization within QHA[40, 65] (see Methods). This represents an advantage compared to established methods such as: (i) compressive sensing, which requires different DFT datasets at different volumes to compute these properties (see, e.g., Refs. [64] for a thorough discussion compressive sensing methods applied to simulations of vibrational properties); (ii) training a conventional (non-foundation) machine learning potential from scratch, which requires a large training dataset to obtain accurate results. Importantly, fMLPs enable zero-shot predictions that are rapidly becoming increasingly more accurate[44]; this is an advantage compared to established (compressive-sensing or non-foundation methods), which require supercell DFT calculations for direct predictions or training.

Another important application of fMLPs is that of high-throughput materials screening, which aims at identifying optimal materials for target applications by systematically calculating and comparing the properties of a large number of materials with different compositions and structures[15, 37, 66, 67, 68, 69, 70, 71]. In the Methods, we demonstrate that SOTA fMLPs can be simultaneously fine-tuned for multiple compositions (rather than independently fine-tuning different compositions), a property that may be highly beneficial to minimize the computational cost of high-throughput screening applications. In particular, we show that simultaneous fine-tuning in three compositions yields conductivity and Grüneisen parameter compatible within 5% with DFT reference. This fine-tuning was particularly data efficient, as it was performed using only three atomic configurations: a single rattled supercell at equilibrium volume per material; two unit-cell-sized configurations at expanded and contracted volume, respectively, to describe stresses.

Conclusions

We introduced a framework that leverages foundation models for atomistic materials chemistry[21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33], the Wigner formulation of heat transport[38, 39], and the quasi-harmonic approach for thermal expansion[40, 43] to address the major challenges of current techniques for designing heat-management materials: high computational cost, limited transferability, or lack of physics awareness. We leverage this framework to quantify how the accuracy of fMLPs in describing many-body interatomic forces connects to their accuracy in predicting experimentally observable thermomechanical properties. We leverage these insights to define physically interpretable, robust, and stringent benchmark metrics that expose how subtle artifacts or inaccuracies in the smoothness of the Born-Oppenheimer potential energy surface (PES) encoded by fMLPs affect the prediction of constant-volume observables such as the thermal conductivity, and volume-deforming properties such as the Grüneisen parameter. Because our metrics test not only single-energy values of the PES but also its derivatives (smoothness), they are physically richer, more robust, and stricter compared to established single-energy metrics[44], as recognized by several works[44, 72, 73, 74] released after the first preprint version of this manuscript.

For these reasons, shortly after the preprint version of this work, our κSRME\kappa_{\mathrm{SRME}} conductivity metric (Eq. (4)) has been included on the Matbench Discovery platform[44] and recognized as a key metric to assess both the utility and physical accuracy of the potential energy surface described by fMLPs[44, 72, 73, 74]. These statements are supported by very recent work from Orbital Materials[73]: they employed the metrics discussed here (in the first preprint version of this manuscript) to develop the Orb-v3 model, which displays significantly improved zero-shot accuracy compared to the model Orb-v1 discussed in this manuscript.

We also demonstrated that zero-shot foundation machine-learning potentials (fMLPs) predictions for conductivity and Grüneisen parameter can be within a factor of 2 accuracy from DFT reference data, and near DFT accuracy (within 2%) can be obtained through fine-tuning fMLPs. Most importantly, we showed how fine-tuned fMLPs also enable the prediction of several thermomechanical properties at significantly reduced computational cost compared to current DFT-based methods (e.g., based on systematic finite-differences[61], compressive sensing[64], or density-functional perturbation theory[75]), as well as compared to conventional (composition-specific, not-foundation) machine-learning potentials that require a computationally expensive training from scratch. Importantly, our physics-aware framework also overcomes the reliability issues (unknown unknowns) of current end-to-end machine-learning methods for predicting vibrational or thermal properties of materials with arbitrary composition and structure.

Finally, we have shown that our framework enables the screening or prediction of materials that violate semiclassical Boltzmann transport, which are crucial for applications ranging from thermal insulation[11, 39] to neuromorphic computing[12]. It is worth mentioning that after posting the preprint of this manuscript on arXiv, the effectiveness of fMLP-based computational frameworks in predicting thermomechanical properties has been discussed also by Microsoft Research for AI, which employed the fMLP MatterSim[25] to perform high-throughput screening of materials with extreme lattice thermal conductivities[76]. Ultimately, we have shown that combining the Wigner Transport Equation with fMLP enables the theory-driven optimization, design, and discovery of materials for next-generation energy and information-management technologies.

Methods

Vibrational properties relevant to compute κ(T)\kappa(T) and γ(T)\gamma(T)

In this section we discuss the relation between inter-atomic vibrational energies and the quantities that appear in the WTE thermal conductivity expression ((1)) and in the formula for the Grüneisen parameter ((2)). The starting point is the Born-Oppenheimer Hamiltonian for atomic vibrations[39] expanded up to anharmonic third-order[55, 77, 78, 79, 80, 81, 82, 83] in the atomic displacements from equilibrium, and accounting for energy perturbations due to presence of isotopes[84, 85, 86, 87, 88]

H^=∑R,b,αp^Rbα22Mb+12∑R,b,αR′,b′,α′∂2V∂uRbα∂uR′b′α′∣equ^Rbαu^R′b′α′+13!∑R,b,αR′,b′,α′R′′,b′′,α′′∂3V∂uRbα∂uR′b′α′∂uR′′b′′α′′∣equ^Rbαu^R′b′α′u^R′′b′′α′′+∑R,b,α(Mbmb−1)p^Rbα22Mb.(9)\hat{H}=\sum_{R,b,\alpha}\frac{\hat{p}_{Rb\alpha}^{2}}{2M_b}+\frac{1}{2}\sum_{\substack{R,b,\alpha\\R',b',\alpha'}}\left.\frac{\partial^{2}V}{\partial u_{Rb\alpha}\partial u_{R'b'\alpha'}}\right|_{\mathrm{eq}}\hat{u}_{Rb\alpha}\hat{u}_{R'b'\alpha'}+\frac{1}{3!}\sum_{\substack{R,b,\alpha\\R',b',\alpha'\\R'',b'',\alpha''}}\left.\frac{\partial^{3}V}{\partial u_{Rb\alpha}\partial u_{R'b'\alpha'}\partial u_{R''b''\alpha''}}\right|_{\mathrm{eq}}\hat{u}_{Rb\alpha}\hat{u}_{R'b'\alpha'}\hat{u}_{R''b''\alpha''}+\sum_{R,b,\alpha}\left(\frac{M_b}{m_b}-1\right)\frac{\hat{p}_{Rb\alpha}^{2}}{2M_b}. \tag*{(9)}

Here, p^Rbα\hat{p}_{Rb\alpha} and u^Rbα\hat{u}_{Rb\alpha} are the momentum and position-displacement operators for the atom bb having isotope-averaged mass MbM_b, position R+τbR+\tau_b (RR is the Bravais-lattice vector and τb\tau_b the position in the crystal’s unit cell), and along the Cartesian direction α\alpha; the last term describes kinetic-energy perturbations induced by isotopes (mbm_b is the exact mass of atom at τb\tau_b, we adopt the common approximation[81, 84] that the mass mbm_b depends only on the index bb and not on the lattice vector RR). The leading (harmonic) term in Eq. (9) determines the vibrational frequencies. In particular, the Fourier transform of the mass-rescaled Hessian of the interatomic potential yields the dynamical matrix at wavevector qq,

D(q)bα,b′α′=∑R∂2V∂uRbα∂uR′b′α′∣eqe−iq⋅(R+τb−τb′)MbMb′,(10)D(q)_{b\alpha,b'\alpha'}=\sum_R\left.\frac{\partial^2V}{\partial u_{Rb\alpha}\partial u_{R'b'\alpha'}}\right|_{\mathrm{eq}}\frac{e^{-iq\cdot(R+\tau_b-\tau_{b'})}}{\sqrt{M_bM_{b'}}}, \tag*{(10)}

and by diagonalizing the dynamical matrix,

∑b′α′D(q)bα,b′α′E(q)s,b′α′=ω2(q)sE(q)s,bα,(11)\sum_{b'\alpha'}D(q)_{b\alpha,b'\alpha'}\mathcal{E}(q)_{s,b'\alpha'}=\omega^2(q)_s\mathcal{E}(q)_{s,b\alpha}, \tag*{(11)}

one obtains from the eigenvalues the phonon energies of the solid ℏω(q)s\hbar\omega(q)_s (ss is a band index ranging from 1 to 3Nat3N_{\mathrm{at}}, where NatN_{\mathrm{at}} is the number of atoms in the unit cell), and from the eigenvectors E(q)s,bα\mathcal{E}(q)_{s,b\alpha} the displacement patterns of atom bb in direction α\alpha for the phonon having wavevector qq and mode ss.

The vibrational frequencies ℏω(q)s\hbar\omega(q)_s are sensitive to the thermally induced volume expansion, as quantified by mode Grüneisen parameter, Eq. (2)[43]. In practice, the volume derivative is calculated using the Hellman-Feynman theorem[65],

γ(q)s=−V2ω(q)s2∑bα,b′α′E∗(q)s,bα∂D(q)bα,b′α′∂VE(q)s,b′α′,(12)\gamma(q)_s=-\frac{\mathcal{V}}{2\omega(q)_s^2}\sum_{b\alpha,b'\alpha'}\mathcal{E}^{*}(q)_{s,b\alpha}\frac{\partial D(q)_{b\alpha,b'\alpha'}}{\partial\mathcal{V}}\mathcal{E}(q)_{s,b'\alpha'}, \tag*{(12)}

where the partial derivative of the dynamical matrix is obtained by computing the numerical differences at volumes expanded or contracted by 1%.

As discussed in detail in Ref.[39], the velocity operator appearing in the WTE conductivity expression (1) is determined by the wavevector derivative of the square root of the dynamical matrix:

νβ(q)s,s′=∑b,α,b′,α′E∗(q)s,bα[∇qβD(q)bα,b′α′]E(q)s′,b′α′.(13)\nu^\beta(q)_{s,s'}=\sum_{b,\alpha,b',\alpha'}\mathcal{E}^{*}(q)_{s,b\alpha}\left[\nabla_q^\beta\sqrt{D(q)}_{b\alpha,b'\alpha'}\right]\mathcal{E}(q)_{s',b'\alpha'}. \tag*{(13)}

The anharmonic linewidths ℏΓa(q)s\hbar\Gamma_{\mathrm{a}}(q)_s, which contribute to the total linewidths ℏΓ(q)s=ℏΓa(q)s+ℏΓi(q)s\hbar\Gamma(q)_s=\hbar\Gamma_{\mathrm{a}}(q)_s+\hbar\Gamma_{\mathrm{i}}(q)_s appearing in Eq. (1), are the energy broadenings due to anharmonic three-phonon interactions[77, 81, 89, 90, 91]. It can be shown[80] that these are determined by the third derivative of the interatomic potential (see Eq. (9)),

ℏΓa(q)s=πNc2∑q′q′′∑s′s′′[2(Nˉ(q′)s′−Nˉ(q′′)s′′)δ(ω(q)s+ω(q′)s′−ω(q′′)s′′)+[1+Nˉ(q′)s′+Nˉ(q′′)s′′]δ(ω(q)s−ω(q′)s′−ω(q′′)s′′)]×\begin{aligned} \hbar\Gamma_{\mathrm{a}}(q)_s=\frac{\pi}{N_c^2}\sum_{q'q''}\sum_{s's''}[2(\bar{N}(q')_{s'}-\bar{N}(q'')_{s''})\delta(\omega(q)_s+\omega(q')_{s'}-\omega(q'')_{s''}) \\ &+[1+\bar{N}(q')_{s'}+\bar{N}(q'')_{s''}]\delta(\omega(q)_s-\omega(q')_{s'}-\omega(q'')_{s''})]\times \end{aligned}
\left|\sum_{\alpha,\alpha',\alpha''}\sum_{b,b',b''}\left.\frac{\partial^3V}{\partial u_{0b\alpha}\partial u_{R'b'\alpha'}\partial u_{R''b''\alpha''}}\right| \mathcal{E}(q)_{s,b\alpha}\mathcal{E}(q')_{s',b'\alpha'}\mathcal{E}(q'')_{s'',b''\alpha''}\sqrt{\frac{\hbar^3\Delta(q+q'+q'')e^{-i[q\cdot\tau_b+q'\cdot(R'+\tau_{b'})+q''\cdot(R''+\tau_{b''})]}}{8\sqrt{M_bM_{b'}M_{b''}\omega(q)_s\omega(q')_{s'}\omega(q'')_{s''}}}\right|^2,} \tag*{(14)}

where Nˉ(q)s=[exp⁡(ℏω(q)s/kBT)−1]−1\bar{N}(q)_s=[\exp(\hbar\omega(q)_s/k_{\mathrm{B}}T)-1]^{-1} is the Bose-Einstein distribution, Δ(q+q′+q′′)\Delta(q+q'+q'') is the Kronecker delta (equal to 1 if q+q′+q′′q+q'+q'' is a reciprocal lattice vector, 0 otherwise), δ\delta is the Dirac delta. The linewidth due to isotopic-mass disorder is determined exclusively by harmonic properties and concentration of isotopes[84]:

ℏΓi(q)s=ℏπ2Nc[ω(q)s]2∑q′,s′δ[ω(q)s−ω(q′)s′]×∑bgb2∣∑αE(q)s,bα∗E(q′)s′,bα∣2,(15)\hbar\Gamma_{\mathrm{i}}(q)_s=\frac{\hbar\pi}{2N_c}[\omega(q)_s]^2\sum_{q',s'}\delta[\omega(q)_s-\omega(q')_{s'}]\times\sum_b g_b^2\left|\sum_\alpha\mathcal{E}(q)^{*}_{s,b\alpha}\mathcal{E}(q')_{s',b\alpha}\right|^2, \tag*{(15)}

where gb2=∑ifi,b(Mb−mi,bMb)2g_b^2=\sum_i f_{i,b}\left(\frac{M_b-m_{i,b}}{M_b}\right)^2 describes the variance of the isotopic masses of atom bb (fi,bf_{i,b} and mi,bm_{i,b} are the mole fraction and mass, respectively, of isotope ii of atom bb; Mb=∑ifi,bmi,bM_b=\sum_i f_{i,b}m_{i,b} is the weighted average mass).

Relation between Grüneisen parameter and third-order force constants

In the quasi-harmonic approximation (QHA), expanding the Hamiltonian in the first-order of applied strain allows one to evaluate the mode Grüneisen parameter[43, 92], which depends linearly on the third-order derivatives of the potential energy surface (PES). In contrast, the thermal conductivity is, in the BTE relaxation-time approximation regime, inversely proportional to the anharmonic linewidths, and the latter depend on the square of the third-order derivatives of the PES. These distinct dependencies of conductivity and Grüneisen parameter from PES derivatives intuitively shows why our SRME[{K(q)s}][\{\mathcal{K}(q)_s\}] and SRME[{G(q)s}][\{\mathcal{G}(q)_s\}] provide complementary information: errors in the PES derivative that are very important for conductivity may have negligible impact on Grüneisen parameter, and vice-versa. Therefore, jointly analyzing thermal conductivity and Grüneisen parameters enables a more comprehensive benchmark of fMLP accuracy, helping in minimizing the risks of overlooking errors that could arise from subtle inaccuracies in modeling the physical PES.

Failures of semiclassical Boltzmann transport

As anticipated in the main text, our automated framework can be used to identify materials in which the semiclassical particle-like BTE fails. This happens when particle-like propagation mechanisms — described by both the BTE and WTE[38, 49, 93], and determining κP\kappa_P in Eq. (1) — do not dominate over wave-like tunneling mechanisms — missing from the BTE but described by the WTE and determining κC\kappa_C in Eq. (1). Failures of the BTE have been discussed to appear in, e.g., strongly anharmonic complex crystals for energy harvesting[94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106], thermal barrier coatings[51, 107, 108, 109], as well as in several disordered functional materials[110, 111, 112, 113, 114, 115]. Our automated framework computes the total Wigner thermal conductivity κP+κC\kappa_P+\kappa_C[39], and therefore can be used to find materials that violate semiclassical Boltzmann thermal transport. In Figure 6 we show that among the 103 compounds analyzed at 300 K, the BTE fails in materials with low thermal conductivity (κTOT≲2\kappa_{\mathrm{TOT}} \lesssim2 W/mK), such as AgBr and AgCl, since the wave-like tunneling conductivity becomes comparable to the particle-like propagation conductivity. We also note that in simple crystals with large conductivity, the BTE is accurate as the particle-like propagation conductivity dominates over the wave-like tunneling conductivity.

Symmetric Relative Mean Error on harmonic and anharmonic vibrational properties

The analyses reported in the main text discuss SRME{K(qs)}\{\mathcal{K}(q_s)\} as a descriptor that is informative of the accuracy of fMLPs in describing microscopic harmonic and anharmonic properties. In this section we provide the tools to resolve whether a large SRME{K(qs)}\{\mathcal{K}(q_s)\} originates from large errors in microscopic harmonic properties, or large errors in anharmonic properties, or from both. To this aim, we employ the expression for the microscopic, single-phonon contributions to the thermal conductivity (see terms inside the square brackets in Eq. (1)) to define descriptors that quantify the accuracy with which harmonic and anharmonic properties are predicted. In particular, we note that the conductivity contribution of a phonon (qsq_s) is a function of: (i) all the phonon frequencies {ω(qs′)}\{\omega(q_{s'})\} at a fixed wavevector qq and variable mode s′=1,…,Nss'=1,\ldots,N_s (NsN_s is the number of phonon bands, equal to 3 times the number of atoms in the crystal’s unit cell[60]); (ii) all velocity-operator elements {v(q)s,s′}\{v(q)_{s,s'}\} at fixed qq, ss and variable s′s'; (iii) all isotopic linewidths at fixed qq and variable s′s', {Γi(qs′)}\{\Gamma^{\mathrm{i}}(q_{s'})\}, which depend solely on harmonic properties, see Eq. (15); (iv) the anharmonic linewidths {Γa(qs′)}\{\Gamma^{\mathrm{a}}(q_{s'})\} at fixed qq and variable s′s'. This can be summarized using the following notation: K(q)s=K(q)s[{ω(q)s′},{v(q)s,s′},{Γi(qs′)},{Γa(qs′)}]\mathcal{K}(q)_s=\mathcal{K}(q)_s[\{\omega(q)_{s'}\},\{v(q)_{s,s'}\},\{\Gamma^{\mathrm{i}}(q_{s'})\},\{\Gamma^{\mathrm{a}}(q_{s'})\}] where the curly brackets denote the set of values of a certain quantity over all the bands at fixed qq (e.g., {ω(q)s′}={ω(q)1,ω(q)2,…,ω(q)Ns}\{\omega(q)_{s'}\}=\{\omega(q)_1,\omega(q)_2,\ldots,\omega(q)_{N_s}\}).

To quantify the impact of errors on the harmonic (har) and anharmonic (anh) properties on the single-phonon conductivity contributions, we define the Symmetric Relative Mean Error (SRME) on these properties as follows:

SRME[har]=2∑qs∣K(q)s[{ωfMLP(q)s′},{vfMLP(q)s,s′},{Γi,fMLP(q)s′},{Γa,DFT(q)s′}]−KDFT(q)s∣∑qs(K(q)s[{ωfMLP(q)s′},{vfMLP(q)s,s′},{Γi,fMLP(q)s′},{Γa,DFT(q)s′}]+KDFT(q)s)(16)\mathrm{SRME}[\mathrm{har}] = 2\frac{\sum_{\boldsymbol{q}s}\left|\mathcal{K}(\boldsymbol{q})_s\left[\{\omega_{\mathrm{fMLP}}(\boldsymbol{q})_{s'}\},\{\boldsymbol{v}_{\mathrm{fMLP}}(\boldsymbol{q})_{s,s'}\},\{\Gamma_{i,\mathrm{fMLP}}(\boldsymbol{q})_{s'}\},\{\Gamma_{a,\mathrm{DFT}}(\boldsymbol{q})_{s'}\}\right]-\mathcal{K}_{\mathrm{DFT}}(\boldsymbol{q})_s\right|}{\sum_{\boldsymbol{q}s}\left(\mathcal{K}(\boldsymbol{q})_s\left[\{\omega_{\mathrm{fMLP}}(\boldsymbol{q})_{s'}\},\{\boldsymbol{v}_{\mathrm{fMLP}}(\boldsymbol{q})_{s,s'}\},\{\Gamma_{i,\mathrm{fMLP}}(\boldsymbol{q})_{s'}\},\{\Gamma_{a,\mathrm{DFT}}(\boldsymbol{q})_{s'}\}\right]+\mathcal{K}_{\mathrm{DFT}}(\boldsymbol{q})_s\right)} \tag*{(16)}
SRME[anh]=2∑qs∣K(q)s[{ωDFT(q)s′},{vDFT(q)s,s′},{Γi,DFT(q)s′},{Γa,fMLP(q)s′}]−KDFT(q)s∣∑qs(K(q)s[{ωDFT(q)s′},{vDFT(q)s,s′},{Γi,DFT(q)s′},{Γa,fMLP(q)s′}]+KDFT(q)s)(17)\mathrm{SRME}[\mathrm{anh}] = 2\frac{\sum_{\boldsymbol{q}s}\left|\mathcal{K}(\boldsymbol{q})_s\left[\{\omega_{\mathrm{DFT}}(\boldsymbol{q})_{s'}\},\{\boldsymbol{v}_{\mathrm{DFT}}(\boldsymbol{q})_{s,s'}\},\{\Gamma_{i,\mathrm{DFT}}(\boldsymbol{q})_{s'}\},\{\Gamma_{a,\mathrm{fMLP}}(\boldsymbol{q})_{s'}\}\right]-\mathcal{K}_{\mathrm{DFT}}(\boldsymbol{q})_s\right|}{\sum_{\boldsymbol{q}s}\left(\mathcal{K}(\boldsymbol{q})_s\left[\{\omega_{\mathrm{DFT}}(\boldsymbol{q})_{s'}\},\{\boldsymbol{v}_{\mathrm{DFT}}(\boldsymbol{q})_{s,s'}\},\{\Gamma_{i,\mathrm{DFT}}(\boldsymbol{q})_{s'}\},\{\Gamma_{a,\mathrm{fMLP}}(\boldsymbol{q})_{s'}\}\right]+\mathcal{K}_{\mathrm{DFT}}(\boldsymbol{q})_s\right)} \tag*{(17)}

where we have used the shorthand notation KDFT(q)s=KDFT(q)s[{ωDFT(q)s′},{vDFT(q)s,s′},{Γi,DFT(q)s′},{Γa,DFT(q)s′}].Intuitively,\mathcal{K}_{\mathrm{DFT}}(\boldsymbol{q})_s = \mathcal{K}_{\mathrm{DFT}}(\boldsymbol{q})_s[\{\omega_{\mathrm{DFT}}(\boldsymbol{q})_{s'}\},\{\boldsymbol{v}_{\mathrm{DFT}}(\boldsymbol{q})_{s,s'}\},\{\Gamma_{i,\mathrm{DFT}}(\boldsymbol{q})_{s'}\},\{\Gamma_{a,\mathrm{DFT}}(\boldsymbol{q})_{s'}\}]. Intuitively,\mathrm{SRME}[\mathrm{har}]$ (16) is large when harmonic vibrational properties (ω(q)s\omega(\boldsymbol{q})_s, or v(q)s,s′\boldsymbol{v}(\boldsymbol{q})_{s,s'}, or Γi(q)s\Gamma_i(\boldsymbol{q})_s) differ between DFT and fMLP, while SRME[anh]\mathrm{SRME}[\mathrm{anh}] (17) is large when the anharmonic linewidths (Γa(q)s\Gamma_a(\boldsymbol{q})_s) differ between DFT and fMLP. We show in Figure 7 that small SRME[har]\mathrm{SRME}[\mathrm{har}] and small SRME[anh]\mathrm{SRME}[\mathrm{anh}] (e.g., as in BeO), imply a small SRE[κ]\mathrm{SRE}[\kappa]. Figure 7 also shows that often large SRME[har]\mathrm{SRME}[\mathrm{har}] and large SRME[anh]\mathrm{SRME}[\mathrm{anh}] (e.g. LiBr) translate into large SRE[κ]\mathrm{SRE}[\kappa]; however there are also cases (e.g. BeTe) in which large SRME[har]\mathrm{SRME}[\mathrm{har}] and large SRME[anh]\mathrm{SRME}[\mathrm{anh}] can also compensate each other and result in a small SRE[κ]\mathrm{SRE}[\kappa]. This confirms that SRE[κ]\mathrm{SRE}[\kappa] is not a reliable descriptor for the capability of fMLPs to capture the harmonic and anharmonic physics underlying heat conduction. In contrast, having both small SRME[har]\mathrm{SRME}[\mathrm{har}] and small SRME[anh]\mathrm{SRME}[\mathrm{anh}] is a sufficient condition to accurately capture harmonic and anharmonic vibrational properties, as well as conductivity.

Overall, these tests illustrate the importance of benchmarking the accuracy of fMLPs in predicting both microscopic harmonic and anharmonic vibrational properties, further motivating the introduction of the SRME[{K(q)s}]\mathrm{SRME}[\{\mathcal{K}(\boldsymbol{q})_s\}].

Influence of isotopic scattering on conductivity

To analyze the influence of isotope scattering on the conductivity, we compare thermal conductivity values with and without considering the linewidths from isotope mass-disorder (Eq. (15)). The results are summarized in Table II. In wurtzite BeO and rocksalt LiBr, the impact of isotope scattering is minimal or small, respectively. However, in zincblende BeTe isotopic and anharmonic linewidths have comparable values (Figure 2b), and therefore both affect the conductivity. In this material, MACE-MP-0 significantly overestimates some linewidths compared to DFT-PBE, and this compensates for the overestimation in other regions. Consequently, MACE-MP-0 shows a significantly larger error in thermal conductivity when isotope scattering is not considered.

wurtzite BeOzincblende BeTerocksalt LiBr
DFTMACEDFTMACEDFTMACE
with Γi(q)s\Gamma_i(\boldsymbol{q})_s286.391259.44378.24669.1381.7890.948
without Γi(q)s\Gamma_i(\boldsymbol{q})_s291.963263.956289.374402.0301.8030.951

Table II. Influence of isotopic scattering on the thermal conductivity at 300 K for BeO, BeTe, and LiBr.

Sensitivity of specific heat to fMLP accuracy

To show that inaccuracies in the PES described by SOTA fMLPs have a small (almost negligible) impact on the specific heat, we consider as paradigmatic cases the materials discussed in Figure 2, namely wurtzite BeO, zincblende BeTe and rocksalt LiBr. Figure 8 presents a comparison of the specific heat computed from DFT and zero-shot MACE-MP-0 for these materials, illustrating that specific heat — and more broadly, phonon spectrum–derived observables — exhibit reduced sensitivity to phonon band structure and hence to the accuracy of the model. This can be intuitively understood from the expression for the specific heat:

Specific heat of wurtzite BeO, zincblende BeTe, and rocksalt LiBr

Figure 8. Specific heat of a) wurtzite BeO, b) zincblende BeTe c) rocksalt LiBr. The zero-shot MACE-MP-0 predictions are in good agreement with DFT confirming that optical phonon mode softening has small (unimportant) effect on the specific heat.

C=1VNc∑qsℏ2ω2(q)skBT2N‾(q)s[N‾(q)s+1](18)C = \frac{1}{V N_c} \sum_{q s} \frac{\hbar^2 \omega^2(q)_s}{k_B T^2} \overline{N}(q)_s\left[\overline{N}(q)_s+1\right] \tag*{(18)}

where q\mathbf{q} and ss are the phonon wavevector and mode index, VV is the unit cell volume and NcN_c the number of wavevectors appearing in the sum; ω(q)s\omega(q)_s is the phonon energy, and N‾(q)s=[exp⁡(ℏω(q)s/kBT)−1]−1\overline{N}(q)_s=[\exp(\hbar\omega(q)_s/k_B T)-1]^{-1} the Bose-Einstein distribution at temperature TT. Eq. (18) shows that at finite TT the quantum specific heat is more sensitive to low-energy (acoustic) phonon modes and becomes decreasingly less sensitive to modes with increasingly higher frequency (e.g., optical phonon modes). SOTA fMLPs describe relatively well the low-energy phonons and suffer from a systematic softening of optical phonons[36], and because the latter inaccuracy involves predominantly high-energy phonons that are unimportant for the specific heat, the specific heat is reasonably well described by SOTA fMLPs already at the zero-shot level.

Automated Wigner conductivity workflow

The workflow, outlined in Figure 9, includes structure relaxation, force constant computation, thermal conductivity evaluation and analysis. To automate the calculation of interatomic forces, we developed an interface between

Automated Wigner Conductivity Workflow

Figure 9. Automated Wigner Conductivity Workflow. The first part of the workflow, involving force evaluation, is performed on a GPU. The input structures undergo relaxation in two stages. In the first stage, initial symmetries are enforced while simultaneously relaxing atomic positions and cell parameters. In the second stage, the structure is further relaxed by removing the symmetry constraint. If the symmetry group of the final structure changes, the structure with enforced symmetry relaxation is used instead. After initializing the displacement structures, the force sets are calculated for each displacement. CPUs handle the remainder of the workflow. Force constants are constructed from the displacement force sets, and thermal conductivity is calculated using phono3py. Structures with imaginary phonon frequencies are discarded from the final analysis (conducted on a PC).

the packages ase[116] (version 3.23.0) and phono3py[47, 117] (version 3.2.0). Atomic positions and unit-cell lattice parameters were simultaneously relaxed to minimize energy, forces, and stresses. We note that the determination of irreducible q\mathbf{q} points needed for thermal conductivity calculations depends on the crystal symmetry, and to evaluate SRME{K(q)s}\{\mathcal{K}(q)_s\} (Eq. (4)) it is essential to compare structures that have the same crystal symmetry when studied using DFT or fMLPs.

We constructed a two-stage relaxation protocol, employing the FretchetCellFilter and the FIRE algorithms in both stages to simultaneously optimize atomic positions and cell parameters. We used force convergence threshold of 10−410^{-4} eV/Å and a maximum number of steps of 300. During the first stage, symmetries are explicitly enforced as constraints, whereas in the second stage, instead, the constraint on the crystal symmetry is removed, to test whether the fMLP correctly preserves the physical crystal symmetry or not. If the symmetry is preserved, the final relaxed structure is used. If symmetry is broken, the structure from the first stage, with enforced symmetry, is retained. We detect symmetries utilizing the spglib package[118] with the default precision parameters. As discussed in Figure 4c, in some cases retaining the symmetric structure yielded structural instabilities (imaginary phonon frequencies), which disappeared when the structure was allowed to break the symmetry and relax. While these structurally stable and broken-symmetry structures are not useful for evaluating SRME[{K(q)}\{\mathcal{K}(\boldsymbol{q})\}], they provide information about shortcomings of fMLPs that could addressed in future studies or used as benchmark metric for accuracy in the description of crystal structures.

In the supercell force-constant calculations, we used the same parameters used in the DFT reference data[47, 48]. Thermal conductivities were computed using phono3py following Refs.[47, 117]. A 19×19×1919 \times19 \times19 q\boldsymbol{q}-mesh was used for rocksalt and zincblende structures, and a 19×19×1519 \times19 \times15 q\boldsymbol{q}-mesh for wurtzite structures. The collision operator was computed using the tetrahedron method. For Figure 2, linewidths were calculated on a 9×9×99 \times9 \times9 q\boldsymbol{q}-mesh, for graphical clarity. We used version 3 of phono3py for anharmonic linewidth calculations and employed the Gonze method[119] for the NAC term, as implemented in the phonopy[65, 117] package (version 2.26.6). For consistency with the phononDB-PBE dataset, in our data release we also include the conductivity results computed with the Wang method[120]. When structures with imaginary phonon frequencies were found with mp-fMLPs, these were discarded from subsequent thermal-conductivity analysis.

Automated Grüneisen parameter workflow

The workflow to compute Grüneisen consists of structure relaxation, force constant computation for three structures (equilibrium volume, volume expanded 1%, and volume contracted 1%), Grüneisen parameter calculation, and analysis. The structure relaxation is carried out using the same two-stage protocol described in the previous section. To generate the configurations at volume expanded or contracted by 1%, we scale the unit cells while keeping the relative atomic positions fixed, then relax the structures using the same procedure, with the cell volume held fixed at each step. Harmonic force constants for the three volumes are computed using 0.03 Å displacements and the same supercells used in the conductivity workflow. The phonopy package[65, 117] is then used to calculate the mode Grüneisen parameters over the same q\boldsymbol{q}-mesh as in the conductivity calculations.

DFT reference and zero-shot mp-fMLP calculations

Computational details of the DFT calculations underlying the phononDB-PBE database can be found in Refs.[47, 48]. Applying the same settings on the same set of materials, we computed the Grüneisen parameters by calculating the phonon band structures the same way as detailed in the previous section. These calculations employed the Perdew, Burke, and Ernzerhof (PBE) exchange-correlation functional[58], consistent with the approach used for generating the MD dataset[45], used to train mp-fMLPs.

Figure 10 shows that in wurtzite BeO, the zero-shot MACE-MP-0-L2 calculation has a computational cost that is more than 3 orders of magnitude lower than the conventional DFT calculation, and yields a conductivity compatible within 7% with DFT reference values (Figure 2a). Importantly, we note that speedups much stronger than those highlighted in Figure 10 are expected in disordered materials without symmetries[50, 110]. In fact, BeO has a simple unit cell containing just 4 atoms and with 12 crystal symmetries[122], which allow to reduce the number of displacements[117]. In contrast, in disordered materials symmetries are broken and consequently a larger number of finite-displacement calculations are needed. Moreover, standard DFT approaches scale with

the cube of the number of atoms, while fMLPs have a linear scaling. Based on these considerations, we expect even stronger speedups in disordered materials[123, 124, 125, 126, 127, 128].

We note that the DFT parameters used to generate the phononDB-PBE database are similar but not always exactly equal to those used for DFT calculations in the MP database. In the regime of computational convergence, variations of the DFT parameters at fixed functional are expected to yield unimportant effects on the conductivity. In particular, Ref.[129] shows that for the prototypical case of silicon, computing force constants using DFT parameters within the range of ‘computationally converged values’ found in the literature produces conductivity variations that are within 7 %, and also shows that in ideal cases varying DFT parameters within the domain of computational convergence should yield conductivity variations smaller than 2 %. As such, the significant discrepancies shown in Figure 3 are expected to be not caused by possible small differences in the DFT parameters.

The computational details on the mp-fMLPs can be found in the following references: M3GNet[21], CHGNet[22], MACE-MP-0[23], SevenNet[24], and ORB-v1-MPtraj[46]. M3GNet was used through the matgl[130] (version 1.1.2) package with the M3GNet-MP-2021.2.8-PES model. The CHGNet version 0.3.8 was used. The MACE-MP-0 2024-01-07-mace-128-L2 model was run through LAMMPS[131]. SevenNet was used through the sevenn package (version 0.9.2) with the SevenNet-0_11July2024 model. ORB-v1-MPtraj was used through the orb-models package (version 0.3.0, commit d142854) with the ORB-v1-MPtraj-only model. The details on the violin plots presented in Figure 4 can be found in Ref.[132].

The ORB-v1-MPtraj, SevenNet, MACE, and CHGNet calculations were performed on an Nvidia Tesla A100 SXM4 80GB GPU, with the relaxation and force calculations for all 103 structures costing approximately ∼3\sim3 GPU hours per model. M3GNet was executed on AMD EPYC 7702 CPUs, with relaxation and force calculations costing around ∼30\sim30 CPU hours. Thermal conductivity calculations for all 103 structures required about ∼50\sim50 CPU hours per model, and similar resources were needed for evaluating thermal conductivity from the DFT force constants. Overall, the analysis of these five models required approximately ∼12\sim12 GPU hours and ∼330\sim330 CPU hours.

Details on fine-tuning

In this section we discuss how the size of the dataset and the number of fine-tuning epochs influence the training error for energies, forces, and stresses in LiBr. Results are shown for different fine-tuning datasets in Figure 11. The three datasets used for fine-tuning contain DFT single-point calculations in 4×4×44 \times4 \times4 supercells of LiBr (same size used for the phonon calculation in the phononDB-PBE[47, 48] database) perturbed from equilibrium with atomic rattling and in some cases also volume variations from 0.25% to 5%. All DFT calculations are performed employing the Perdew, Burke, and Ernzerhof (PBE) exchange-correlation functional[58], and the projector augmented wave (PAW) formalism for electronic minimization, implemented within the VASP (Vienna ab initio simulation package) code[133, 134]. Following Ref.[48], we use a 600 eV plane-wave-basis cutoff (20% higher than the default one), and the Brillouin zone is sampled at the Gamma point. Using these DFT settings, the

LiBr structure in the phononDB-PBE database shows a residual negative stress of 960 Bar. Therefore, to make the most accurate possible comparison between fine-tuned MACE-MP-0 and reference DFT data, we relaxed the fine-tuned MACE-MP-0 structure exactly to the same residual stress present in the reference DFT-PBE data.

The MACE-MP-0 2024-01-07-mace-128-L2 model https://github.com/ACEsuit/mace-mp/commits/mace_mp_0 was fine-tuned using the mace_run_train script in the main branch of https://github.com/ACEsuit/mace. The training started from the last checkpoint of MACE-MP-0 L2. The three datasets discussed above were used to perform three independent trainings. For trainings on the 1-frame and 3-frame datasets we employed a batch size of 1, and for trainings on the 20-frame dataset we used a batch of size 4. Since our fine-tuning datasets were computed with the DFT-PBE functional, which was also employed to generate the MP database[45] used to train MACE-MP-0, in the fine-tuning we kept the E0s fixed at their original MACE-MP-0 value. Representative examples of training scripts are available at: https://github.com/MSimoncelli/fMLP_conductivity.git. The fine-tuning was performed on Nvidia Tesla A100 SXM4 80GB GPU.

Hybrid fMLP-DFT approach to include non-analytical correction term. We propose a hybrid fMLP-DFT method to incorporate the non-analytical correction (NAC) term[59] computed from density functional perturbation theory (DFPT)[75, 135] into phonon band structures predicted by fMLPs. To limit the computational cost, we neglect changes in the Born effective charges that may originate from small structural differences between the DFT-relaxed and fMLP-relaxed primitive cell; the accuracy of this approximation will be discussed a posteriori. We demonstrate the effectiveness of this hybrid approach in LiBr, using MACE-MP-0 potential fine-tuned on 3 frames for over 200 epochs, and combining them with DFT-quality Born effective charges from the phononDB-PBE database[48]. As shown in Figure 12, the hybrid approach with fine-tuned fMLPs and DFT Born effective charges is in good agreement with the DFT reference calculation, confirming the accuracy of our hybrid fMLP-DFT approach, and, a posteriori, the validity of the approximations performed therein.

Figure 12. Hybrid fMLP-DFT phonon band structure including non-analytical correction (NAC) term. The hybrid fMLP-DFT approach shows good agreement with reference band structure obtained from DFT. As an example, we used the MACE-MP-0 model fine-tuned on 3 frames trained over 200 epochs.

Influence of Born effective charges on phonon dispersion, conductivity, and Grüneisen parameter. The long-range non-analytical correction (NAC) term to the phonon band structure can be described using two methodologies: (i) the one proposed by Gonze and Lee[59] using density-functional perturbation theory in reciprocal space, or using the mixed-space approach by Wang et al.[120]. To evaluate the numerical difference between these two methods, we employed both of them to calculate the phonon band structures of LiBr, also checking how these two methods numerically converge with the size of the supercell used (all the other parameters were kept fixed and consistent parameters to those of the phononDB-PBE dataset). Figure 13 shows that increasing the supercell size from 2×2×22 \times2 \times2 to 6×6×66 \times6 \times6 yields a strong effect in the band structures that do not include the NAC term (this was expected and is explained in detail in Ref.[120]), and a noticeable effect in the band structure computed accounting for NAC with Wang’s method. In contrast, Gonze’s method shows significantly smaller variations with respect to varying supercell size, displaying acceptable numerical convergence already in supercells of size 4×4×44 \times4 \times4. Therefore, in the plots of this work we account for NAC using the approach by Gonze and Lee[42].

Figure 13. Numerical performance of different approaches to account for the effect of the long-range NAC term on phonon dispersion. We show the phonon band structure of LiBr without non-analytical correction (NAC) (a), considering NAC using Wang’s[120] approach (b) or Gonze’s[59] method (c). These three cases are compared in simulations that use supercells of different sizes.

In Figure 14 we show how thermal conductivity and Grüneisen parameter change considering or not the NAC term, reporting the cumulative distribution of the relative variation for all the 103 materials analyzed. Overall, the materials analyzed do not show a significant dependence on considering or not the NAC term. In particular, the maximum variation for the Grüneisen parameter is 1.4% in rocksalt AgCl, while for conductivity is 11.8% in wurtzite CdSe. The cumulative distribution of the relative NAC effect on conductivity and Grüneisen parameter is presented in Figure 14.

Cumulative distribution of the non-analytical correction term's effect on conductivity and Grüneisen parameter

Figure 14. Cumulative distribution of non-analytical correction term’s effect on conductivity and Grüneisen parameter, and comparison between the methods by Gonze and Lee[59], and Wang et al.[120]

fMLPs to describe thermal expansion and bulk modulus within the quasi harmonic approximation. To demonstrate that fine-tuned fMLPs can accurately describe not only thermal conductivity and Grüneisen parameter, but also other technologically relevant thermomechanical properties, we employed the fine-tuned fMLP to compute the volumetric thermal expansion and temperature-dependent bulk modulus of LiBr using free energy minimization within the quasi-harmonic approximation (QHA), as implemented in the phonopy package[65]. The capability of fine-tuned fMLPs to describe multiple different thermomechanical properties provides an advantage in terms of computational cost over the following established simulation approaches: (i) compressive sensing[64, 94], which require separate DFT datasets at contracted and expanded unit-cell volumes to describe both conductivity and Grüneisen parameter; (ii) traditional non-foundation machine learning potentials, which rely on much larger training datasets to reach accuracy comparable to fine-tuned fMLPs[110].

The DFT reference data were generated with the same computational settings as the fine-tuning dataset, applying volume variations of ±1%\pm1\%, ±2%\pm2\%, ±3%\pm3\%, ±4%\pm4\%, and ±5%\pm5\%, alongside 0.03 displacements for interatomic force constants and the default Vinet equation of state.

constants and the default Vinet equation of state. The MACE-MP-0 model was evaluated both in a zero-shot setting and after fine-tuning on three frames over 500 epochs. Figure 15 presents the resulting volumetric thermal expansion coefficients and temperature-dependent bulk modulus compared to the DFT reference. The fine-tuned MACE-MP-0 model closely reproduces the DFT results for thermal expansion within 1.5% and for bulk modulus within 3% across the full temperature range, whereas the zero-shot model significantly overestimates the thermal expansion coefficient and underestimates the bulk modulus. Experimental bulk modulus values[136] were plotted for comparison. We note that the bulk modulus exhibits a clear temperature dependence over the plotted range, emphasizing that thermal expansion cannot be divided into independent single-phonon contributions. This is due to the temperature dependence of both the bulk modulus[43] and the Grüneisen parameter, the phonon-dependent terms appearing in the expression for the volumetric thermal expansion, β=γCVBTV\beta= \frac{\gamma C_{V}}{B_{T}V}, as discussed in the main text.

Figure 15. Volumetric thermal expansion and temperature-dependent bulk modulus of LiBr within the quasi-harmonic approximation (QHA) from free energy minimization. Fine-tuning the MACE-MP-0 model on three supercell frames for 500 epochs yields volumetric thermal expansion coefficients in close agreement with DFT reference values across the full temperature range considered.

Simultaneous fine-tuning for multiple materials

To demonstrate the applicability and generalizability of fine-tuning fMLPs for thermomechanical properties, we simultaneously fine-tuned models for three distinct materials: wurtzite AlP, zincblende BN, and rocksalt LiBr, targeting both thermal conductivity and the Grüneisen parameter. For each material, three structural configurations were used: a rattled supercell (atomic displacements sampled with a standard deviation of 0.1 Å), and two unit cells uniformly scaled by ±1%\pm1\% in volume. This setup enhances the accuracy of stress predictions and yields relaxed structures that more closely align with DFT references.

Figure 16 shows that the fMLP effectively learns across multiple materials in parallel, improving accuracy while reducing the per-observable training cost. Fine-tuning was performed using the same settings as for LiBr, on a Nvidia Tesla A100 SXM4 80GB GPU.

Figure 16. Improving fMLP accuracy via simultaneous fine-tuning of three materials. The thermal conductivity, Grüneisen parameter and the corresponding SRME values for wurtzite AlP, zincblende BN and rocksalt LiBr after fine-tuning simultaneously on 3 frames (1 rattled supercell and 2 isotropically scaled unit cell) against the number of epochs.

Data availability. The Matbench Discovery results are available at https://matbench-discovery.materialsproject.org/. The phononDB-PBE dataset including the displacements generated by phono3py and the corresponding force sets are available at https://github.com/atztogo/phonondb[47, 48]. The dataset needed to reproduce the findings of this study is available on the Materials Cloud Archive[137]: https://doi.org/10.24435/materialscloud:2d-4b; such dataset contains thermal conductivity resolved for each phonon mode, as well as data for fine-tuning.

Code availability. The phono3py and phonopy packages are available at https://github.com/phonopy/; the ase package is available at https://gitlab.com/ase/ase; the spglib package is available at https://github.com/spglib/spglib. matgl containing the M3GNet model is available at https://github.com/materialsvirtuallab/matgl; the CHGNet model and package is available at https://github.com/CederGroupHub/chgnet; the SevenNet model and package is available at https://github.com/MDIL-SNU/SevenNet; the ORB-v1-MPtraj model and package is available at https://github.com/orbital-materials/orb-models. The MACE used through LAMMPS[131] is available at https://github.com/ACEsuit/lammps, and the MACE-MP-0 model is available at https://github.com/ACEsuit/mace-mp. The Hiphive package is available at https://gitlab.com/materials-modeling/hiphive. The source code used to reproduce the results presented in this work is publicly available. The k_SRME code, which enables the thermal conductivity benchmarks, is accessible at https://github.com/MPA2suite/k_SRME. The gamma_SRME code, used for computing Grüneisen parameter benchmarks, is available at https://github.com/MPA2suite/gamma_SRME.

Acknowledgements. The computational resources were provided by: (i) the Kelvin2 HPC platform at the NI-HPC Centre (funded by EPSRC and jointly managed by Queen’s University Belfast and Ulster University); (ii) the UK National Supercomputing Service ARCHER2, for which access was obtained via the UKCP consortium and funded by EPSRC [EP/X035891/1]. M. S. acknowledges support from Gonville and Caius College. P.A. acknowledges support from SNSF through Post.Doc mobility fellowship P500PN_206693. We thank William C. Witt, Ilyes Batatia, Andrew D. P. Smith, and Janosh Riebesell for the useful discussions. We gratefully acknowledge Atsushi Togo for having provided harmonic and anharmonic force constants computed using density functional theory (PBE functional) from the phononDB-PBE database[47, 48].

Author contributions

Author contributions. M.S. conceived and supervised the project. B.P. developed the automated computational framework and performed the numerical calculations with inputs from P.A., G.C., and M.S. The results were analyzed and organized by B.P. and M.S., with inputs from P.A. All authors contributed to discussing the results, writing and editing the manuscript.

Competing interests

Competing interests. C.G. has equity interest of Symmetric Group LLP that licenses force fields commercially and also in Ångstrom AI. The other authors declare that they have no competing interest.

Inclusion & ethics statement

Inclusion & ethics statement. All the authors of this study have fulfilled the criteria for authorship re-

References

quired by Nature Portfolio journals, and their participation was essential for the design and implementation of the study. Roles and responsibilities were discussed and agreed among collaborators. Local and regional research relevant to our study has been considered in the citations.

References

  1. [1]T. B. Blank, S. D. Brown, A. W. Calhoun, and D. J. Doren, The Journal of Chemical Physics 103, 4129 (1995).
  2. [2]J. Behler and M. Parrinello, Physical Review Letters 98, 146401 (2007).
  3. [3]A. P. Bartók, M. C. Payne, R. Kondor, and G. Csányi, Physical Review Letters 104, 136403 (2010).
  4. [4]M. Rupp, A. Tkatchenko, K.-R. Müller, and O. A. von Lilienfeld, Physical Review Letters 108, 058301 (2012).
  5. [5]L. Zhang, J. Han, H. Wang, R. Car, and W. E., Physical Review Letters 120, 143001 (2018).
  6. [6]R. Drautz, Physical Review B 99, 014104 (2019).
  7. [7]A. Seko, A. Togo, and I. Tanaka, Physical Review B 99, 214108 (2019).
  8. [8]V. L. Deringer, A. P. Bartók, N. Bernstein, D. M. Wilkins, M. Ceriotti, and G. Csányi, Chemical Reviews 121, 10073 (2021).
  9. [9]S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nature Communications 13, 2453 (2022).
  10. [10]E. Kocer, T. W. Ko, and J. Behler, Annual Review of Physical Chemistry 73, 163 (2022).
  11. [11]X. Qian, J. Zhou, and G. Chen, Nature Materials 20, 1188 (2021).
  12. [12]G. F. Nataf, S. Volz, J. Ordonez-Miranda, J. Íñiguez González, R. Rural, and B. Dkhil, Nature Reviews Materials, 1 (2024).
  13. [13]R. Okabe, A. Chrottanapituk, A. Boonkird, N. Andrejevic, X. Fu, T. S. Jaakkola, Q. Song, T. Nguyen, N. Drucker, S. Mu, Y. Wang, B. Liao, Y. Cheng, and M. Li, Nature Computational Science 4 (2024).
  14. [14]Z. Guo, P. Roy Chowdhury, Z. Han, Y. Sun, D. Feng, G. Lin, and X. Ruan, npj Computational Materials 9, 1 (2023).
  15. [15]J. Qijh, M. Al-Fahdi, Y. Yao, J. Hu, and M. Hu, Journal of Materials Chemistry A 12, 8502 (2024).
  16. [16]A. Rodriguez, C. Lin, H. Yang, M. Al-Fahdi, C. Shen, K. Choudhary, Y. Zhao, J. Hu, B. Cao, H. Zhang, et al., npj Computational Materials 9, 20 (2023).
  17. [17]Y. Srivastava and A. Jain, Journal of Applied Physics 134 (2023).
  18. [18]V. D. Rose, A. Kozachinskiy, C. Rojas, M. Petrache, and P. Barceló, Three iterations of $(d-1)$-$WL test distinguish non isometric clouds of $d$-dimensional points (2023).
  19. [19]J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, in Proceedings of the 34th International Conference on Machine Learning (PMLR, 2017) pp. 1263–1272.
  20. [20]I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, and G. Csányi, MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields (2023).
  21. [21]C. Chen and S. P. Ong, Nature Computational Science 2 (2022).
  22. [22]B. Deng, P. Zhong, K. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder, Nature Machine Intelligence 5, 1031 (2023).
  23. [23]I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kovács, J. Riebesell, X. R. Advincula, M. Asta, M. Avaylon, W. J. Baldwin, F. Berger, N. Bernstein, A. Bhowmik, S. M. Blau, V. Cărare, J. P. Darby, S. De, F. Della Pia, V. L. Deringer, R. Elijošiūš, Z. El-Machachi, F. Falcioni, E. Fako, A. C. Ferrari, A. Gneiritch-Schriever, J. George, R. E. A. Goodall, C. P. Grey, P. Grigorev, S. Han, W. Handley, H. H. Heenen, K. Hermansson, C. Holm, J. Jaafar, S. Hofmann, K. S. Jakob, H. Jung, V. Kapil, A. D. Kaplan, N. Karimitaraji, J. R. Kermode, N. Kroupa, J. Kullgren, M. C. Kuner, D. Kurylа, G. Lliopunite, J. T. Margraf, I.-B. Magdău, A. Michaelides, J. H. Moore, A. A. Naik, S. P. Niblett, S. W. Norwood, N. O’Neill, C. Ortner, K. A. Persson, K. Reuter, A. S. Rosen, L. L. Schaaf, C. Schran, B. X. Shi, E. Sivonxay, T. K. Stenczel, V. Svahn, C. Sutton, T. D. Swinburne, J. Tilly, C. van der Oord, E. Varga-Umbrich, T. Vegge, M. Vondrák, Y. Wang, W. C. Witt, F. Zills, and G. Csányi, A foundation model for atomistic materials chemistry (2024), arXiv:2401.00096.arxiv.org/abs/2401.00096
  24. [24]Y. Park, J. Kim, S. Hwang, and S. Han, Journal of Chemical Theory and Computation 20, 4857 (2024).
  25. [25]H. Yang, C. Hu, Y. Zhou, X. Liu, Y. Shi, J. Li, G. Li, Z. Chen, S. Chen, C. Zeni, M. Horton, R. Pinsler, A. Fowler, D. Zügner, T. Xie, J. Smith, L. Sun, Q. Wang, L. Kong, C. Liu, H. Hao, and Z. Lu, MatterSim: A deep learning atomistic model across elements, temperatures and pressures (2024), arXiv:2405.04967 [cond-mat.mtrl-sci].arxiv.org/abs/2405.04967
  26. [26]A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Nature 624, 80 (2023).
  27. [27]K. Choudhary and B. DeCost, npj Computational Materials 7, 185 (2021).
  28. [28]C. Chen, W. Ye, Y. Zuo, C. Zheng, and S. P. Ong, Chemistry of Materials 31, 3564 (2019).
  29. [29]L. Ward, R. Liu, A. Krishma, V. I. Hegde, A. Agrawal, A. Choudhary, and C. Wolverton, Physical Review B 96, 024104 (2017).
  30. [30]Y. Zuo, M. Qin, C. Chen, W. Ye, X. Li, J. Luo, and S. P. Ong, Materials Today 51, 126 (2021).
  31. [31]T. Xie and J. C. Grossman, Physical Review Letters 120, 145301 (2018).
  32. [32]J. Gibson, A. Hire, and R. G. Hennig, npj Computational Materials 8, 211 (2022).
  33. [33]R. E. A. Goodall, A. S. Parackal, F. A. Faber, R. Armiento, and A. A. Lee, Science Advances 8, eabn4117 (2022).
  34. [34]J. Riebesell, R. E. A. Goodall, P. Benner, Y. Chiang, B. Deng, G. Ceder, M. Asta, A. A. Lee, A. Jain, and K. A. Persson, Nature Machine Intelligence 7, 836
  35. [35]H. Yu, M. Giantomassi, G. Materzanini, J. Wang, and G.-M. Rignanese, Materials Genome Engineering Advances 2, e58 (2024), e-print: https://onlinelibrary.wiley.com/doi/pdf/10.1002/mega.58.
  36. [36]B. Deng, Y. Choi, P. Zhong, J. Riebesell, S. Anand, Z. Li, K. Jun, K. A. Persson, and G. Ceder, npj Computational Materials 11, 9 (2025), publisher: Nature Publishing Group.
  37. [37]H. Lee, V. I. Hegde, C. Wolverton, and Y. Xia, Materials Today Physics 53, 101688 (2025).
  38. [38]M. Simoncelli, N. Marzari, and F. Mauri, Nature Physics 15, 809 (2019).
  39. [39]M. Simoncelli, N. Marzari, and F. Mauri, Physical Review X 12, 041011 (2022).
  40. [40]E. T. Ritz, S. J. Li, and N. A. Benedek, Journal of Applied Physics 126, 171102 (2019).
  41. [41]T. Barron, J. Collins, and G. White, Advances in Physics 29, 609 (1980).
  42. [42]M. Born and K. Huang, Dynamical Theory of Crystal Lattices, Oxford Classic Texts in the Physical Sciences (Oxford University Press, Oxford, New York, 1998).
  43. [43]D. C. Wallace, Thermodynamics of Crystals (Dover Publications, 1972).DOI
  44. [44]J. Riebesell, R. E. A. Goodall, P. Benner, Y. Chiang, B. Deng, A. A. Lee, A. Jain, and K. A. Persson, Matbench-discovery website (2024).DOI
  45. [45]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, APL Materials 1, 011002 (2013).
  46. [46]O. M. Ltd., orb-models, https://github.com/orbital-materials/orb-models (2024).
  47. [47]A. Togo, L. Chaput, and I. Tanaka, Phys. Rev. B 91, 094306 (2015).
  48. [48]A. Seko, A. Togo, H. Hayashi, K. Tsuda, L. Chaput, and I. Tanaka, Physical Review Letters 115, 205901 (2015).
  49. [49]M. Simoncelli, N. Marzari, and A. Cepellotti, Phys. Rev. X 10, 011019 (2020).
  50. [50]M. Simoncelli, F. Mauri, and N. Marzari, npj Computational Materials 9, 1 (2023).
  51. [51]A. Pazhedath, L. Bastonerо, N. Marzari, and M. Simoncelli, Phys. Rev. Appl. 22, 022064 (2024).
  52. [52]N. P. Padтure, Nature Materials 15, 804 (2016).
  53. [53]H. Lehmann, D. Pitzer, G. Pracht, R. Vassen, and D. Stöver, Journal of the American Ceramic Society 86, 1338 (2003), e-print: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1151-2916.2003.tb03473.x.
  54. [54]R. Vassen, X. Cao, F. Tietz, D. Basu, and D. Stöver, Journal of the American Ceramic Society 83, 2023 (2004).
  55. [55]J. J. Plata, P. Nath, D. Unsamzaj, J. Carrete, C. Toher, M. de Jong, M. Asta, M. Fornari, M. B. Nardelli, and S. Curtarolo, npj Computational Materials 3, 1 (2017).
  56. [56]F. Knoop, T. A. Purcell, M. Scheffler, and C. Carbogno, Physical Review Letters 130, 236301 (2023).
  57. [57]S. Bandi, C. Jiang, and C. A. Marianetti, Machine Learning: Science and Technology (2024).
  58. [58]J. P. Perdew, K. Burke, and M. Ernzerhof, Physical Review Letters 77, 3865 (1996).
  59. [59]X. Gonze and C. Lee, Physical Review B 55, 10355 (1997).
  60. [60]J. M. Ziman, Electrons and phonons: the theory of transport phenomena in solids (Oxford university press, 1960).DOI
  61. [61]A. Togo, L. Chaput, and I. Tanaka, Phys. Rev. B 91, 094306 (2015).
  62. [62]A. D. Kaplan, R. Liu, J. Qi, T. W. Ko, B. Deng, J. Riebesell, G. Ceder, K. A. Persson, and S. P. Ong, A Foundational Potential Energy Surface Dataset for Materials (2025), arXiv:2503.04070 [cond-mat].arxiv.org/abs/2503.04070
  63. [63]B. Hakansson and R. G. Ross, Journal of Physics: Condensed Matter 1, 3977 (1989).
  64. [64]F. Eriksson, E. Fransson, and P. Erhart, Advanced Theory and Simulations 2, 1800184 (2019).
  65. [65]A. Togo, Journal of the Physical Society of Japan 92, 012001 (2023).
  66. [66]N. Mounet, M. Gibertini, P. Schwaller, D. Campi, A. Merkys, A. Marrazzo, T. Sohier, I. E. Castelli, A. Cepellotti, G. Pizzi, and N. Marzari, Nature Nanotechnology 13, 246 (2018), publisher: Nature Publishing Group.
  67. [67]Q. Liang, S. Dwaraknath, and K. A. Persson, Scientific Data 6, 135 (2019).
  68. [68]M. Bagheri and H.-P. Komsa, Scientific Data 10, 80 (2023).
  69. [69]M. Uhrin, S. P. Huber, J. Yu, N. Marzari, and G. Pizzi, Comput. Mater. Sci. 187, 110086 (2021).
  70. [70]D. Roy, S. C. Mandal, and B. Pathak, ACS Applied Materials & Interfaces 13, 56151 (2021).
  71. [71]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, APL Materials 1, 011002 (2013).
  72. [72]X. Fu, B. M. Wood, L. Barroso-Luque, D. S. Levine, M. Gao, M. Dzamba, and C. L. Zitnick, Learning smooth and expressive interatomic potentials for physical property prediction (2025), arXiv:2502.12147 [physics] version: 1.arxiv.org/abs/2502.12147
  73. [73]B. Rhodes, S. Vandenhaute, V. Šimkus, J. Gin, J. Godwin, T. Duignan, and M. Neumann, Orb-v3: atomistic simulation at scale (2025), arXiv:2504.06231 [cond-mat].arxiv.org/abs/2504.06231
  74. [74]EGIP: Another step forward in ML-first materials | Radical AI (2025).
  75. [75]S. Baroni, S. de Gironcoli, A. Dal Corso, and P. Giannozzi, Reviews of Modern Physics 73, 515 (2001).
  76. [76]J. Li, Z. Chen, Q. Wang, H. Yang, Z. Lu, G. Li, S. Chen, Y. Zhu, X. Liu, J. Tan, M. Tang, Y. Zhou, C. Zeni, A. Fowler, D. Zügner, R. Pinsler, M. Horton, T. Xie, T.-Y. Liu, H. Liu, T. Qin, B. Lv, D. Donadio, and H. Hao, Probing the limit of heat transfer in inorganic crystals with deep learning (2025).arxiv.org/abs/2503.11568
  77. [77]J. Carrete, B. Vermeersch, A. Katre, A. van Roekeghem, T. Wang, G. K. Madsen, and N. Mingo, Computer Physics Communications 220, 351 (2017).
  78. [78]T. Tadano, Y. Gohda, and S. Tsuneyuki, J. Phys. Condens. Matter 26, 225402 (2014).
  79. [79]G. Barbalinardo, Z. Chen, N. W. Lundgren, and D. Donadio, Journal of Applied Physics 128, 135104 (2020).
  80. [80]L. Paulatto, F. Mauri, and M. Lazzeri, Physical Review B 87, 214303 (2013).
  81. [81]G. Fugallo, M. Lazzeri, L. Paulatto, and F. Mauri, Phys. Rev. B 88, 045430 (2013).
  82. [82]A. Cepellotti, J. Coulter, A. Johansson, N. S. Fedorova, and B. Kozinsky, Journal of Physics: Materials 5, 035003 (2022).
  83. [83]I. Savić, D. Donadio, F. Gygi, and G. Galli, Applied Physics Letters 102 (2013).
  84. [84]S.-i. Tamura, Phys. Rev. B 27, 858 (1983).
  85. [85]C. J. Foss and Z. Aksamija, Physical Review Materials 4, 124006 (2020).
  86. [86]K. R. Hahn, C. Melis, F. Bernardini, and L. Colombo, Frontiers in Mechanical Engineering 7, 712989 (2021).
  87. [87]O. Abou El Kheir, L. Bonati, M. Parrinello, and M. Bernasconi, npj Computational Materials 10, 33 (2024).
  88. [88]D. A. Stewart, I. Savić, and N. Mingo, Nano Letters 9, 81 (2009).
  89. [89]A. Togo and A. Seko, The Journal of Chemical Physics 160, 211001 (2024).
  90. [90]T. Tadano, Y. Gohda, and S. Tsuneyuki, Journal of Physics: Condensed Matter 26, 225402 (2014).
  91. [91]W. Li, J. Carrete, N. A. Katcho, and N. Mingo, Computer Physics Communications 185, 1747 (2014).
  92. [92]O. Hellman and I. A. Abrikosov, Physical Review B 88, 144301 (2013).
  93. [93]J. Dragašević and M. Simoncelli, arXiv:2303.12777 (2023).arxiv.org/abs/2303.12777
  94. [94]E. Di Lucente, M. Simoncelli, and N. Marzari, Physical Review Research 5, 033125 (2023).
  95. [95]A. Jain, Phys. Rev. B 102, 201201 (2020).
  96. [96]Y. Xia, V. I. Hegde, K. Pal, X. Hua, D. Gaines, S. Patel, J. He, M. Aykol, and C. Wolverton, Phys. Rev. X 10, 041029 (2020).
  97. [97]X. Shen, N. Ouyang, Y. Huang, Y.-H. Tung, C.-C. Yang, M. Faizan, N. Perez, R. He, A. Sotnikov, K. Willa, C. Wang, Y. Chen, and E. Guilmeau, Advanced Science 11, 10.1002/advs.202400258 (2024).
  98. [98]T. Pandey, M.-H. Du, D. S. Parker, and L. Lindsay, Materials Today Physics 28, 100881 (2022).
  99. [99]A. Fiorentino and S. Baroni, Physical Review B 107, 054311 (2023).
  100. [100]S. Zhou, E. Xiao, H. Ma, K. Gofryk, C. Jiang, M. E. Manley, D. H. Hurley, and C. A. Marianetti, Physical Review Letters 132, 160502 (2024).
  101. [101]J. Zheng, C. Lin, C. Lin, G. Hautier, R. Guo, and B. Huang, npj Computational Materials 10 (2024).
  102. [102]F. Jia, S. Zhao, J. Wu, L. Chen, T.-H. Liu, and L.-M. Fu, Angewandte Chemie 135, e202315642 (2023).
  103. [103]T. Tadano and W. A. Saidi, Physical Review Letters 129, 185901 (2022).
  104. [104]T. Bernges, R. Hanus, B. Wankmiller, K. Imasato, S. Lin, M. Ghidiu, M. Gerlitz, M. Peterlechner, S. Graham, G. Hautier, Y. Pei, M. R. Hansen, G. Wilde, G. J. Snyder, J. George, M. T. Agne, and W. G. Zeier, Advanced Energy Materials 12, 2200717 (2022).
  105. [105]X. Yang, J. Tiwari, and T. Feng, Materials Today Physics 24, 100689 (2022).
  106. [106]Z. Zeng, X. Shen, R. Cheng, O. Perez, N. Ouyang, Z. Fan, P. Lemoine, B. Raveau, E. Guilmeau, and Y. Chen, Nature Communications 15 (2024).
  107. [107]G. Caldarelli, M. Simoncelli, N. Marzari, F. Mauri, and L. Benfatto, Physical Review B 106, 024312 (2022).
  108. [108]Y. Xia, D. Gaines, J. He, K. Pal, Z. Li, M. G. Kanatzidis, V. Ozoliņš, and C. Wolverton, Proceedings of the National Academy of Sciences 120, e2302541120 (2023).
  109. [109]Y. Luo, X. Yang, T. Feng, J. Wang, and X. Ruan, Nature Communications 11, 2554 (2020).
  110. [110]A. F. Harper, K. Iwanowski, W. C. Witt, M. C. Payne, and M. Simoncelli, Physical Review Materials 8, 043601 (2024).
  111. [111]Y. Liu, H. Liang, L. Yang, G. Yang, H. Yang, S. Song, Z. Mei, G. Csányi, and B. Cao, Advanced Materials 35, 2210873 (2023).
  112. [112]M. Ndour, P. Jund, and L. Chaput, Journal of Non-Crystalline Solids 621, 122618 (2023).
  113. [113]N. W. Lundgren, G. Barbalinardo, and D. Donadio, Physical Review B 103, 024204 (2021).
  114. [114]S. Kielar, C. Li, H. Huang, R. Hu, C. Slebodnick, A. Alatas, and Z. Tian, Nature Communications 15, 6981 (2024).
  115. [115]M. Simoncelli, D. Fournier, M. Marangolo, E. Balan, K. Bén eut, B. Baptiste, B. Doisneau, N. Marzari, and F. Mauri, Temperature-invariant heat conductivity from compensating crystalline and glassy transport: from the Steinbach meteorite to furnace bricks (2024).arxiv.org/abs/2405.13161
  116. [116]A. H. Larsen, J. J. Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P. C. Jennings, P. B. Jensen, J. Kermode, J. R. Kitchin, E. L. Kolsbjerg, J. Kubal, K. Kaasbjerg, S. Lyngsgaard, J. B. Maronsson, T. Maxson, T. Olsen, L. Pastewka, A. Peterson, C. Rostgaard, J. Schiøtz, O. Schütt, M. Strange, K. S. Thygesen, T. Vegge, L. Vilhelmsen, M. Walter, Z. Zeng, and K. W. Jacobsen, Journal of Physics: Condensed Matter 29, 273002 (2017).
  117. [117]A. Togo, L. Chaput, T. Tadano, and I. Tanaka, Journal of Physics: Condensed Matter 35, 353001 (2023).
  118. [118]A. Togo and I. Tanaka, $\mathrm{Spglib}$: a software library for crystal symmetry search (2018).
  119. [119]X. Gonze and C. Lee, Physical Review B 55, 10355 (1997).
  120. [120]Y. Wang, J. J. Wang, W. Y. Wang, Z. G. Mei, S. L. Shang, L. Q. Chen, and Z. K. Liu, Journal of Physics: Condensed Matter 22, 202201 (2010).
  121. [121]The Cambridge Service for Data Driven Discovery (CSD3), cost of CPU and GPU hours (2024-09-02).
  122. [122]A. Togo, K. Shinohara, and I. Tanaka, Science and Technology of Advanced Materials: Methods 0, 2384822 (2024).
  123. [123]L. Reicht, L. Legenstein, S. Wieser, and E. Zojer, Molecules 29, 3724 (2024).
  124. [124]I. Kriuchevskyi, V. V. Palyulin, R. Milkus, R. M. Elder, T. W. Sirk, and A. Zaccone, Phys. Rev. B 102, 024108 (2020).
  125. [125]H. de Araujo Oliveira, Z. Fan, A. Harju, and L. F. C. Pereira, ACS Applied Nano Materials (2024).
  126. [126]E. Bosoni, D. Campi, D. Donadio, G. C. Sosso, J. Behler, and M. Bernasconi, Journal of Physics D: Applied Physics 53, 054001 (2020).
  127. [127]F. DeAngelis, M. G. Muraleedharan, J. Moon, H. R. Seyf, A. J. Minnich, A. J. H. McGaughey, and A. Henry, Nanoscale and Microscale Thermophysical Engineering 23, 81 (2019).
  128. [128]M. Guerboub, S. D. Wansi Wendji, C. Massobrio, A. Bouzid, M. Boero, G. Ori, and E. Martin, The Journal of Chemical Physics 158 (2023).
  129. [129]A. Jain and A. J. McGaughey, Computational Materials Science 110, 115 (2015).
  130. [130]T. W. Ko, M. Nassar, S. Miret, E. Liu, and S. P. Ong, Zenodo (2023), https://doi.org/10.5281/zenodo.8025189.
  131. [131]S. Plimpton, Journal of Computational Physics 117, 1 (1995).
  132. [132]J. L. Hintze and R. D. Nelson, The American Statistician 52, 181 (1998).
  133. [133]G. Kresse and J. Hafner, Physical Review B 47, 558 (1993).
  134. [134]G. Kresse and J. Furthmüller, Computational Materials Science 6, 15 (1996).
  135. [135]P. Ghosez, J.-P. Michenaud, and X. Gonze, Physical Review B 58, 6224 (1998), publisher: American Physical Society.
  136. [136]B. J. Marshall and C. R. Cleavelin, Journal of Physics and Chemistry of Solids 30, 1905 (1969).
  137. [137]L. Talirz, S. Kumbhar, E. Passaro, A. V. Yakutovich, V. Granata, F. Gargiulo, M. Borrelli, M. Uhrin, S. P. Huber, S. Zoupanos, C. S. Adorf, C. W. Andersen, O. Schütt, C. A. Pignedoli, D. Passerone, J. Van-deVondele, T. C. Schulthess, B. Smit, G. Pizzi, and N. Marzari, Scientific Data 7, 299 (2020), publisher: Nature Publishing Group.

Paper details

vlnc:W_ac2kybed
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents