Ferroelectric quantum critical point in superconducting hydrides: The case of H 3 ​ S {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} Marco Cherubini, Abhishek Raghav, Michele Casula Abstract H 3 S sulfur hydride has been widely investigated for its high superconducting critical temperature T c T_{c} of 203 K at about p c = 155 p_{c}=155 GPa. Despite being the precursor of superconducting hydrides, a detailed picture of its phase diagram in an extended temperature and pressure range is still missing. To determine it with inclusion of both thermal and quantum effects, we carry out path integral molecular dynamics combined to a MACE neural network potential trained on BLYP density functional theory configurations. The resulting H 3 S phase diagram is characterized by the ferroelectric transition between the Im 3 ¯ \bar{3} m and R3m phases, which originates from a quantum critical point (QCP) located at p QCP ≈ 134 p_{\mathrm{QCP}}\approx 134 GPa. We show that the experimental T c T_{c} peak falls into a paraelectric region of large nuclear quantum fluctuations above the ferroelectric QCP, as measured by local phonon Green’s functions resolved in imaginary time, where fluctuating dipole moments are at play. We study the critical behavior of the system in the proximity of the QCP by a finite-size scaling analysis, showing that it belongs to the 4D Ising universality class. We finally discuss its implications for the superconducting state. Superconductivity was first discovered more than a century ago 51 . Since then, the quest for systems with high superconducting critical temperature ( T_{c} ) has been one of the frontiers in both experimental and theoretical physics, with high values of T_{c} expected in systems containing light atoms 4 ; 33 . In 2015, a T_{c} of 203 K at about 150 GPa was measured in sulfur hydride ( {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} ) 21 ; 23 ; 49 ; 44 ; 52 , exceeding the highest values found in cuprates 9 ; 27 ; 12 . In {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} , the superconducting critical temperature shows a dome-like shape as a function of pressure 21 , similarly to what happens in cuprates 35 ; 43 or in certain dilute metals like {}\mathrm{SrTiO}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}} 50 ; 38 ; 58 . The T_{c} peak at 155 GPa was originally associated with a ferroelectric phase transition 44 from the body-centered-cubic Im \bar{3} m structure to the lower-symmetry trigonal R3m phase 22 ; 26 ; 32 , triggered by the hydrogen (H) atoms displaced from their midpoint position between two flanking sulfur (S) atoms. However, theoretical attempts to study this transition failed to locate it at a pressure compatible with the position of the T_{c} peak, the most reliable theories placing the transition pressure more than 40 GPa below 24 ; 10 ; 57 . Nevertheless, a detailed picture of the {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} phase diagram is still missing, particularly its temperature evolution, which remains largely unexplored. In this Letter, we fill this gap by studying the system over a large temperature and pressure range and by analyzing its quantum criticality in the proximity of the ferroelectric transition. By path integral molecular dynamics calculations (PIMD) with machine learning interatomic potential (MLIP) trained on density functional theory (DFT-BLYP) 8 ; 39 , we unveil the presence of a ferroelectric quantum critical point (QCP) at 134 \pm 2 GPa belonging to the 4D Ising universality class (Fig. 1 ). According to our picture, the dome shape of the T_{c} evolution has a peak falling in the paraelectric Im \bar{3} m phase, and it is located in a region of strong quantum fluctuations surrounding the QCP, with superconductivity possibly enhanced by them. Figure 1: \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{S} phase diagram, computed by PIMD using a MACE-MLIP trained with a DFT-BLYP dataset for L=4 : ferroelectric transition (gray points, dashed line); classical molecular dynamics (MD) transition (red empty circles, solid line); extrapolated QCP (filled purple circle) and ferroelectric line (black squares, solid line); experimental superconducting T_{c} (green diamonds) 21 ; 23 . The colormap refers to the \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} ratio plotted in Fig. 2 (d); it measures the relative intensity of the retarded fluctuations of local moments. Data are interpolated using a weighted sum of radially symmetric basis functions. Blue and red shaded areas are the ferroelectric and paraelectric phases, respectively. Figure 2: Order parameter and local observables for L=4 . Panel (a): Order parameter for the ferroelectric transition. Full and empty symbols are the absolute and the real values of the order parameter, \Delta_{\textrm{abs}} and \Delta , respectively. The variance of \Delta is in panel (b). Different temperatures are reported: 50 K (black triangles), 100 K (pink diamonds), 150 K (cyan circles), 200 K (orange squares) and 300 K (green reversed triangles). The vertical dashed lines indicate the transition pressures. Panel (c): Proton centroids distribution along the S-S direction for several pressures at T=50 K. Thick black and red curves indicate the ferroelectric transition and the unimodal-to-bimodal variation, respectively. Panel (d): \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} ratio, where \sigma^{2}_{g(\tau)} is the variance of the local Green function in imaginary time. Dashed lines indicate its maximum. In systems containing light atoms, like {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} , nuclear quantum fluctuations play a pivotal role. In this work, we use the PIMD formalism to treat nuclear quantum effects (NQE) exactly. However, its computational cost increases significantly as the temperature is lowered. Therefore, the combination of PIMD with ab initio electronic structure methods, such as DFT, to explore the full phase diagram of sulfur hydride is unfeasible. Here, to maintain the DFT-level accuracy at a much lower computational cost, we trained a MLIP using MACE 7 ; 6 . Then, we ran NVT-PIMD simulations at different volumes V and temperatures T=1/(k_{B}\beta) , using the i-PI package 16 in combination with the generated MLIP. Quantum nuclei are necklaces made of L_{\tau} beads, such that \beta=L_{\tau}\,\delta\tau , with \delta\tau the imaginary time step assuring convergence. The simulations are up to 400 ps long, for systems made of L\times L\times L primitive cells with L\in\{2,3,4\} , necessary for an accurate determination of phase boundaries via finite-size scaling. Further details about the PIMD simulations and the MLIP generation can be found in the Supplemental Material (SM) 1 . The ferroelectric Im \bar{3} m and paraelectric R3m structures differ by the average protons position with respect to their neighboring S atoms. To distinguish the two symmetries, we define the global proton displacement \Delta^{(j)}=\frac{1}{N_{\mathrm{H}}}\sum_{i=1}^{N_{\mathrm{H}}}\bigl[\vec{r}^{(j)}_{\mathrm{H}_{i}\mathrm{S}_{i_{1}}}\cdot\hat{r}^{(j)}_{\mathrm{S}_{i_{1}}\mathrm{S}_{i_{2}}}-d^{(j)}_{\mathrm{S}_{i_{1}}\mathrm{S}_{i_{2}}}/2], (1) evaluated at the j -th PIMD iteration. For every hydrogen atom \mathrm{H}_{i} in the supercell, we identify its two flanking S atoms, \mathrm{S}_{i_{1}} and \mathrm{S}_{i_{2}} . Then, we project the vector \vec{r}_{\mathrm{H}_{i}\mathrm{S}_{i_{1}}} , connecting \mathrm{H}_{i} to one of the two S atoms, onto the \mathrm{S}_{i_{1}} - \mathrm{S}_{i_{2}} direction. Finally, the distance between the projection and the \mathrm{S}_{i_{1}} - \mathrm{S}_{i_{2}} midpoint is summed over the N_{\mathrm{H}} hydrogen atoms in the supercell. \mathrm{H}_{i} , \mathrm{S}_{i_{1}} and \mathrm{S}_{i_{2}} are taken here as centroid positions. Equivalent results can be obtained for bead positions (see SM 1 ). To appropriately define a scalar order parameter for the ferroelectric transition, we establish a reference for the positive direction of these displacements. To do so, we project them onto the eight different degenerate orientations that the {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} molecule can assume in sulfur hydride, and we take the one that maximizes the order parameter \Delta , defined as the average of the global displacements over the entire PIMD trajectory made of N steps, i.e., \Delta=\frac{1}{N}\sum_{j=1}^{N}\Delta^{(j)} . For this orientation, we compute also \Delta_{\textrm{abs}}=\frac{1}{N}\sum_{j=1}^{N}|\Delta^{(j)}| . All global displacements \Delta^{(j)} belonging to the same PIMD trajectory are computed with respect to the same reference orientation. To better identify the ferroelectric transition, we exploit a combined analysis of \Delta and \Delta_{\textrm{abs}} (Fig. 2 (a)). At the transition, the protons freeze in one of the degenerate minima of the underlying potential energy surface (PES). When this happens, the order parameter \Delta and its absolute value \Delta_{\textrm{abs}} become indistinguishable, because the fluctuations are suppressed. Simultaneously, the variance of the order parameter, \sigma^{2}_{\Delta} , suddenly drops (Fig. 2 (b)). By employing these two criteria, i.e. order parameter saturation and frozen fluctuations, we unambiguously identify the ferroelectric transition line as a function of temperature (Fig. 1 ). The critical pressure increases as the temperature is lowered, with a pressure-versus-temperature slope that decreases upon cooling the system. A ferroelectric quantum critical point (QCP) emerges at T=0 K, where the transition is purely driven by quantum fluctuations. Here, the PIMD formalism cannot be used to describe the system. However, we can identify the location of the QCP at p_{\mathrm{QCP}}\approx 134\pm 2 GPa by extrapolating in temperature and size. To further characterize the \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{S} ferroelectric transition, we can also look at the local proton displacement, or local dipole moment \Delta^{(j)}_{i}=\vec{r}^{(j)}_{\mathrm{H}_{i}\mathrm{S}_{i_{1}}}\cdot\hat{r}^{(j)}_{\mathrm{S}_{i_{1}}\mathrm{S}_{i_{2}}}-d^{(j)}_{\mathrm{S}_{i_{1}}\mathrm{S}_{i_{2}}}/2, (2) defined for the centroid coordinates, as in Eq. 1 . Its distribution, gathered for all atoms H_{i} and PIMD iterations, is shown in Fig. 2 (c) for T=50 K and different pressures. In a scenario characterized by a displacive transition, the H atoms display very different behaviors according to the underlying PES shape. Deep in the paraelectric phase, at very high pressures, the potential exhibits a single well and the H atoms stay on average at the S-S segment midpoint, resulting in a unimodal symmetric distribution.Upon decompression, the potential acquires a displaced global minimum which is eight-fold degenerate. In the distribution of the local parameter \Delta^{(j)}_{i} , this is reflected in the development of a bimodal shape, or peak-split density, whose lobes correspond to the positive and negative \mathrm{H}_{i} projections along the corresponding \mathrm{S}_{i_{1}} - \mathrm{S}_{i_{2}} direction. It is the regime of pre-formed local dipole moments, fluctuating across a central barrier. This corresponds to increased fluctuations of the order parameter, detected by \sigma^{2}_{\Delta} (Fig. 2 (b)), while \Delta is still zero on average. Upon further decompression, the potential barrier increases to the point where protons can no longer cross it, freezing in one of the degenerate minima; the distribution of the local proton displacement returns unimodal but shifted from the S-S midpoint. This leads to the creation of a permanent local dipole moment, generated by the displaced protons in the R3m phase. Interestingly, although \Delta^{(j)}_{i} is a local property, it can be used as an additional probe for the ferroelectric transition, because in \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{S} the local moments, once frozen, are long-range ordered. The other temperatures show the same phenomenology: by reducing the pressure, we observe first a crossover from unimodal to bimodal \Delta^{(j)}_{i} distributions with zero order parameter \Delta , and then an off-centered unimodal shape at the ferroelectric transition (see SM 1 ). The density peak splitting associated with the local moment formation does not show a significant temperature dependence. The ferroelectric transition line approaches the local moment formation pressures only at the lowest temperatures, owing to reduced thermal fluctuations. At T=200 K the ferroelectric critical pressure is 124 \pm 2 GPa, significantly lower than the experimental T_{c} peak located at about 155 GPa. This mismatch is aligned with other theoretical predictions 10 ; 57 . Moreover, a comparison with classical MD results, also reported in Fig. 1 , shows that the ferroelectric transition pressure is strongly reduced by NQE ( \approx 50 GPa shift from the classical value at T=200 K). Figure 3: Soft optical mode frequencies as a function of pressure at different temperatures computed for L=4 from PIMD phonons 47 (see also the SM 1 ). Colors and symbols are the same as Fig. 2 (a)-(b). Dashed vertical lines indicate the ferroelectric transition. One of the properties most sensitive to structural transitions is lattice dynamics, encoded by the imaginary-time phonon Green function: G_{x_{i},x_{j}}(\tau)=-\sqrt{m_{i}m_{j}}\langle{\cal T}\delta x_{i}(\tau)\delta x_{j}(0)\rangle, (3) with m_{i} the mass of the i- th atom, \cal{T} the time-ordered operator and \delta x_{i}(\tau) the displacement operator evaluated at the imaginary time \tau along the direction x . The quantum correlator \langle\delta x_{i}(\tau)\delta x_{j}(0)\rangle is accessible by PIMD, where the atomic displacement, \delta x_{i}(\tau)=x_{i}(\tau)-\langle x_{i}\rangle , is defined with respect to \langle x_{i}\rangle , the quantum thermal average of the coordinate x_{i} , and evaluated on the L_{\tau} points in the imaginary time interval [0,\beta[ . Within PIMC, one can easily compute the Kubo-transformed version of Eq. 3 , which corresponds to the static limit of the Matsubara Green function 47 . Diagonalizing its inverse yields the anharmonic phonons frequencies in the static approximation 47 ; 48 . Dispersive phonon branches are obtained by Fourier transforming the spatial coordinates x . In Fig. 3 , we focus on the optical phonons at the \Gamma point that correspond to the modes driving the ferroelectric transition. Their patterns are a combination of H shuttling modes along the S-S direction and S-H-S bending modes, leading to the formation of distinct \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{S} molecules in the crystal. A phonon softening is apparent when approaching the transition from the paraelectric phase, while a sharp jump takes place at the transition. The phonon softening is related to a vanishing curvature of the PES; the sharp frequency increase is due to a stiffer potential experienced by protons once frozen in the ferroelectric configuration. Note how at each temperature the frequency jump happens exactly at the corresponding critical pressure, as detected from \Delta and its variance \sigma^{2}_{\Delta} . This is the typical picture of a second-order phase transition, which goes through a sign change of the PES curvature. The softening of these modes has already been associated to the ferroelectric transition 10 ; 57 . However, we provide here an unprecedented resolution of their temperature dependence, thanks to the MLIP acceleration and a larger statistics. We characterize the universality class of the transition at the QCP by performing a finite-size scaling analysis. Close to a QCP, observables \mathcal{O} are expected to scale as 54 ; 37 ; 46 \mathcal{O}=L^{-x_{\mathcal{O}}}\tilde{\mathcal{O}}(uL^{1/\nu},L_{\tau}/L^{z}), (4) where u=\frac{V-V_{c}}{V_{c}} is the reduced coupling, with V_{c} the critical volume, and x_{\mathcal{O}} the scaling dimension of the observable \mathcal{O} . \nu is the correlation length exponent, while the dynamical critical exponent z governs the temperature dependence. When applying Eq. 4 to the order parameter, i.e. \mathcal{O}=\Delta , we assume Lorentz invariance ( z=1 ) and vanishing anomalous dimension ( x_{\Delta}=1 ), fulfilled by the Ising universality class 5 ; 54 . In fact, the computed grid of temperatures and supercell sizes, limited by the expensive cost of PIMD simulations, prevents the direct determination of z from our data. Therefore, we assume the validity of the Ising universality class and, in order to check whether our data are compatible with this scenario, we focus on the critical exponent \nu . Given z=1 , we fix the L_{\tau}/L ratio to be a constant, such that the scaling function in Eq. 4 depends only on one variable. Thus, the scaling analysis is performed using simulations at T=50 K ( L_{\tau}=100 , L=4 ), 67 K ( L_{\tau}=75 , L=3 ), and 100 K ( L_{\tau}=50 , L=2 ) for different volumes. To obtain \nu , we look for the value that makes L\Delta best collapse onto a unique curve (Fig. 4 ). We identify it by minimizing the cost function C(\nu)=\sum_{j}|y_{j+1}-y_{j}|/(\textrm{max}(y_{j})-\textrm{min}(y_{j}))-1 , where y_{j} are the values of the order parameter for all temperatures, volumes and sizes, ordered according to the uL^{1/\nu} product 56 ; 3 ; 46 . We found that the optimal value is \nu\sim 0.483\pm 0.014 (see inset of Fig. 4 ), consistent with the 4D Ising universality class ( \nu=\frac{1}{2} ). Accordingly, we accurately extrapolated the ferroelectric transition line to the thermodynamic limit, by including the logarithmic corrections appearing at the upper critical dimension 13 ; 2 ; 36 , such that p_{\mathrm{QCP}}(L)=p_{\mathrm{QCP}}(\infty)+\alpha L^{-2}\ln(L)^{-1/6} at fixed temperature (Fig. 1 and SM 1 ). Figure 4: Finite-size scaling of \Delta at T=50 K (black triangles), 67 K (blue circles) and 100 K (red stars), and supercell sizes chosen to keep L_{\tau}/L constant. We use g[uL^{1/\nu}] as a functional form, where u=\frac{V-V_{c}}{V_{c}} , with V_{c} the critical volume. Inset: the critical exponent \nu is obtained by minimizing the cost function C(\nu) defined in the text. Having established the accurate location of the \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{S} ferroelectric transition and its 4D Ising universality class, we then quantify the magnitude of quantum fluctuations across the transition. This information is encoded in the imaginary time evolution, accessible in PIMD. Given the importance of the shuttling modes (Fig. 3 ) in driving the transition, it is worth studying the full time dependence of the phonon Green function projected on the proton local displacements (Eq. 2 ), g(\tau)\equiv G_{ii}(\tau) , where i is the H atom index. We thus extended the approach of Ref. 47 in order to compute the displacement-displacement correlators in Eq. 3 by fully retaining their imaginary time dependence. In the case of a local moment formation, g(\beta/2) saturates to a finite value, due to the freezing of local quantum fluctuations, while in absence of preformed local moments g(\beta/2)\propto\exp(-k\beta)\simeq 0 . Therefore, the variance of g(\beta/2) , i.e. \sigma^{2}_{g(\beta/2)} , bears information on the formation of stable local moments. More specifically, we compute the normalized variance \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} , in order to isolate the contribution of retardation effects at \tau=\beta/2 from the instantaneous moments formation at \tau=0 . Fig. 2 (d) shows the pressure dependence of \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} for all temperatures studied. Its maximum at T=50 K matches the pressures where the unimodal-to-bimodal variation occurs in the local density (Fig. 2 (c)), while the \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} ratio decreases sharply in the ferroelectric phase, when the H atoms freeze into one of the displaced minima, where fluctuations are strongly suppressed. At higher temperatures, the \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} peak is shallower but still well defined, indicating that retardation effects are still relevant, although weaker. Its location is nearly temperature-independent. To have a more comprehensive view of the retardation effects, in Fig. 1 we plot \sigma^{2}_{g(\beta/2)}/\sigma^{2}_{g(0)} as a heatmap in the (p,T) plane. It shows how the ferroelectric phase is characterized by suppressed quantum fluctuations. On the other hand, large fluctuations, only marginally irrelevant in 4D, are present in the paraelectric side, stretching over a wide (p,T) domain. It is intriguing that the experimental superconducting T_{c} peak is quite far from the ferroelectric transition line, while falling in a paraelectric region with strong retardation effects spawned by the QCP. It is already known that NQE can enhance the superconducting critical temperature of a BCS superconductor, as shown in the atomic phase of pristine hydrogen 20 . This requires going beyond the standard Migdal-Eliashberg framework 42 by taking into account phonon renormalization effects 41 ; 25 through the inclusion of the full frequency dependence of the phonon Green function, beside the multi-band and the k -dependence nature of the electron-phonon interaction 55 ; 40 , and its vertex corrections 53 ; 34 ; 45 . In the case of {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} , source of these effects is the proximity of the ferroelectric QCP, which then calls for a more appropriate evaluation of the superconducting T_{c} in this material. At the same time, it would be desirable to estimate the electron-phonon coupling within a non-perturbative framework, such as the one proposed in Ref. 11 . Indeed, the presence of a ferroelectric QCP has been invoked to explain the enhancement of superconductivity in some dilute metals like {}\mathrm{SrTiO}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}} 50 ; 58 ; 19 ; 28 ; 38 ; 59 , by relying on an electron-pairing mechanism beyond linear order in the electron-phonon vertex, namely mediated by two optical phonons 50 . Although it remains to verify whether {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} meets similar conditions, the application of the latter picture to this material is tempting, owing to the presence of optical phonons with strong spectral weight 15 , significantly softened at the ferroelectric transition (Fig. 3 ), and to the strong renormalization of both one-body and two-body phonon Green functions (Fig. 1 and SM 1 ). To conclude, we have characterized the phase diagram of {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} over a wide range of pressures and temperatures, by unveiling the 4D Ising universality class of its ferroelectric quantum criticality, the location of the QCP and its transition line from BLYP-MLIP driven PIMD. The emerging scenario for the high- T_{c} superconductivity in {}{}\mathrm{H}{\vphantom{\mathrm{X}}}_{\smash[t]{\mathrm{3}}}\mathrm{S} appears more complex than previously thought. Indeed, according to our calculations the maximum of the superconducting T_{c} falls in a paraelectric region dominated by quantum fluctuations and retardation effects surrounding the QCP, which could both affect superconductivity.