Article https://doi.org/10.

1038/s41467-023-38855-1

Realistic phase diagram of water from “first

principles” data-driven quantum simulations
1 1,2,3,4
Received: 10 January 2023 Sigbjørn Løland Bore & Francesco Paesani

Accepted: 12 May 2023

Since the experimental characterization of the low-pressure region of water’s

phase diagram in the early 1900s, scientists have been on a quest to under-
Check for updates
stand the thermodynamic stability of ice polymorphs on the molecular level. In
this study, we demonstrate that combining the MB-pol data-driven many-body
1234567890():,;
1234567890():,;

potential for water, which was rigorously derived from “first principles” and
exhibits chemical accuracy, with advanced enhanced-sampling algorithms,
which correctly describe the quantum nature of molecular motion and ther-
modynamic equilibria, enables computer simulations of water’s phase diagram
with an unprecedented level of realism. Besides providing fundamental
insights into how enthalpic, entropic, and nuclear quantum effects shape the
free-energy landscape of water, we demonstrate that recent progress in “first
principles” data-driven simulations, which rigorously encode many-body
molecular interactions, has opened the door to realistic computational studies
of complex molecular systems, bridging the gap between experiments and
simulations.

Arguably, water is the single most important molecule on Earth, being hydrogen bonded to four other water molecules9. While constraining
an essential component of life1 and being directly involved in several the spatial arrangement of water molecules to tetrahedral geometries,
fundamental biological and chemical processes2. From a scientific a vast space of energetically favorable solids exists.
standpoint, one of the most intriguing aspects of water is the contrast The phase diagram of water keeps expanding with time through
between its simple chemical formula and its complex behavior3. Liquid a close synergy between experiment and simulation. Pioneering
water exhibits several anomalous properties4, including the well- measurements by Angell and co-workers10,11 and subsequent com-
known density maximum at 4 °C which allows fish to thrive at the puter simulations have led to several hypotheses about the exis-
bottom of icy lakes. Ordinary ice, i.e., hexagonal ice (ice Ih), is an tence of a liquid–liquid critical point at deeply supercooled
extraordinary solid5. It has a lower density than liquid water, which temperatures12–15, which have stimulated several experimental
makes ice float on liquid water. Ice is slippery when one walks, skates, measurements for the past two decades16–20. Similarly, while
or just stands still on it, but is sticky when one touches it6. Strictly experiments continue to make progress in exploring the phase
speaking, ordinary ice is not even a crystalline material since it exhibits diagram of water21, with 20 different crystalline ice polymorphs22,23
orientational disorder7, which stabilizes the lattice structure and, and 3 amorphous forms24 discovered to date, computer simulations
consequently, raises the melting point by ~100 K compared to the have generated a plethora of energetically viable ice polymorph
melting points of other similar chemical compounds5. The origin of candidates25. Despite significant advancements in computer simu-
these unusual properties can be traced to the ability of the water lations, reproducing the phase diagram of water experimentally
molecules to form directional hydrogen bonds whose strength and determined by Bridgeman and Taman in the early 1900s26,27 still
orientation fluctuate in time and space depending on temperature and remains a challenge. Current state-of-the-art simulations can only
pressure8. As determined in the 1930s, the structure of ice follows the qualitatively account for the equilibria between liquid water and the
so-called Bernal-Fowler rules, which state that every water molecule is different ice polymorphs28–36. This is symptomatic of the difficulties

1
Department of Chemistry and Biochemistry, University of California, San Diego, La Jolla, CA 92093, USA. 2Materials Science and Engineering, University of
California San Diego, La Jolla, CA 92093, USA. 3Halicioğlu Data Science Institute, University of California San Diego, La Jolla, CA 92093, USA. 4San Diego
Supercomputer Center, University of California San Diego, La Jolla, CA 92093, USA. e-mail: [email protected]

Nature Communications | (2023)14:3349 1

Article https://doi.org/10.1038/s41467-023-38855-1

for existing water models to correctly represent the free-energy combination of machine-learned representations of short-range quan-
landscape of water in regions of the phase diagram that most closely tum-mechanical interactions and mean-field representations of many-
resemble conditions encountered for aqueous solutions on Earth37. body effects95,96. Fully derived from CCSD(T) data, MB-pol accurately
The accuracy of a computer simulation in predicting the proper- predicts structural, thermodynamic, dynamical, and spectroscopic
ties of water across the entire phase diagram depends on the ability of properties of water from gas-phase clusters97–99 to the liquid100–105 and
the model used in the simulation to accurately capture the underlying ice106–109 phases, bypassing the accuracy limitations of DFT-based
molecular interactions, as well as on the extent to which the simulation models. The MB-pol potential is thus uniquely positioned to provide
exhaustively samples the free-energy landscape of water over a wide realistic, molecular-level insights into the phase diagram of water.
range of thermodynamic conditions. On the one hand, the free-energy For a precise determination of the phase diagram of water, equally
landscape of water is particularly complex. For example, the average important to the accurate representation of the underlying molecular
molecular dipole moment of water increases by 30–50% moving from interactions is the exhaustive sampling of the associated free-energy
the gas to the condensed phases38. Furthermore, water molecules can landscape30. The most common approach to characterizing coex-
form highly directional hydrogen bonds whose strength is determined istence equilibria is thermodynamic integration, which allows for cal-
by many-body effects that vary significantly depending on the local culating free-energy differences by performing a series of simulations
three-dimensional structural arrangement39. In addition, due to the that connect a phase of known free energy to the phase of interest110. In
light mass of the hydrogen atoms, the properties of water are modu- a seminal work28, Sanz et al. used thermodynamic integration in
lated by nuclear quantum effects, which are responsible for several combination with the Einstein Molecule method and Gibbs-Duhem
differences in the behavior of light (H2O) and heavy (D2O) water40,41. On integration to calculate the phase diagram of water using different
the other hand, since some ice polymorphs are separated energetically empirical, pairwise-additive models, providing an important bench-
by only 0.06 kJ mol−142,43, computer simulations of water’s phase dia- mark for the ability of computer simulations to reproduce the
gram require highly precise determination of the associated free- experimental phase diagram. It should, however, be noted that cal-
energy landscape. culating the phase diagram of water using the Einstein Molecule
By construction, “first principles” (or ab initio) simulations pro- method is not devoid of challenges. In particular, for proton-
vide the most rigorous, although still approximate, description of a disordered ice polymorphs, the Einstein Molecule method requires
molecular system by solving the corresponding Schrödinger equation exact knowledge of the molecular configuration that minimizes the
“on the fly”44. Different “first principles” methods, however, exhibit associated free energy as determined by the water model used in the
significantly different accuracy and predictive power depending on the simulations. This is a daunting task to accomplish for partially-
approximations that they rely on, ranging from the Hartree–Fock disordered ice phases, such as ice III and ice V, because determining
method45–47, which scales with the fourth power of the number of basis the corresponding minimum free-energy configuration requires
functions (that is proportional to the system size) but neglects electron extremely long simulations due to the extremely slow transition
correlation, to coupled-cluster methods48,49, such as CCSD(T), i.e., a from one configuration to another111,112. Recent simulation studies
coupled-cluster method that includes single, double, and perturbative carried out with the TIP4P/2005 and TIP4P/Ice force fields have
triple excitations, which currently represents the “gold standard” for shown that the Einstein Molecule method largely underestimates
molecular interactions but scales with the seventh power of the the thermodynamic stability of ice III compared to direct-coexistence
number of basis functions50–52. In practice, an accuracy-cost compro- and enhanced-coexistence simulations33,112,113. As discussed in the ori-
mise has to be made in “first principles” simulations. In this context, ginal references33,112,113, since both direct-coexistence and enhanced-
density functional theory (DFT)53,54, which formally scales with the coexistence simulations explicitly simulate the crystallization process,
third power of the number of basis functions, remains the only viable they do not rely on any approximation for the entropic contributions
“first principles” method for computer simulations of condensed- associated with proton disorder. This allows for correctly determining
phase systems55. Besides being still computationally too expensive for the free-energy difference between liquid water and a given ice poly-
a complete exploration of water’s phase diagram, DFT, however, suf- morph, independently of the extent of proton disorder present in the
fers from inherent limitations due to the use of approximate exchange- ice polymorph (i.e., direct-coexistence and enhanced-coexistence
correlation functionals and electron densities56–64, which manifest in simulations inherently sample the relevant regions of the underlying
both functional-driven and density-driven errors65–69. A recent study multidimensional free-energy landscape.)
has shown that even the most accurate DFT models exhibit errors that In this study, we report the phase diagram of water calculated at
are similar in magnitude to the relative differences in lattice energies of the fully quantum-mechanical level using the “first principles” MB-pol
ice polymorphs70. These findings also imply that neural network data-driven many-body potential. Using a multi-stage approach that
potentials of water derived from DFT-based simulations35,71–79, which leverages the computational efficiency of a deep neural network
are gaining popularity in computational molecular sciences, exhibit potential trained on MB-pol data (DNN@MB-pol) and rigorous free-
the same limitations of the parent DFT models. Given the short- energy sampling techniques (see Methods for details), we demonstrate
comings associated with DFT-based simulations, it is thus not sur- that MB-pol reproduces the phase diagram of water with an unprece-
prising that pairwise-additive water models such as TIP4P/200580 and dented level of realism, thus closing the gap between experimental
TIP4P/Ice81, which were empirically parameterized to reproduce a measurements and simulation predictions.
subset of experimental thermodynamic data, still provide some of the
most reasonable representations of the phase diagram of water28–33,82. Results
The development of efficient algorithms for correlated electronic Liquid-ice coexistence
structure methods has recently enabled routine coupled-cluster cal- While it is, in principle, possible to determine each coexistence line
culations of interaction energies for water clusters83,84. This has given from a single melting point of the relevant ice polymorph, we calcu-
rise to a new class of “first principles” data-driven potentials for lated a total of 15 melting points. This allowed us to average the
water85–92 that rigorously decompose the interaction energy of an coexistence lines obtained from all the melting points, resulting in
arbitrary water system into individual many-body contributions93, more accurate estimates. Additionally, it allowed us to compare the
which can be efficiently calculated at the coupled-cluster level of the- melting lines calculated using Gibbs-Duhem integration to the melting
ory. Among these “first principles” data-driven many-body potentials, points directly determined from enhanced-coexistence simulations
MB-pol89–91 exploits the “nearsightedness of electronic matter”94 to using the DNN@MB-pol potential. Figure 1a shows that the melting
accurately describe CCSD(T) interaction energies through a points determined from the enhanced-coexistence simulations lie

Nature Communications | (2023)14:3349 2

Article https://doi.org/10.1038/s41467-023-38855-1

Fig. 1 | Liquid-ice melting points and coexistence lines. Melting points calculated level with MB-pol (c) are indicated by stars, while thermodynamically stable and
at the classical level with DNN@MB-pol (a) and MB-pol (b), and at the quantum metastable line segments are represented by solid and dotted lines, respectively.

Fig. 2 | Phase diagram of water. The phase diagrams calculated at the classical diagram is adapted from ref. 5. The regions of stability for ice Ih, II, III, V, and VI and
level with DNN@MB-pol (a) and MB-pol (b), and at the quantum level with MB-pol liquid water are shown as areas colored in green, orange, gray, brown, pink, and
(c) are compared with the experimental phase diagram (d). The experimental phase blue, respectively.

precisely on the estimated coexistence lines. It should be noted that, the triple points reported in Supplementary Tables 2–7, and then
although the coexistence lines were traced using the DNN@MB-pol determine the DNN@MB-pol and MB-pol phase diagrams shown in
potential, the consistency between the melting lines computed using Fig. 2. The phase diagram calculated at the classical level with the
Gibbs-Duhem integration and the melting points determined from DNN@MB-pol potential (Fig. 2a) correctly locates the regions of sta-
enhanced-coexistence simulations is equally good for the corre- bility of ice Ih, ice II, and ice VI, but does not predict any region of
sponding estimates obtained at both classical (Fig. 1b) and quantum stability for ice III and ice V. In contrast, the MB-pol phase diagram
(Fig. 1c) levels using the MB-pol potential upon applying thermo- (Fig. 2b) obtained at the classical level from thermodynamic pertur-
dynamic perturbation theory and thermodynamic integration by mass, bation of the corresponding DNN@MB-pol phase diagram displays
respectively. The comparisons shown in Fig. 1 thus demonstrate that distinct regions of stability for all ice polymorphs, achieving semi-
directly tracing the melting lines using Gibbs-Duhem integration from quantitative agreement with the experimental phase diagram (Fig. 2c).
single melting points calculated for each ice polymorph is indeed a Accounting for nuclear quantum effects further expands the regions of
reliable approximation. stability associated with ice III and ice V, effectively bringing the MB-
pol phase diagram calculated at the quantum level to a quantitative
Phase diagram agreement with the experimental phase diagram. This trend is con-
Starting from the triple points of the liquid-ice coexistence lines, we sistent with previous observations derived from simulations carried
performed additional Gibbs-Duhem integration calculations to obtain out with different water models32,35, which highlighted the importance

Nature Communications | (2023)14:3349 3

Article https://doi.org/10.1038/s41467-023-38855-1

potential, we hypothesize that the phase diagram of water reported in

ref. 36, which predicts ice III to be thermodynamically unstable, may
benefit from thermodynamic perturbation calculations connecting the
neural network potential to the actual reference DFT model. It should,
however, be noted that the original phase diagram of ref. 36, which was
calculated using the Einstein Molecule method, may also change sig-
nificantly when the free-energy differences are calculated using direct-
coexistence or enhanced-coexistence simulations as discussed in
Supplementary Note 5.
As shown in Fig. 2, accounting for nuclear quantum effects leads
to a quantitative agreement between the MB-pol and experimental
phase diagrams. Figure 3 indicates that the corrections to the dif-
ferences in chemical potential calculated at the quantum-mechanical
level are positive for the liquid-ice Ih and liquid-ice II equilibria,
negligible for the liquid-ice III equilibrium, and negative for the
liquid-ice V and liquid-ice VI equilibria, in line with previous obser-
Fig. 3 | Corrections to chemical potentials. Average corrections calculated from vations based on simulations carried out with the pairwise-additive
thermodynamic perturbation (i.e., classical DNN@MB-pol → classical MB-pol) and TIP4PQ/200532 model and the revPBE0-D3 DFT model35. The differ-
thermodynamic integration by mass (i.e., classical MB-pol → quantum MB-pol), and ent magnitude of the quantum corrections likely depends on the
corresponding net corrections applied to the differences in chemical potentials delicate interplay between competing nuclear quantum effects116 and
between the liquid phase and each ice polymorph. Each error bar is defined as the different hydrogen-bonding topologies of different ice phases117,
variance of the correction over the thermodynamic conditions used in the
which are further modulated by temperature and pressure. The
calculations.
investigation of the different impact that nuclear quantum effects
have on the difference in chemical potential between liquid water
and different ice polymorphs will be the focus of a future study.
of nuclear quantum effects for a correct representation of free-energy Interestingly, the largest shifts between the coexistence lines calcu-
differences involving ice III and ice V. lated at the classical and quantum levels with MB-pol are on the order
The systematic improvement in the description of the phase of ~5 K, which are significantly smaller than the shift of ~20 K
boundaries observed when moving from classical DNN@MB-pol to obtained from simulations with the TIP4PQ/2005 model. These dif-
classical MB-pol simulations, and then from classical MB-pol to quan- ferences can possibly be attributed to the competition between
tum MB-pol simulations provides fundamental insights into the level inter- and intra-molecular nuclear quantum effects116, which is cor-
of accuracy necessary for achieving a realistic representation of the rectly represented by realistic water models such as MB-pol102 but
phase behavior of water. In this regard, Fig. 3 reports the corrections exaggerated by empirical pairwise additive models such as TIP4PQ/
applied to the chemical potential calculated with the DNN@MB-pol 2005118.
potential for each ice polymorph with respect to the liquid phase,
which were necessary to elevate the DNN@MB-pol results to the actual Comparison with current state-of-the-art simulations
MB-pol values (i.e., classical DNN@MB-pol → classical MB-pol, and Figure 4 compares the MB-pol phase diagram calculated at the
classical MB-pol → quantum MB-pol). While DNN@MB-pol demon- quantum-mechanical level with current state-of-the-art phase dia-
strates remarkable consistency with MB-pol for energies and forces grams reported in the literature for various water models. Empirical
calculated for molecular configurations extracted from MB-pol simu- pairwse-additive water models belonging to the TIP4P family (such as
lations carried out over a wide range of thermodynamic conditions TIP4P/200533 and TIP4P/Ice113 shown in Fig. 4a and b, respectively) and
(see Supplementary Note 1), the DNN@MB-pol phase diagram calcu- the polarizable iAMOEBA model34 (Fig. 4c) are able to qualitatively
lated at the classical level (Fig. 2a) is qualitatively different from the capture some features of the experimental phase diagram. However,
corresponding MB-pol phase diagram (Fig. 2b). Figure 3 shows that the none of the regions of stability for the different ice polymorphs is
corrections associated with thermodynamic perturbation calculations correctly represented, except that for ice Ih. In particular, both TIP4P/
that connect DNN@MB-pol to MB-pol (blue bars) overall favor the ice 2005 and TIP4P/Ice largely overestimate the region of stability of ice
polymorphs over the liquid phase. This leads to a significant contrac- III, which consequently leads to the shrinking of the region of stability
tion of the region of stability for the liquid phase and, consequently, for ice VI and pushes the region of stability for ice II down to tem-
provides space for ice III and ice V, improving the agreement with the peratures below 100–150 K33,113. Similar performance is exhibited by
experimental phase diagram. The shifts in the stability of the different iAMOEBA that places the region of stability for ice II and ice VI at
ice polymorphs are due to the high sensitivity of the free-energy significantly lower temperatures (below 200 K) and higher pressures
landscape of water to the level of accuracy achieved in the description (above 1.2 GPa), respectively, compared to the experimental phase
of the underlying molecular interactions, which emphasizes that the diagram.
DNN@MB-pol potential, by construction, is not an exact clone of the Both DFT-based phase diagrams calculated at the classical level
MB-pol potential. In this regard, it has recently been shown that neural with the DNN@SCAN potential (Fig. 4d) and at the quantum level with
network potentials, such as DNN@MB-pol, are intrinsically limited in the revPBE0-D3 model (Fig. 4e) predict ice Ih, ice II, ice V, and ice VI to
their transferability across different phases and thermodynamic con- be stable. However, the predicted regions of stability are significantly
ditions, being unable to correctly represent individual many-body different from those observed experimentally, with the revPBE0-D3
contributions to the underlying energy landscape114, which is particu- model predicting ice VI to be only stable above 1 GPa and below 250 K.
larly important in determining the relative stability of different ice In addition, both DNN@SCAN and revPBE0-D3 do not predict ice III to
phases106,115. be a stable phase, which is in clear disagreement with the experimental
Our results are in line with previous observations of proton- observations. As discussed in Supplementary Note 5, the absence of a
ordered ice II being destabilized relative to the other proton- region of stability for ice III in the DNN@SCAN and revPBE0-D3 phase
disordered ice phases when long-range interactions are properly diagrams is likely an artifact of the Einstein Molecule method28 and the
accounted for35. Based on the results obtained with the MB-pol closely related Debye Crystal method119, respectively. While calculating

Nature Communications | (2023)14:3349 4

Article https://doi.org/10.1038/s41467-023-38855-1

Fig. 4 | Comparison among state-of-the-art simulations. a Classical phase dia- Crystal method. f Quantum phase diagram of MB-pol calculated in this study using
gram of TIP4P/2005 from ref. 33 calculated using direct-coexistence simulations. enhanced-coexistence simulations. The phase diagrams that are shown in
b Classical phase diagram of TIP4P/Ice from ref. 113 calculated using enhanced- a, c, d, and e were digitized from the original references. In each panel, the regions
coexistence simulations. c Classical phase diagram of iAMOEBA from ref. 34 cal- of stability for ice Ih, II, III, V, and VI and liquid water are shown as areas colored in
culated using direct-coexistence simulations. d Classical phase diagram of green, orange, gray, brown, pink, and blue, respectively, and the experimental
DNN@SCAN from ref. 36 calculated using the Einstein Molecule method. phase diagram5 is shown using dotted red lines.
e Quantum phase diagram of revPBE0-D3 from ref. 35 calculated using the Debye

the DNN@SCAN and revPBE0-D3 phase diagrams using direct- diagram. Importantly, the deviations between the MB-pol and
coexistence or enhanced-coexistence simulations will likely result in experimental coexistence lines are always on the order of ~10 K
the appearance of a region of stability for ice III, this will also be (~0.02 kcal/mol), demonstrating that MB-pol consistently predicts
accompanied by the shrinking of the region of stability for ice II, which the properties of “real” water across different phases and thermo-
is already underestimated by both water models. In addition, since the dynamic condition with an accuracy that is well within chemical
region of stability for ice VI is independent of the method used to accuracy (1 kcal/mol)120.
calculate the phase diagram (Supplementary Fig. 30), the revPBE0-D3 Accounting for nuclear quantum effects lowers the melting
phase diagram calculated using direct-coexistence or enhanced- point predicted by MB-pol for ice Ih at 1 atm by 3.9 K. While, on the
coexistence simulations will likely still not be able to predict the cor- absolute temperature scale, this shift results in slightly worse
rect region of stability for ice VI. Figure 4 clearly demonstrates that MB- agreement with the experimental value, the relative difference
pol outperforms all models that have, to date, been used to simulate between the classical and quantum melting points of ice Ih at 1 atm
the low-pressure region of the phase diagram of water. Combined with
previous findings97–109, the comparisons shown in Fig. 4 provide further
Table 1 | Melting point (Tm) and heat of fusion of ice Ih (Hfus) at
support to the notion that MB-pol currently provides the most realistic
1 atm
representation of water across different phases and thermodynamic
conditions. Method Tm/K Hfus/kJ mol−1
H2O, experiment 273.15 6.01
Thermodynamic transferability D2O, experiment 276.95 6.22
While demonstrating remarkable accuracy in predicting the proper- TIP4P/Ice 269.8 5.39
ties of water across the entire phase diagram, MB-pol is still a com-
DNN@SCAN 312 7.6
puter model and, therefore, by definition, does not exactly
H2O, revPBE0-D3 276 6.8
correspond to “real” water. For example, due to a nearly constant
shift of ~10 K in the liquid–ice Ih coexistence line, MB-pol slightly D2O, revPBE0-D3 282 7.4
underestimates the melting points of the ice polymorphs. As a result, H2O, MB-pol 262.3 5.83
the melting points predicted by TIP4P/Ice and revPBE0-D3 for ice Ih Classical, MB-pol 266.2 6.42
at 1 atm appear to be in closer agreement with the experimental value Comparisons between the melting points and the heats of fusion determined from computer
(Table 1). Among all water models, MB-pol, however, clearly displays simulations with TIP4P/Ice122,137, the DNN@SCAN potential121, the revPBE0-D3 model75, and the
better transferability across different phases and is the only model MB-pol PEF. We estimated the revPBE0-D3 values from the updated chemical potentials that
correct a sign error in the original calculations75 as described in ref. 35.
that correctly reproduce the overall shape of the experimental phase

Nature Communications | (2023)14:3349 5

Article https://doi.org/10.1038/s41467-023-38855-1

predicted by MB-pol is in remarkable agreement with the difference Data availability

of 3.8 K between the melting points of D2O and H2O ice Ih measured Example files used for the enhanced-coexistence simulations are
experimentally at 1 atm. This agreement is consistent with the notion available at PLUMED-NEST (https://www.plumed-nest.org), as plu-
that classical simulations more closely describe the behavior of mID:23.001. Input and data files for all simulations, as well as the
heavy water40. Moreover, the MB-pol classical and quantum heats of DNN@MB-pol potential along with the corresponding training set, are
fusion determined at 1 atm from the corresponding chemical available on Zenodo131. Any other data generated and analyzed for this
potentials (Table 1) are within 3% of the experimental values mea- study are available from the authors upon request.
sured for H2O and D2O ice Ih, respectively. To put the MB-pol results
in context, DNN@SCAN overestimates the heats of fusion of H2O and Code availability
D2O ice Ih by 27% and 22%, respectively121, revPBE0-D3 overestimates All analysis files are available at the Paesani Lab data Repository:
the heats of fusion of H2O and D2O ice Ih by 13% and 18%, https://github.com/paesanilab/Data_Repository. MB-pol reference
respectively75, and TIP4P/Ice underestimates the heats of fusion of energies and forces were computed using the MBX software132, which
H2O and D2O ice Ih by 10% and 15%, respectively122. is available at: https://github.com/paesanilab/MBX. The DNN@MB-pol
potential was trained using DeePMD-kit133, which is available at: https://
Discussion github.com/deepmodeling/deepmd-kit. All classical molecular
We have demonstrated that the “first pinciples” MB-pol data-driven dynamics (MD) simulations were carried out with LAMMPS134 patched
many-body potential, which was rigorously derived from the many- with PLUMED135 and DeePMD-kit133. All path-integral molecular
body expansion of the energy calculated at the “gold standard” dynamics (PIMD) simulations were carried out with i-PI136, which is
coupled-cluster level of theory, predicts the low-pressure region of the available at: https://github.com/i-pi/i-pi. Any additional codes not lis-
phase diagram of water in quantitative agreement with experiment, ted here are available from the authors upon request.
exhibiting an unprecedented level of realism for molecular-level
computer simulations. Besides marking an important milestone in References
computer simulations of water, both accuracy and transferability 1. Franks, F. Water: a matrix of life, vol. 21 (Royal Society of Chem-
demonstrated by MB-pol across a wide range of thermodynamic istry, 2000).
conditions provide support for the reliability and validity of MB-pol 2. Bagchi, B. Water in biological and chemical processes: from
simulations of water under conditions that are difficult to access by structure and dynamics to function (Cambridge University
experiments19,123. In this context, the close resemblance of MB-pol to Press, 2013).
the long-sought-after “universal model” of water, as defined in ref. 124, 3. Eisenberg, D. & Kauzmann, W. The structure and properties of
provides a more realistic basis for “in silico” studies of supercooled water (OUP Oxford, 2005).
water and ice nucleation, which have long puzzled the scientific 4. Pettersson, L. G. M., Henchman, R. H. & Nilsson, A. Water – the
community due to inconsistent or inconclusive results from existing most anomalous liquid. Chem. Rev. 116, 7459–7462 (2016).
water models. For example, computer simulations with various water 5. Petrenko, V. F. & Whitworth, R. W. Physics of ice (OUP
models have yielded a wide range of predictions for a possible liquid- Oxford, 1999).
liquid critical point125 and rates of homogeneous ice nucleation122. 6. Rosenberg, R. Why is ice slippery? Phys. Today 58, 50 (2005).
Importantly, since the data-driven many-body formalism originally 7. Pauling, L. The structure and entropy of ice and of other crystals
adopted in the development of the MB-pol potential has recently been with some randomness of atomic arrangement. J. Am. Chem. Soc.
extended to generic molecules126–130, our results also indicate that it will 57, 2680–2684 (1935).
soon be possible to perform realistic molecular simulations of com- 8. Maréchal, Y. The hydrogen bond and the water molecule: the
plex systems, thus bridging the gap between computer modeling and physics and chemistry of water, aqueous and bio-media (Else-
experiments. vier, 2006).
9. Bernal, J. D. & Fowler, R. H. A theory of water and ionic solution,
Methods with particular reference to hydrogen and hydroxyl ions. J. Chem.
The phase diagram of water was calculated with MB-pol using a Phys. 1, 515–548 (1933).
multi-stage approach as described in detail in the Supplementary 10. Speedy, R. & Angell, C. Isothermal compressibility of supercooled
Information. Briefly, we first developed a deep neural network water and evidence for a thermodynamic singularity at -45 oC. J.
potential (DNN@MB-pol) trained on MB-pol data which enables MB- Chem. Phys. 65, 851–858 (1976).
pol-level molecular dynamics simulations at a fraction of the com- 11. Angell, C., Sichina, W. & Oguni, M. Heat capacity of water at
putational cost associated with actual MB-pol simulations. This extremes of supercooling and superheating. J. Phys. Chem. 86,
speedup was critical to enabling extensive enhanced-coexistence 998–1002 (1982).
simulations of the relevant liquid-ice polymorph equilibria, which 12. Speedy, R. J. Stability-limit conjecture. An interpretation of the
would have been otherwise unaffordable with MB-pol. Second, we properties of water. J. Phys. Chem. 86, 982–991 (1982).
calculated the DNN@MB-pol melting points for the relevant ice 13. Poole, P. H., Sciortino, F., Essmann, U. & Stanley, H. E. Phase
polymorphs using enhanced-coexistence simulations which, as dis- behaviour of metastable water. Nature 360, 324–328 (1992).
cussed in the Introduction, are not affected by possible artifacts 14. Poole, P. H., Sciortino, F., Grande, T., Stanley, H. E. & Angell, C. A.
arising from approximate definitions of proton disorder33,112,113. Third, Effect of hydrogen bonds on the thermodynamic behavior of
starting from the DNN@MB-pol melting points, we determined the liquid water. Phys. Rev. Lett. 73, 1632 (1994).
corresponding MB-pol melting points using thermodynamic pertur- 15. Sastry, S., Debenedetti, P. G., Sciortino, F. & Stanley, H. E.
bation theory (Supplementary Note 2). Fourth, we used thermo- Singularity-free interpretation of the thermodynamics of super-
dynamic integration by mass to account for nuclear quantum effects cooled water. Phys. Rev. E 53, 6144 (1996).
in the liquid-ice polymorph equilibria and thus calculate the MB-pol 16. Pallares, G. et al. Anomalies in bulk supercooled water at
quantum melting points for the different ice polymorphs. Finally, we negative pressure. Proc. Natl Acad. Sci. USA 111, 7936–7941
used Gibbs–Duhem integration to trace the coexistence lines con- (2014).
necting all melting points and determine the phase diagrams of water 17. Caupin, F. Escaping the no man’s land: Recent experiments
(Fig. 2) at the classical level with DNN@MB-pol and MB-pol, and at on metastable liquid water. J. Non Cryst. Solids 407, 441–448
the quantum level with MB-pol. (2015).

Nature Communications | (2023)14:3349 6

Article https://doi.org/10.1038/s41467-023-38855-1

18. Kim, K. H. et al. Maxima in the thermodynamic response and 41. Ceriotti, M. et al. Nuclear quantum effects in water and aqueous
correlation functions of deeply supercooled water. Science 358, systems: experiment, theory, and current challenges. Chem. Rev.
1589–1593 (2017). 116, 7529–7550 (2016).
19. Kim, K. H. et al. Experimental observation of the liquid-liquid 42. Whalley, E. Energies of the phases of ice at zero temperature and
transition in bulk supercooled water under pressure. Science 370, pressure. J. Chem. Phys. 81, 4087–4092 (1984).
978–982 (2020). 43. Gillan, M. J., Alfe, D. & Michaelides, A. Perspective: how good is
20. Pathak, H. et al. Enhancement and maximum in the isobaric DFT for water? J. Chem. Phys. 144, 130901 (2016).
specific-heat capacity measurements of deeply supercooled 44. Car, R. & Parrinello, M. Unified approach for molecular dynamics
water using ultrafast calorimetry. Proc. Natl Acad. Sci. USA 118, and density-functional theory. Phys. Rev. Lett. 55, 2471 (1985).
e2018379118 (2021). 45. Hartree, D. R. The wave mechanics of an atom with a non-
21. Salzmann, C. G. Advances in the experimental exploration of Coulomb central field. Part I. Theory and methods. Proc. Camb.
water’s phase diagram. J. Chem. Phys. 150, 060901 (2019). Philos. Soc. 24, 89–110 (1928).
22. Millot, M. et al. Nanosecond X-ray diffraction of shock- 46. Fock, V. Näherungsmethode zur lösung des quantenmecha-
compressed superionic water ice. Nature 569, 251–255 (2019). nischen mehrkörperproblems. Z. Phys. 61, 126–148 (1930).
23. Gasser, T. M., Thoeny, A. V., Fortes, A. D. & Loerting, T. Structural 47. Slater, J. C. Note on hartree’s method. Phys. Rev. 35, 210 (1930).
characterization of ice XIX as the second polymorph related to ice 48. Coester, F. Bound states of a many-particle system. Nucl. Phys. 7,
VI. Nat. Commun. 12, 1–10 (2021). 421–424 (1958).
24. Rosu-Finsen, A. et al. Medium-density amorphous ice. Science 49. Coester, F. & Kümmel, H. Short-range correlations in nuclear wave
379, 474–478 (2023). functions. Nucl. Phys. 17, 477–485 (1960).
25. Zhu, C. et al. Computational prediction of novel ice phases: a 50. Bartlett, R. J. & Musiał, M. Coupled-cluster theory in quantum
perspective. J. Phys. Chem. Lett. 11, 7449–7461 (2020). chemistry. Rev. Mod. Phys. 79, 291 (2007).
26. Tammann, G. Ueber die Grenzen des festen Zustandes IV. Ann. 51. Rezac, J. & Hobza, P. Describing noncovalent interactions beyond
Phys. 307, 1–31 (1900). the common approximations: how accurate is the “gold stan-
27. Bridgman, P. W. Water, in the liquid and five solid forms, under dard,” CCSD (T) at the complete basis set limit? J. Chem. Theory
pressure. Proc. Am. Acad. Arts Sci. 47, 441–558 (1912). Comput. 9, 2151–2155 (2013).
28. Sanz, E., Vega, C., Abascal, J. L. F. & MacDowell, L. G. Phase dia- 52. Rezac, J. & Hobza, P. Benchmark calculations of interaction
gram of water from computer simulation. Phys. Rev. Lett. 92, energies in noncovalent complexes and their applications. Chem.
255701 (2004). Rev. 116, 5038–5071 (2016).
29. Vega, C., Abascal, J., Sanz, E., MacDowell, L. & McBride, C. Can 53. Hohenberg, P. & Kohn, W. Inhomogeneous electron gas. Phys.
simple models describe the phase diagram of water? J. Phys. Rev. 136, B864 (1964).
Condens. Matter. 17, S3283 (2005). 54. Kohn, W. & Sham, L. J. Self-consistent equations including
30. Vega, C., Sanz, E., Abascal, J. & Noya, E. Determination of phase exchange and correlation effects. Phys. Rev. 140, A1133 (1965).
diagrams via computer simulation: methodology and applications 55. Marx, D. & Hutter, J. Ab initio molecular dynamics: basic theory and
to water, electrolytes and proteins. J. Phys. Condens. Matter. 20, advanced methods (Cambridge University Press, 2009).
153101 (2008). 56. Parr, R. G. Density functional theory. Annu. Rev. Phys. Chem. 34,
31. Vega, C., Abascal, J. L., Conde, M. & Aragones, J. What ice can 631–656 (1983).
teach us about water interactions: a critical comparison of the 57. Kohn, W., Becke, A. D. & Parr, R. G. Density functional theory of
performance of different water models. Faraday Discuss. 141, electronic structure. J. Phys. Chem. 100, 12974–12980 (1996).
251–276 (2009). 58. Cohen, A. J., Mori-Sánchez, P. & Yang, W. Insights into current
32. McBride, C., Noya, E. G., Aragones, J. L., Conde, M. M. & Vega, C. limitations of density functional theory. Science 321,
The phase diagram of water from quantum simulations. Phys. 792–794 (2008).
Chem. Chem. Phys. 14, 10140–10146 (2012). 59. Perdew, J. P., Ruzsinszky, A., Constantin, L. A., Sun, J. & Csonka, G.
33. Conde, M. M., Gonzalez, M., Abascal, J. & Vega, C. Determining the I. Some fundamental issues in ground-state density functional
phase diagram of water from direct coexistence simulations: the theory: a guide for the perplexed. J. Chem. Theory Comput. 5,
phase diagram of the TIP4P/2005 model revisited. J. Chem. Phys. 902–908 (2009).
139, 154505 (2013). 60. Cohen, A. J., Mori-Sánchez, P. & Yang, W. Challenges for density
34. Wang, L.-P. et al. Systematic improvement of a classical molecular functional theory. Chem. Rev. 112, 289–320 (2012).
model of water. J. Phys. Chem. B 117, 9956–9972 (2013). 61. Burke, K. Perspective on density functional theory. J. Chem. Phys.
35. Reinhardt, A. & Cheng, B. Quantum-mechanical exploration of the 136, 150901 (2012).
phase diagram of water. Nat. Commun. 12, 588 (2021). 62. Medvedev, M. G., Bushmarinov, I. S., Sun, J., Perdew, J. P. & Lys-
36. Zhang, L., Wang, H., Car, R. & Weinan, E. Phase diagram of a deep senko, K. A. Density functional theory is straying from the path
potential water model. Phys. Rev. Lett. 126, 236001 (2021). toward the exact functional. Science 355, 49–52 (2017).
37. Michard, G. Natural aqueous solutions in the earth. In The physics 63. Mardirossian, N. & Head-Gordon, M. Thirty years of density func-
and chemistry of aqueous ionic solutions, 379–397 tional theory in computational chemistry: an overview and
(Springer, 1987). extensive assessment of 200 density functionals. Mol. Phys. 115,
38. Coulson, C. A. & Eisenberg, D. Interactions of H2O molecules in ice 2315–2372 (2017).
I. The dipole moment of an H2O molecule in ice. Proc. R. Soc. Lond. 64. Verma, P. & Truhlar, D. G. Status and challenges of density func-
291, 445–453 (1966). tional theory. Trends Chem. 2, 302–318 (2020).
39. Xantheas, S. S. Ab initio studies of cyclic water clusters (H2O)n, 65. Mori-Sánchez, P., Cohen, A. J. & Yang, W. Localization and delo-
n = 1–6. II. Analysis of many-body interactions. J. Chem. Phys. 100, calization errors in density functional theory and implications for
7523–7534 (1994). band-gap prediction. Phys. Rev. Lett. 100, 146401 (2008).
40. Paesani, F. & Voth, G. A. The properties of water: Insights 66. Kim, M.-C., Sim, E. & Burke, K. Understanding and reducing errors
from quantum simulations. J. Phys. Chem. B 113, 5702–5719 in density functional calculations. Phys. Rev. Lett. 111,
(2009). 073003 (2013).

Nature Communications | (2023)14:3349 7

Article https://doi.org/10.1038/s41467-023-38855-1

67. Dasgupta, S., Lambros, E., Perdew, J. P. & Paesani, F. Elevating 87. Wang, Y. & Bowman, J. M. Ab initio potential and dipole moment
density functional theory to chemical accuracy for water simula- surfaces for water. II. Local-monomer calculations of the infrared
tions through a density-corrected many-body formalism. Nat. spectra of water clusters. J. Chem. Phys. 134, 154510 (2011).
Commun. 12, 1–12 (2021). 88. Babin, V., Medders, G. R. & Paesani, F. Toward a universal water
68. Sim, E., Song, S., Vuckovic, S. & Burke, K. Improving results by model: first principles simulations from the dimer to the liquid
improving densities: density-corrected density functional theory. phase. J. Phys. Chem. Lett. 3, 3765–3769 (2012).
J. Am. Chem. Soc. 144, 6625–6639 (2022). 89. Babin, V., Leforestier, C. & Paesani, F. Development of a “first
69. Palos, E. et al. Assessing the interplay between functional-driven principles” water potential with flexible monomers: dimer poten-
and density-driven errors in DFT models of water. J. Chem. Theory tial energy surface, VRT spectrum, and second virial coefficient. J.
Comput. 18, 3410–3426 (2022). Chem. Theory Comput. 9, 5395–5403 (2013).
70. Della Pia, F., Zen, A., Alfè, D. & Michaelides, A. DMC-ICE13: Ambi- 90. Babin, V., Medders, G. R. & Paesani, F. Development of a “first
ent and high pressure polymorphs of ice from diffusion Monte principles” water potential with flexible monomers. II: Trimer
Carlo and density functional theory. J. Chem. Phys. 157, potential energy surface, third virial coefficient, and small clus-
134701 (2022). ters. J. Chem. Phys. 10, 1599–1607 (2014).
71. Morawietz, T., Sharma, V. & Behler, J. A neural network potential- 91. Medders, G. R., Babin, V. & Paesani, F. Development of a “first-
energy surface for the water dimer based on environment- principles” water potential with flexible monomers. III. Liquid
dependent atomic energies and charges. J. Chem. Phys. 136, phase properties. J. Chem. Theory Comput. 10, 2906–2910 (2014).
064103 (2012). 92. Yu, Q. et al. q-AQUA: a many-body CCSD(T) water potential
72. Morawietz, T. & Behler, J. A density-functional theory-based neural including four-body interactions demonstrates the quantum nat-
network potential for water clusters including van der Waals ure of water from clusters to the liquid phase. J. Phys. Chem. Lett.
corrections. J. Phys. Chem. A 117, 7356–7366 (2013). 13, 5068–5074 (2022).
73. Morawietz, T., Singraber, A., Dellago, C. & Behler, J. How van der 93. Hankins, D., Moskowitz, J. & Stillinger, F. Water molecule interac-
Waals interactions determine the unique properties of water. Proc. tions. J. Chem. Phys. 53, 4544–4554 (1970).
Natl Acad. Sci. USA 113, 8368–8373 (2016). 94. Prodan, E. & Kohn, W. Nearsightedness of electronic matter. Proc.
74. Morawietz, T. et al. The interplay of structure and dynamics in the Natl Acad. Sci. USA 102, 11635–11638 (2005).
Raman spectrum of liquid water over the full frequency and 95. Reddy, S. K. et al. On the accuracy of the MB-pol many-body
temperature range. J. Phys. Chem. Lett. 9, 851–857 (2018). potential for water: Interaction energies, vibrational frequencies,
75. Cheng, B., Engel, E. A., Behler, J., Dellago, C. & Ceriotti, M. Ab initio and classical thermodynamic and dynamical properties from
thermodynamics of liquid and solid water. Proc. Natl Acad. Sci. clusters to liquid water and ice. J. Chem. Phys. 145, 194504 (2016).
USA 116, 1110–1115 (2019). 96. Paesani, F. Getting the right answers for the right reasons: toward
76. Gartner III, T. E. et al. Signatures of a liquid–liquid transition in an predictive molecular simulations of water with many-body
ab initio deep neural network model for water. Proc. Natl Acad. Sci. potential energy functions. Acc. Chem. Res. 49, 1844–1851 (2016).
USA 117, 26040–26046 (2020). 97. Richardson, J. O. et al. Concerted hydrogen-bond breaking by
77. Wohlfahrt, O., Dellago, C. & Sega, M. Ab initio structure and quantum tunneling in the water hexamer prism. Science 351,
thermodynamics of the RPBE-D3 water/vapor interface by neural- 1310–1313 (2016).
network molecular dynamics. J. Chem. Phys. 153, 144710 (2020). 98. Cole, W. T., Farrell, J. D., Wales, D. J. & Saykally, R. J. Structure and
78. Zhang, C. et al. Modeling liquid water by climbing up Jacob’s torsional dynamics of the water octamer from THz laser spectro-
ladder in density functional theory facilitated by using deep neural scopy near 215 μm. Science 352, 1194–1197 (2016).
network potentials. J. Phys. Chem. B 125, 11444–11456 (2021). 99. Brown, S. E. et al. Monitoring water clusters “melt” through
79. Gartner III, T. E., Piaggi, P. M., Car, R., Panagiotopoulos, A. Z. & vibrational spectroscopy. J. Am. Chem. Soc. 139,
Debenedetti, P. G. Liquid-liquid transition in water from first prin- 7082–7088 (2017).
ciples. Phys. Rev. Lett. 129, 255702 (2022). 100. Medders, G. R. & Paesani, F. Infrared and Raman spectroscopy of
80. Abascal, J. L. F. & Vega, C. A general purpose model for the liquid water through “first-principles” many-body molecular
condensed phases of water: TIP4P/2005. J. Chem. Phys. 123, dynamics. J. Chem. Theory Comput. 11, 1145–1154 (2015).
234505 (2005). 101. Medders, G. R. & Paesani, F. Dissecting the molecular structure of
81. Abascal, J. L. F., Sanz, E., García Fernández, R. & Vega, C. A the air/water interface from quantum simulations of the sum-
potential model for the study of ices and amorphous water: TIP4P/ frequency generation spectrum. J. Am. Chem. Soc. 138,
Ice. J. Chem. Phys. 122, 234511 (2005). 3912–3919 (2016).
82. McBride, C., Vega, C., Noya, E. G., Ramírez, R. & Sesé, L. M. 102. Reddy, S. K., Moberg, D. R., Straight, S. C. & Paesani, F.
Quantum contributions in the ice phases: the path to a new Temperature-dependent vibrational spectra and structure of
empirical model for water – TIP4PQ/2005. J. Chem. Phys. 131, liquid water from classical and quantum simulations with the mb-
024506 (2009). pol potential energy function. J. Chem. Phys. 147, 244504 (2017).
83. Adler, T., Knizia, G. & Werner, H. A simple and efficient CCSD(T)- 103. Moberg, D. R., Straight, S. C. & Paesani, F. Temperature depen-
F12 approximation. J. Chem. Phys. 127, 221106–221106 (2007). dence of the air/water interface revealed by polarization sensitive
84. Knizia, G., Adler, T. B. & Werner, H.-J. Simplified CCSD(T)-F12 sum-frequency generation spectroscopy. J. Phys. Chem. B 122,
methods: theory and benchmarks. J. Chem. Phys. 130, 4356–4365 (2018).
054104 (2009). 104. Muniz, M. C. et al. Vapor–liquid equilibrium of water with the MB-
85. Bukowski, R., Szalewicz, K., Groenenboom, G. C. & Van der Avoird, pol many-body potential. J. Chem. Phys. 154, 211103 (2021).
A. Predictions of the properties of water from first principles. 105. Gartner III, T. E. et al. Anomalies and local structure of liquid water
Science 315, 1249–1252 (2007). from boiling to the supercooled regime as predicted by the many-
86. Wang, Y., Huang, X., Shepler, B. C., Braams, B. J. & Bowman, J. M. body MB-pol model. J. Phys. Chem. 13, 3652–3658 (2022).
Flexible, ab initio potential, and dipole moment surfaces for water. 106. Pham, C. H., Reddy, S. K., Chen, K., Knight, C. & Paesani, F. Many-
I. Tests and applications for clusters up to the 22-mer. J. Chem. body interactions in ice. J. Chem. Theory Comput. 13,
Phys. 134, 094509 (2011). 1778–1784 (2017).

Nature Communications | (2023)14:3349 8

Article https://doi.org/10.1038/s41467-023-38855-1

107. Moberg, D. R., Straight, S. C., Knight, C. & Paesani, F. Molecular 129. Bull-Vulpe, E., Riera, M., Götz, A. & Paesani, F. MB-Fit: software
origin of the vibrational structure of ice Ih. J. Phys. Chem. Lett. 8, infrastructure for data-driven many-body potential energy func-
2579–2583 (2017). tions. J. Chem. Phys. 155, 124801 (2021).
108. Moberg, D. R., Sharp, P. J. & Paesani, F. Molecular-level inter- 130. Bull-Vulpe, E. F., Riera, M., Bore, S. L. & Paesani, F. Data-driven
pretation of vibrational spectra of ordered ice phases. J. Phys. many-body potential energy functions for generic molecules:
Chem. B 122, 10572–10581 (2018). linear alkanes as a proof-of-concept application. J. Chem. Theory
109. Moberg, D. R. et al. The end of ice I. Proc. Natl Acad. Sci. USA 116, Comput. https://doi.org/10.1021/acs.jctc.2c00645 (2022).
24413–24419 (2019). 131. Bore, S. L. & Paesani, F. Realistic phase diagram of water from “first
110. Frenkel, D. & Ladd, A. J. New Monte Carlo method to compute the principles” data-driven quantum simulations. Zenodo https://doi.
free energy of arbitrary solids. Application to the fcc and hcp org/10.5281/zenodo.7863744 (2023).
phases of hard spheres. J. Chem. Phys. 81, 3188–3193 (1984). 132. Paesani group (UC San Diego). MBX: an energy and force calcu-
111. MacDowell, L. G., Sanz, E., Vega, C. & Abascal, J. L. F. Combina- lator for data-driven many-body potential energy functions.
torial entropy and phase diagram of partially ordered ice phases. J. https://github.com/paesanilab/MBX (2021).
Chem. Phys. 121, 10145–10158 (2004). 133. Wang, H., Zhang, L., Han, J. & Weinan, E. DeePMD-kit: a deep
112. Espinosa, J. R. et al. Ice Ih vs. ice III along the homogeneous learning package for many-body potential energy representation
nucleation line. Phys. Chem. Chem. Phys. 21, 5655–5660 (2019). and molecular dynamics. Comput. Phys. Commun. 228,
113. Bore, S. L., Piaggi, P. M., Car, R. & Paesani, F. Phase diagram of the 178–184 (2018).
TIP4P/Ice water model by enhanced sampling simulations. J. 134. Thompson, A. P. et al. LAMMPS – a flexible simulation tool for
Chem. Phys. 157, 054504 (2022). particle-based materials modeling at the atomic, meso, and con-
114. Zhai, Y., Caruso, A., Bore, S. L., Luo, Z. & Paesani, F. A “short tinuum scales. Comput. Phys. Commun. 271, 108171 (2022).
blanket” dilemma for a state-of-the-art neural network potential 135. Tribello, G. A., Bonomi, M., Branduardi, D., Camilloni, C. & Bussi, G.
for water: Reproducing experimental properties or the physics of PLUMED 2: new feathers for an old bird. Comput. Phys. Commun.
the underlying many-body interactions? J. Chem. Phys. 158, 185, 604–613 (2014).
084111 (2023). 136. Kapil, V. et al. i-PI 2.0: a universal force engine for advanced
115. Melko, R. G., den Hertog, B. C. & Gingras, M. J. Long-range order at molecular simulations. Comput. Phys. Commun. 236,
low temperatures in dipolar spin ice. Phys. Rev. Lett. 87, 214–223 (2019).
067203 (2001). 137. Conde, M., Rovere, M. & Gallo, P. High precision determination of
116. Habershon, S., Markland, T. E. & Manolopoulos, D. E. Competing the melting points of water TIP4P/2005 and water TIP4P/Ice
quantum effects in the dynamics of a flexible water model. J. models by the direct coexistence technique. J. Chem. Phys. 147,
Chem. Phys. 131, 024501 (2009). 244506 (2017).
117. Li, X.-Z., Walker, B. & Michaelides, A. Quantum nature of the
hydrogen bond. Proc. Natl Acad. Sci. USA 108, 6369–6373 (2011). Acknowledgements
118. Noya, E. G., Vega, C., Sesé, L. M. & Ramírez, R. Quantum effects on We are grateful to Pablo Piaggi for his aid in setting up the environments,
the maximum in density of water as described by the TIP4PQ/ including both oxygen and hydrogen atoms, for the enhanced-
2005 model. J. Chem. Phys. 131, 124518 (2009). coexistence simulations involving ice II. We thank Bingqing Cheng for
119. Kapil, V., Engel, E., Rossi, M. & Ceriotti, M. Assessment of providing the updated figure of the revPBE0-D3 chemical potentials and
approximate methods for anharmonic free energies. J. Chem. making the i-PI input files for the revPBE0-D3 simulations available on
Theory Comput. 15, 5845–5857 (2019). GitHub, which were of great help in calculating the quantum corrections
120. Pople, J. A. Nobel lecture: quantum chemical models. Rev. Mod. to the chemical potentials, and Carlos Vega and Venkat Kapil for their
Phys. 71, 1267 (1999). comments on the initial version of our manuscript and insights about the
121. Piaggi, P. M., Weis, J., Panagiotopoulos, A. Z., Debenedetti, P. G. & Einstein Molecule and Debye Crystal methods. This research was sup-
Car, R. Homogeneous ice nucleation in an ab initio machine- ported by the Air Force Office of Scientific Research grant no. FA9550-
learning model of water. Proc. Natl Acad. Sci. USA 119, 20-1-0351 (S.L.B. and F.P.). Computational resources were provided by
e2207294119 (2022). the Department of Defense High Performance Computing Moderniza-
122. Espinosa, J., Sanz, E., Valeriani, C. & Vega, C. Homogeneous ice tion Program (HPCMP), the Extreme Science and Engineering Discovery
nucleation evaluated for several water models. J. Chem. Phys. 141, Environment (XSEDE), which is supported by the National Science
18C529 (2014). Foundation through grant no. 1548562, the Advanced Cyberinfras-
123. Kringle, L., Thornley, W. A., Kay, B. D. & Kimmel, G. A. Reversible tructure Coordination Ecosystem: Services & Support (ACCESS) pro-
structural transformations in supercooled liquid water from 135 to gram, which is supported by National Science Foundation grants nos.
245 K. Science 369, 1490–1492 (2020). 2138259, 2138286, 2138307, 2137603, and 2138296, the Triton Shared
124. Keutsch, F. N. & Saykally, R. J. Water clusters: untangling the Computing Cluster (TSCC) at the San Diego Supercomputer Center
mysteries of the liquid, one molecule at a time. Proc. Natl Acad. (SDSC), and the Scientific Computing Core at the Flatiron Institute, a
Sci. USA 98, 10533–10540 (2001). division of the Simons Foundation.
125. Gallo, P. et al. Water: a tale of two liquids. Chem. Rev. 116,
7463–7500 (2016).
126. Riera, M., Yeh, E. P. & Paesani, F. Data-driven many-body models Author contributions
for molecular fluids: CO2/H2O mixtures as a case study. J. Chem. F.P. conceived the research. S.L.B. and F.P. designed the research. S.L.B.
Theory Comput. 16, 2246–2257 (2020). performed the simulations. S.L.B and F.P. analyzed and discussed the
127. Riera, M., Hirales, A., Ghosh, R. & Paesani, F. Data-driven many- results, and wrote the paper. F.P. acquired funding and administered the
body models with chemical accuracy for CH4/H2O mixtures. J. project.
Phys. Chem. A 124, 11207–11221 (2020).
128. Lambros, E. et al. General many-body framework for data-driven
potentials with arbitrary quantum mechanical accuracy: water as a Competing interests
case study. J. Chem. Theory Comput. 17, 5635–5650 (2021). The authors declare no competing interests.

Nature Communications | (2023)14:3349 9

Article https://doi.org/10.1038/s41467-023-38855-1

Additional information Open Access This article is licensed under a Creative Commons
Supplementary information The online version contains Attribution 4.0 International License, which permits use, sharing,
supplementary material available at adaptation, distribution and reproduction in any medium or format, as
https://doi.org/10.1038/s41467-023-38855-1. long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license, and indicate if
Correspondence and requests for materials should be addressed to changes were made. The images or other third party material in this
Francesco Paesani. article are included in the article’s Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not
Peer review information Nature Communications thanks Ji Chen and the included in the article’s Creative Commons license and your intended
anonymous, reviewer(s) for their contribution to the peer review of this use is not permitted by statutory regulation or exceeds the permitted
work. A peer review file is available. use, you will need to obtain permission directly from the copyright
holder. To view a copy of this license, visit http://creativecommons.org/
Reprints and permissions information is available at licenses/by/4.0/.
http://www.nature.com/reprints
© The Author(s) 2023
Publisher’s note Springer Nature remains neutral with regard to jur-
isdictional claims in published maps and institutional affiliations.

Nature Communications | (2023)14:3349 10