The origin of complex behavior of liquid carbon: an insight from computer simulation Yu. D. Fomin Abstract In the present paper we perfomrm molecular dynamics simulation of liquid carbon with a machine-learning potential GAP-20. We show that within the framework of this model carbon demonstrates a relatively low critical temperature, which can affect the results of experimental measurements of melting point of graphite. Keywords :speed of sound, binary mixture, phase diagram, Frenkel line pacs 61.20.Gy, 61.20.Ne, 64.60.Kw I Introduction Melting of graphite is one of a long-standing problems of physics. Although the first experiments were undertaken in the beginning of XX century 1911 (see, for instance, Chapter 1 of the book savvatimsky for a historical review), the exact location of the melting point is still controversial. In spite of numerous works in this field different experiments give the estimation of melting point of graphite from T_{m}=4000 K savvatimsky up to T_{m}=6700 K rakhel . Apparently, such dispersion of the results cannot be recognized as satisfactory. Therefore, novel ideas are required in the field to find the reasons for such spread of the results. In the present paper we try to find out possible reasons for the large difference between the results of different experiment. In order to do it we analyze the thermodynamic properties of graphite and liquid carbon obtained in different experiments and theoretical models. In order to find some reference points on the phase diagram we start from the discussion of graphite-liquid-vapor triple and critical points of carbon. The estimations of the triple point suffer on great uncertainty savvatimsky . At the same time there are just a few evaluations of the critical point of liquid-gas transition in carbon. The most well recognized paper is the one by Leider, Krikorian and Young leider . In order to evaluate the critical temperature of carbon, the authors of this paper had to make several assumptions. According to the available for them experimental results the specific volume of liquid carbon had to be substantially larger than the one of graphite, but they did not have exact values of the density of molten carbon. For this reason they arbitrary assumed that the volume jump upon melting of carbon is 20\% . This assumption gave them the density of liquid carbon at the triple point \rho_{liq-tr}=1.611 g/ml . The second assumption of Ref. leider is that thermal expansion coefficient of liquid carbon is the same that the one of graphite. Using the law of rectilinear diameters the authors estimate the critical point parameters as T_{c}=6810 K, \rho_{c}=0.64 g/ml and P_{c}=2200 atm. It is seen from the description above that two suspicious assumptions are introduced in Ref. leider . For instance, the later experimental result state that the volume jump of carbon upon graphite melting is much larger: V_{l}/V_{s}=1.7 (see paragraph 5.2 of Ref. savvatimsky ). In Fig. 1 we show the boiling line of carbon obtained by rescaling the boiling line of Lennard-Jones system according to the law of corresponding states. The critical parameters of carbon were taken from Ref. leider . In the P-T plane we show also the melting curve of graphite from a seminal work by Bundy bundy and the melting curve obtained by Korobenko, Savvatimskiy and Cheret ksch . By extrapolating the curve to the zero pressure limit we obtain that the boiling curve does not cross the melting line from the Bundy work, i.e. no triple point is identified. If we take the melting curve from Ref. ksch we obtain that the triple point takes place at P=180 bar and T=4800 K. The triple point obtained in Ref. leider is in the range 4600-4800 K and P=103 bar. A careful analysis of the most of the available data given by Savvatimsky in Ref. savvatimsky also gives the location of the triple point at P\approx 120 bar and T\approx 4800-4900 K. Figure 1: Boiling curve of carbon obtained by the corresponding states law with the critical parameters from Ref. leider in (a) \rho-T and (b) P-T planes. The red star in the panel (b) shows an extrapolation of the boiling curve to the limit of zero pressure. The curves ’Bundy’ and KSCh in the panel (b) show the melting curves of graphite from Refs. bundy and ksch correspondingly. The analysis of experimental data also requires some parameters which are not known with sufficient accuracy. For instance, it is supposed in Ref. ksch that the heat capacity of graphite prior to the melting point is c_{P}=3.2 \frac{J}{g\cdot K} for the temperature range from T=4200 K to 5200 K. At the same time the authors compare their results for the heat capacity with the literature data: c_{P}=3.27 \frac{J}{g\cdot K} at 4400 K and increases to 4.11 \frac{J}{g\cdot K} in Ref. cp-1 and c_{P}=3.77 \frac{J}{g\cdot K} at T=4500 K in Ref. cp-2 . Although the difference between different data is not so crucial, the amount of absorbed heat is determined by an integral of the heat capacity, therefore, this disagreement can result in completely different estimation of the melting point. Another kind of inconsistencies take place in a number of other works. For instance, in Ref. rakhel experimental measurements of ’liquid carbon’ density at T=7000 K are reported. However, this temperature exceeds the T_{c} from Ref. leider , therefore no ’liquid’, but supercritical fluid takes place in this region. At the same time, measurements of Ref. rakhel show that the isochoric heat capacity of ’liquid’ carbon is c_{V}/R=3.4\pm 0.4 . This value looks to be very small for a region of phase diagram close to the critical point, where the heat capacity demonstrate strong grows. We remind that according to Ref. leider the critical temperature T_{c}=6810 K is the estimation of the upper limit of critical temperature of carbon. It means that the real critical temperature can be much lower. However, strong maximum of isochoric heat capacity preserves up to the temperatures of T\approx 3T_{c} widom-lj , therefore the heat one should still expect large c_{V} . The examples above demonstrate an important lack in the analysis of the experimental data: although experiments on graphite melting are performed at extremely high temperatures, in most cases influence of the gas phase is not taken into account. Interestingly, the authors of experimental papers often mention that the presence of sublimated carbon influence the results of measurements, for instance, worsening the data of pyrometry, but the presence of gas phase is not taken into account in the thermodynamic calculations to develop the experimental data. At the same time it looks clear that sublimation of carbon should influence the heat capacity of the sample. In the present paper we try to find out the location of the critical point of carbon by molecular dynamics method. Having determined the location of the critical point we calculate the points of maxima of thermodynamic response functions (Widom lines) and discuss their influence on the experimental measurements of carbon melting. II System and Methods In the present work we simulate a system of 4096 carbon atoms in a cubic box with periodic boundary conditions. GAP-20 potential gap20 is used to model the interactions between the particles. The same model was used in Ref. or-gap for simulation of fluid state of carbon at high temperature. The system was modeled in canonical ensemble (constant number of particles N, volume V and temperature T). The timestep was set to dt=2.0 fs. The equilibration period was 10 ps. The production period was set to 20 ps. Since we are interested in liquid-gas critical point we had to simulate a system in a large number of state points. The density varied from \rho_{min}=0.05 g/ml to \rho_{max}=2.0 g/ml with step \Delta\rho=0.05 g/ml . The temperatures varied from T=4000 K to 5000 K with step \Delta T=200 K. Additionally we calculated the T=4100 , 6000 and 7000 K isotherms. During the course of simulation we sampled the thermodynamic quantities such as pressure and internal energy. We constructed the isotherms of the system and identified the critical point as a point of inflection on the isotherm. We approximate the thermodynamic data from simulation by an equation X=\sum_{i,j=0,4}a_{ij}T^{i}\rho^{j} , where X stands for internal energy or pressure and a_{i,j} are fitting coefficients. In this approximation we used only the data for \rho\leq 1.0 g/ml . This density interval was sufficient for investigation of near critical anomalies. The fitting coefficients are given in Supplementary materials. Using these fits we calculated the thermodynamic response functions such as isobaric heat capacity c_{P}=\left(\frac{\partial H}{\partial T}\right)_{P} , where H=U+PV is the enthalpy of the system, thermal expansion coefficient \alpha_{P}=\frac{1}{\rho}\left(\frac{\partial\rho}{\partial T}\right)_{P} and isothermal compressibility \beta_{T}=\frac{1}{\rho}\left(\frac{\partial\rho}{\partial P}\right)_{T} . We calculated the location of the maxima of these quantities. All simulations were performed using the LAMMPS simulation package lammps . III Results and Discussion First of all we calculate the equation of state (EoS) of carbon and fit it to the polynomial function described above. The examples of EoS along isotherms and fit of the data are shown in Fig. 2 (a) - (c). The MD data are rather noisy. The GAP-20 potential is rather computationally heavy. For this reason, the MD simulation was relatively short which did not allow to collect very good statistics. The fitting curves are in good agreement with the results of simulation. Therefore, we assume that one can use the fitting polynomial to determine the thermodynamic properties of carbon. Figure 2: Equation of state and fitting curve of GAP-20 model of carbon at (a) T=4100 K, (b) T=4400 K and (c) T=5000 K. Figure 3 shows several isotherms in the vicinity of the critical temperature. From these isotherms we estimate the critical point as T_{c}=4235 K, \rho_{c}=0.1887 g/ml and P_{c}=106.61 bar. It is much lower than the one from Ref. leider . The estimation of graphite-liquid-gas triple point in the same paper is T_{t}=4765 K, which is above our estimation of T_{c} . At the same time, experimental evaluation of triple point of carbon suggests the values from T_{t}=3670 K to 4650 K (see table 3.1 in Ref. savvatimsky ), therefore T_{c} of our simulations can be consistent with some experimental works. Any way, we would like to stress that evaluation of the critical point is a difficult task which requires special care. Even tiny mistakes in interaction potential can lead to strong deviation of calculated critical point from the experimental one. The GAP-20 model has been fitted to simulate condensed phases of carbon and it can be not good enough for the critical point evaluation. However, this result is valuable, since it demonstrates that accounting for the critical phenomena is required in analysis of experimental data on graphite melting. At the same time at T=4100 K and \rho=1.55 g/ml we observe that system demonstrates some traces of crystallinity. It means that the critical point is located rather close to the melting line in the density-temperature plane. It is well known that near the critical point some quantities (heat capacity, thermal expansion coefficient, etc.) demonstrate maxima (Widom lines). This case when the melting process is complete the system can approach the lines of maxima of these quantities, which should strongly influence the thermodynamic response of the system, for instance, the heat absorption of molten carbon. We remind that in many publications the heat capacity of the liquid carbon is assumed to be constant in a wide range of temperatures. This assumption is violated in the vicinity of the critical point. Moreover, since the thermodynamic function demonstrate strong dependence on pressure and temperature, even small uncertainty in these quantities can lead to a large discrepancy of the results, which can be responsible for a large body of conflicting measurements of the carbon melting point. Figure 3: Several isotherms next to the critical temperature of carbon obtained from the polynomial fit. Figure 4: A snaphot of a sample at T=4100 K and \rho=1.55 g/ml . Traces of crystallization are visible in the bottom part of the Figure. Based on the discussion above it is impotant to calculate the near-critical maxima of different thermodynamic functions of GAP-20 carbon. Examples of isobaric heat capacity, thermal expansion coefficient and isothermal compressibility of GAP-20 carbon are given in Fig. 5 . As it is expected, all these quantities demonstrate maxima close to the critical point, which rapidly disappear as we move away from it. We recall that in Ref. rakhel large heat capacity of liquid carbon at T=7000 K was reported: c_{V}/R=3.8\pm 0.4 . Later on in Ref. or-gap an attempt to explain the high value of c_{V} was undertaken. The authors assumed that the reason for large c_{V} can be due to transition from high density sp2 hybridized liquid to an sp hybridized low density one. They also observe voids in the ’liquid’ and negative slope of the density dependence of isothermal expansion coefficient \beta_{T} . All these results are compatible with the assumption that ’liquid’ carbon is indeed in supercritical state (below we use the abbreviation SCF for ’supercritical fluid’). While gaseous carbon is mostly sp hybridized, the liquid one has sp2 hybridization. A smooth crossover from gas-like to liquid-like state takes place at supercritical temperatures. SCF demonstrates clusters at low densities and voids at higher one. Finally, the region of densities studied in Ref. or-gap belongs to decreasing branch of isothermal compressibility, whose maximum takes place at the densities about \rho=0.2 g/ml (see Fig. 5 (c)). Figure 5: (a) Isobaric heat capacity, (b) thermal expansion coefficient and (c) isothermal compressibility of GAP-20 model of carbon next the critical point. The locations of maxima of thermodynamic response functions of GAP-20 model of carbon are shown in Fig. 6 . This figure looks qualitatively similar to the Widom lines of other substances: van der Waals gas vdw , Lennard-Jones system lj-wid , square well sw-wid , carbon dioxide co2-wid , etc. The maxima of \beta_{T} take place only at the temperatures slightly exceeding the critical one and rapidly disappear. At the same time the maxima of c_{P} and \alpha_{P} preserve even at rather large temperatures. We recall that in the case of simple liquid these maxima take place up to the temperatures about 3T_{c} . We have not done the simulations at such high temperatures. Figure 6: Location of supercritical maxima of GAP-20 model of carbon. Calculation of location of triple point of carbon requires estimation of both melting and boiling lines, which goes beyond the scope of the present paper. In our simulation we observe spontaneous crystallization of liquid carbon into graphite phase at T=4100 K, \rho_{gr}=1.55 g/ml , P_{gr}=7225 bar and T=4200 K, \rho_{g}r=1.8 g/ml , P_{gr}=20262 bar. These values are close to the critical point in temperature, but exceed it in pressure. However, the density rapidly decreases upon isobaric heating which can result in approaching of the Widom line in the experimental studies of liquid carbon. It is of special interest to have a look on the maxima of isobaric heat capacity and thermal expansion coefficient (Fig. 5 (a) and (b)). It is seen, that in the close vicinity of the critical point the isobaric heat capacity reaches as much as 25k_{B} per particle. It rapidly drops on moving away from the critical point. Any way, it is clear that the assumptions of constant heat capacity completely fail. The same speculation takes place for the thermal expansion coefficient, which is supposed to be close to the one of graphite: in the vicinity of the critical point \alpha_{P} is much larger than it would be expected from the one of graphite. We would like to stress once again: the results of this work are obtained with the help of GAP-20 model, which was developed to describe condensed phases of carbon. For this reason, these results cannot be directly projected on the experimental data and can be used for qualitative speculations only. However, such speculations allow us to propose a possible explanation of large spread of experimental data on melting point of graphite. The melting temperature of graphite appears to be very close to the critical one. As a result when the crystal structure collapses, the system appears to be in supercritical state with strong fluctuations of thermodynamic quantities (see Fig. 6 ). These fluctuations play two negative roles in the analysis of the experimental data. First of all, to the best of our knowledge, they are not accounting for when the raw data are analyzed. The second point is that even small experimental uncertainties lead to great errors. For instance, when the system is close to the maximum of isothermal compressibility even small error in pressure leads to strong error in density. IV Problems of the critical point calculations In the present paper we discuss calculation of critical point of carbon. The results obtained in our calculations look very odd, which is, however, consistent with general understanding of the critical point calculations. The GAP-20 potential employed in the present study is a machine-learning model fitted to reproduce the DFT calculations. Therefore, it cannot be more accurate than the DFT calculations themselves. However, calculation of critical point of gas-liquid transition is a challenging problem for DFT and indeed the DFT results are not trustable themselves. Several problems take place to calculate the location critical point in DFT. First of all, the critical point of carbon is located at very high temperature, which looks unreasonably high for DFT. Also van der Waals (vdW) forces are of great importance in the vicinity of the critical point. The optB88-vdW functional used for fitting the GAP-20 model employees the van der Waals corrections, which are important for the reliable graphite calculations. This functional shows good results for solids dft-1 ; dft-2 and surfaces of crystals dft-2 . However, it was not tested for diluted phases such as gaseous carbon. Therefore, large errors can be expected for gases and supercritical fluids. More problems take place in the vicinity of the critical point, such critical fluctuations and critical slowing down which i) require to simulate rather large system and ii) require to perform very long simulations to get a lot of data for proper averaging. Although GAP-20 potential works much faster than ab-initio calculations, we are still unable to overcome these problems (the system size is 4096 atoms and the production run is 20 ps). Moreover, since large fluctuations take place next to the critical point, even tiny error in some parameters strongly affects the others. For instance, the well recognized model of water TIP4P/2005 gives the critical temperature T_{c}=640 K (the experimental value is T_{c}^{exp}=647 K), critical density \rho_{c}=0.31 g/ml (the experimental one \rho_{c}^{exp}=0.322 g/ml ), while the critical pressure is P_{c}=146 bar ( P_{c}^{exp}=220.64 bar) tip4p , i.e., a relatively small error in the critical density leads to a great effect on critical pressure: P_{c}^{exp}/P_{c}=1.5 . This example shows that even in the case of well developed systems with numerous experimental data available the evaluation of the critical point parameters is still challenging. In the case of substances with low critical temperature the location of the critical point is used as one of the fitting parameters. For instance, Lennard-Jones potential parameters of noble gases can be selected in such a way that the critical point of the model corresponds to the experimental critical point of the substance of interest. This is, however, not applicable to the case of carbon. Based on this discussion, we are sure that the location of the critical point reported in the present study does not correspond to the critical point of real carbon. At the same time we believe that the calculation of the present paper make sense. These calculations let us to propose that one of two possibilities can take place in the case of experimental investigations of melting of graphite. The first option is the one obtained in our simulations: after the heating of graphite the system transforms into SCF. This case it experiences strong fluctuations, including maxima of heat capacity (which influences the estimation of the temperature) and thermal expansion coefficient (which leads to wrong evaluation of the density). The second option which should take place if the critical point is at the temperatures above the ones reported in the melting experiments is that after the melting of graphite the system can be into the liquid-gas two phase region. However, the system in two phase region is also very sensitive to tiny perturbations, since they change the amount of particles in the phases. In both cases one should be very careful to the thermodynamic properties of the system and the usual assumptions like constant heat capacity or constant thermal expansion coefficient should give erroneous results. Novel measurements of the graphite melting appeared recently rakhel-new . According to the results of this paper, the melting temperature of graphite in the pressure range of 0.2-2.5 GPa is about 6000 K and demonstrate very weak dependence on pressure. The estimation of the volume of the solid and liquid carbon at the melting point is V_{s}/V_{0}\approx 1.3 and V_{l}/V_{0}\approx 1.8 , where V_{0}=0.4425 ml/g is the volume of HOPG at ambient conditions. It gives V_{l}/V_{s}\approx 1.38 , which is different from the values used in leider to evaluate the critical point. The authors estimate the isochoric heat capacity of liquid carbon next to the melting line to be c_{V}=2.9 J/(g\cdot K) , which corresponds to c_{V}/k_{B}\approx 4.19 . This value is much larger than the one corresponding to Dulong-Petit law, which can happen either in the near-critical fluctuations region, or if some structural crossovers in the liquid phase take place highcv-1 ; highcv-2 . At the same time the authors write that the thermal expansion coefficient is nearly constant in the temperature range of 4000-6000 K, which is not consistent with the assumptions above. Once again we see that even in the case of careful measurements no self-consistent result is achieved. The authors of Ref. rakhel-new also compare their results with the ones of the quantum molecular dynamics simulations. The simulations give the melting temperatures of graphite of the order of 3800 K. The GAP potential employed in the present study should give the melting point of the same order of magnitude. Comparing different experimental and computational works one can find that simulations give too low melting points comparing to the experiments. One can assume that some drawbacks of quantum simulations lead to underestimation of the melting points. Further development of the simulation methods is definitely required. V Conclusions In the present paper we perform molecular dynamics simulation of carbon with GAP-20 machine learning potential. We show that this model demonstrates a very low critical temperature of gas-liquid transition: T_{c}=4235 K, which is very close to the results for the melting temperature of graphite. For the first time we calculate the Widom lines of carbon in the framework of GAP-20 model. Based on these results we propose a possible mechanism of inconsistencies between different experiments on graphite melting: they might be related to the strong fluctuations of thermodynamic quantities in the vicinity of the critical point. We note that the results are obtained in the framework of GAP-20 model, which was constructed for condensed phases of carbon. For this reason the results are not applicable to direct comparison with experiment. At the same time comparing different computational and experimental works, one can assume that the simulations give qualitatively correct results, but shifted with respect to the experimental ones to lower temperatures. This case if can happen that the idea of the present paper is correct, but the critical point is at much higher temperatures. This assumption is consistent with the critical temperature of carbon from Ref. leider ( T_{c}=6810 K) and the results for the melting temperature of graphite obtained in the recent work rakhel-new T_{m}\approx 6100 K. We believe that construction of novel models of carbon, fitted to the properties of liquid, gas and graphite phases at the same time is required for more elaborate simulation of the system. VI