A quick point of view from an experimentalist. I think it is important to keep in mind that melting is a first order phase transition where extreme superheating and supercooling is common. Your simulations have most likely no defect where condensation or melting can start. That means you stay for quite a while in a relative free energy minimum without transitioning into the global free energy minimum which determines the equilibrium. In reality there are always condensation cores. If you have an ideal situation without defects in experiment you would be surprised how far you can, e.g., heat up water until it actually transitions into vapor. This is by the way why the chemists always use little stones when boiling water. They are well aware of the potentially dangerous boiling retardation.

Thank you for your answer, Christian. I am aware of superheating. Actually, I am using Z method, a method which uses the fact of overheating to get the melting curve (For a given volume (NVE ensemble), you increase the energy until it melts. Facts show that the system drops from maximum overheating temperature to the melting temperature). In these simulations, I getting melting temperatures are also higher than in the reported melting curve.

Actually, what I want is to compare the structure factor and radial distribution function for liquid structure at certain pressure and temperature. But the pressure and temperature at which they report g(r) and structure factor, produces a solid structure in my simulations, so g(r) is obviously different. So I don't know what to do in order to compare these properties if I don't do it at the same thermodynamic conditions. So what to do?

Dear Felipe:

I am also doing some investigations of high pressure melting. I think you should test the pseudopotential and exchange-correlation functional with experimental EOS. I also don't know whether the experimental data used in your research is from DAC or shock wave loading.

Furthermore, the number of atoms used in your simulations has an important effect on the results. You can try to compare the results of 16, 32, 64, 128 atoms.

Dear Kai:

My simulations use 96 atoms. I re-built the 0 K isotherm using the pseudopotential and there is good agreement. The problem is that there is no agreement in melting temperature. The experimental g(r) is obtained doing x-ray diffraction in a DAC. The sample can melt, as Christian says, due to the presence of defects in the experiment, so how can I compare with the experiment if the melting temperatures differ?

Are you making your sample experimentally or by computational chemistry (Gaussian program)?

If experimentally, then you can try TGA to check the degradation point.

No, Abdullah, as the title says, I am doing ab-initio simulations, with the VASP program. I am trying to compare with experiments. But thanks for the interest.

You can try Semi Imperial method to compare your result.

Like what?

Like LAMMPS. You can try solid-liquid coexistance method to simulate melting with EAM potentials. Indeed, the Z method can only give the up boundary of melting curve.

The problem for LAMMPS is that I would need a classical potential which I don't have. Coexistance method would be to determine the melting curve, which I have already done with Z method. But how can I compare those g(r) and structure factor with experiments if the g(r) of my simulation shows that for the same PT conditions, my structure is solid, while in the experiment, the g(r) correspond to a liquid?

Felipe, that simply means that the experimental Tm and the one you computed are very different. It is not obviuous which one is wrong - we know, for example that the

melting curve for iron has recently moved up by quite some amount (see Anzellini, Science, 2013). The experimental iron melting curve is now corrected upwards by

quite some T shift - it is now consistent with the one obtained from shock-wave experiments and theory (e.g. Belonoshko and Ahuja, PEPI, 1997; Belonoshko et al., PRL 2000; Nature 2003). So, it could be that the experimental curve is just too low. On the other hand, you sample is rather small - if you compute melting for a Close packed structure stable at T=0 - it is fine, but if you ar computing melting T for the structure unstable at T=0 - you might get results that are size dependent up to very large sizes. I would take an empirical potential and check the size dependence of your results for the structure you use. The empirical potential does not have to be of the same queality as DFT but it should correctly reproduce essential behaviour of the DFT system.

I also wonder, if you take a liquid sample and , strating from it, perform an MD run at the experimental melting conditions, shall it become solid?

According to the recent results of my LAMMPS simulations, I think the liquid may transform into a solid phase different from the DFT calculations or experiments. In another words, if the metal is in bcc structure at some given P-T confirmd by experiments, it may transform into a complex structure including several solid phases by using coexistance method. So it may be not appropriate to investigate the solid-solid transition with coexistance method.

However, as Belonoshko said, the emperical potential can reproduce the melting curve obtained from AIMD calculations and shock wave experiments.

Kai, you are quite right. Empirical and DFT models might be very different form each other and produce different results.

Dear Anatoly:

Thank you for your advices and references. That is very helpful. I will compute the melting curve with your Z method, and I will compare with experimental curve before I come back with this problem.

Thanks again.