Phase diagram of CO2-I/III from molecular dynamics simulation using a PBE0-accuracy machine learning potential

We employ the PBC-GEBF-AL workflow, integrating active learning (AL), the linear-scaling generalized energy-based fragmentation approach under periodic boundary conditions (PBC-GEBF), and multiTPU-OPES enhanced sampling, to create a PBE0-D3(BJ)/aug-cc-pVDZ-accuracy machine learning potential for studying the carbon dioxide (CO2) molecular crystal phase diagram. Based on up to 170 ns MLP-based multiTPU-OPES simulations, we obtain the phase diagram across a wide temperature range (250-700 K) and pressure range (10.5-14.0 GPa). It suggests a CO2-I/III coexistence line peaks at ∼12.3 GPa, 525 K, with a negative slope at higher temperatures, closely matching experimental I/III and I/VII transition pressures. Trajectory analysis revealed a concerted CO2-I/III transition mechanism driven by molecular rotation, lattice deformation, and non-monotonic volume changes (expansion followed by contraction). We propose a new Cmca space group structure for the true CO2-III phase, characterized by molecules tilted relative to the ac plane and lattice parameters a > b. It differs from the reported CO2-III structure determined by powder x-ray diffraction but closely resembles the CO2-VII crystal, where CO2 molecules are aligned parallel to the ac plane. This proposed tilted Cmca structure coexists with CO2-VII at high temperatures. We conject that CO2-III and CO2-VII belong to the same phase, with the discrepancies in their experimental Raman spectra primarily caused by slight structural changes due to the thermal effect. This work provides deeper insights into CO2 phase transitions and establishes a generalizable strategy for high-precision MLPs in complex rare event systems.