To investigate phase transitions—such as those in barium titanate (BaTiO₃)—where the crystal structure changes with temperature using atomic simulations, it is necessary to perform long-duration molecular dynamics simulations on systems containing a large number of atoms while varying the temperature. While first-principles molecular dynamics using DFT allows for highly accurate evaluation of interatomic forces, simulating systems on the order of several thousand atoms over hundreds of thousands of steps is extremely computationally expensive. Consequently, machine learning interatomic potentials (MLIPs) are increasingly being adopted as a solution to this problem. MLIPs aim to achieve a balance between descriptive power comparable to first-principles calculations and computational speed similar to classical molecular dynamics by rapidly evaluating interatomic interactions—learned from first-principles calculations—using neural networks and other methods. With Quloud, both material-specific MLIPs and general-purpose MLIPs can be easily utilized within LAMMPS—a software package that can be considered the de facto standard for molecular dynamics (MD). This enables simulations with complex settings—such as pressure, temperature, and heating/cooling processes—to be performed using potentials with first-principles-level accuracy.

HybMD is a hybrid simulation method developed by Quemix that combines high-speed molecular dynamics calculations using machine learning potentials with electronic state evaluations based on first-principles calculations. In ion migration and interfacial reactions occurring under an external electric field, not only do atoms move, but their charge states also change depending on the local environment. HybMD addresses this by rapidly evaluating interatomic forces using a machine learning potential while updating the electron density and atomic charges via Orbital-Free Density Functional Theory (OFDFT), thereby enabling the simultaneous treatment of electric field-driven ion transport and the accompanying charge changes. While general machine learning potentials can rapidly calculate energy and interatomic forces, they do not directly output electron density or atomic charges that vary depending on the local environment. Consequently, there is a challenge in directly representing key phenomena—such as “charge changes accompanying atomic motion,” “redox reactions at interfaces,” and “electric field responses corresponding to local structural changes”—which are crucial when modeling ion transport and interfacial reactions under electric fields. On the other hand, while using Kohn–Sham DFT at each step of MD allows for detailed tracking of electronic states, the computational cost becomes extremely high for large-scale systems and long time scales. HybMD bridges this gap by adopting a division of labor: machine learning potentials are used for interatomic forces, while OFDFT is used for electronic density and charge.

In crystal growth processes, factors such as temperature, partial pressure of feed gases, carrier gases, and surface structure are intricately interrelated, making it difficult to determine which surface states will actually emerge based solely on experimental conditions. A research group from Kyushu University and Nagoya University has devised a method to predict stable surface reconfigurations based on growth conditions by combining the surface structure energy obtained from first-principles calculations with the chemical potential of molecules in the gas phase. They have achieved significant results by applying this method to materials such as gallium nitride. Quloud can also perform the same analysis by integrating multiple computational software packages. For example, first-principles calculations such as Quantum ESPRESSO are used to evaluate the energy of surface slabs and adsorption structures, while thermodynamic quantities derived from the vibrational, rotational, and translational degrees of freedom of gas-phase molecules are evaluated using the quantum chemical calculation software Psi4. This same approach can be applied to various crystal growth and surface reaction processes—such as those involving semiconductors, oxides, and catalyst surfaces—where the gas phase interacts with solid surfaces.