Researchers Develop New Method for Understanding Material Behavior Under Different Pressure-Temperature-Volume Conditions

Strategic Insights -

Researchers have proposed a new data-driven approach to modeling material behavior under various pressure-temperature-volume conditions, which is crucial for understanding complex materials and their applications. The traditional models rely on thermodynamic assumptions and are limited by scalability and extrapolability, whereas the new method, called EOSNN, uses a neural network-based physics-informed deep learning approach that jointly learns multiple equation of state (EOS) surfaces from diverse data sources. This method is shown to outperform traditional and Gaussian process methods in terms of accuracy, flexibility, and extensibility.

Key Takeaways:

  • The equation of state (EOS) is essential for understanding material behavior under different pressure-temperature-volume conditions.
  • Traditional models, such as the Mie-Grüneisen-Debye equation, rely on thermodynamic assumptions and expert knowledge.
  • EOSNN, a neural network-based physics-informed deep learning method, jointly learns multiple EOS surfaces from diverse data sources.
  • The method is shown to outperform traditional and Gaussian process methods in several aspects, including accuracy, flexibility, and extensibility.
  • The method's prediction for the energy off-Hugoniot can reach a F1 score as high as 0.83 and RMSE as low as 0.52 eV/atom with proper selection of regularization.
  • The benefits of the method can be further enhanced with physics-informed regularizations on quantities such as heat capacity, Grüneisen parameter, or bulk modulus.

Statistics:

  • The method's F1 score for the energy off-Hugoniot is as high as 0.83.
  • The method's RMSE for the energy off-Hugoniot is as low as 0.52 eV/atom.
  • The traditional regression method based on the Mie-Grüneisen equation has a F1 score of 0.013 and an RMSE of 1.78 eV/atom when trained on the same data.

Sources:

  • Joint learning equation of state surfaces with uncertainty-aware physically regularized neural networks. Scientific Reports, 2025;15(1):27046.
  • Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, 519082, People's Republic of China.