Predicting material properties at finite temperatures is possible in theory via principles of quantum and statistical mechanics. However, such as ‘first principles’ is computationally prohibitive for general systems. In this talk, I will discuss recent advances in physics-based and machine learning (ML) techniques that enable a rigorous description of quantum statistical mechanics of generic material at finite thermodynamic conditions. These include pretrained ML interatomic potentials that achieve qualitative or semi-quantitative first-principles accuracy across diverse systems [1], fine-tuning protocols that enable quantitative first-principles accuracy for specific materials [2], ML-based predictions of electronic properties such as polarization and polarizability tensors, and ML techniques that incorporate quantum nuclear motion beyond the harmonic approximation at a classical computational cost [3]. Finally, I will demonstrate the capabilities of these approaches in delivering highly accurate and computationally efficient descriptions of molecular materials with broad implications in catalysis and pharma domains.

[1] Batatia, I. et al. arXiv:2401.00096 (2024).

[2] Kaur, H. et al. Faraday Discussions, 256, 120–138 (2025).

[3] Kapil, V. et al. Faraday Discussions, 249, 50–68 (2024).