Numerical modelling of materials

Machine Learning

Post-Image

Figure 1: Active learning procedure for structural search. From the PhD thesis of J.-B. Charraud, Studies of superhydrides by ab-initio and machine learning accelerated random structure searching (2021)

Machine Learning

Our research focuses on the development of predictive atomistic and molecular simulation methodologies to describe materials with the help of machine learning tools. In order to improve the accuracy and transferability of interatomic potentials we develop Machine Learning Potentials (SNAP, GAP, NNP) for the description of different phenomenon and materials, such as for studying the structure of superhydrides [1] or to describe plasticity and phase transition [2] as well as magnetic contributions [3] in iron. Machine learning tools are also employed to characterize defects in crystalline solids [4] and classify structures and dislocations on the fly. On a more fundamental perspective, we applied Invertible Neural Networks to solve inverse problems and determine the Minimum Energy Path associated to a phase transition [5]. The link between atomistic simulations and mesoscopic modeling for chemically reactive systems is achieved with the help of non supervised ML techniques. Finally, in the aim of accelerating ab initio molecular dynamics (AIMD) simulations, which are particularly costly in terms of computation time, we develop the Machine Learning Assisted Canonical Sampling (MLACS) [6,7] approach. This method is based on on-the-fly fitting of machine learning potentials and is able to sample the configuration space for a fraction of the cost of AIMD simulations while keeping the same accuracy.

Publications

  1. J.-B. Charraud, G. Geneste, M. Torrrent, J.-B. Maillet. “Machine learning accelerated random structure searching: Application to yttrium superhydrides” J. Chem. Phys., 156, 204102 (2022) DOI
  2. J.-B Maillet, C. Denoual, G. Csanyi. “Machine-learning based potential for iron: Plasticity and phase transition” AIP Conference Proceedings, 1979, 050011 (2018) DOI
  3. S. Nikolov, M. A. Wood, A. Cangi, J.-B. Maillet, M.-C. Marinica, A.P. Thompson, M.P. Desjarlais, J. Tranchida. “Data-driven magneto-elastic predictions with scalable classical spin-lattice dynamics” npj Computational Materials, 7, 153, (2021) DOI
  4. A.M. Goryaeva, C. Lapointe, A. Cangi, C. Dai, J. Dérès, J.-B. Maillet, and M.-C. Marinica “Reinforcing materials modelling by encoding the structures of defects in crystalline solids into distortion scores” Nat. Comm., 11, 4691, (2020) DOI
  5. M. Ramil, C. Boudier, A.M. Goryaeva, M.-C. Marinica, and J.-B. Maillet. “On Sampling Minimum Energy Path” J. Chem. Theory Comput., 18, 5864 (2022) DOI
  6. A. Castellano, F. Bottin, J. Bouchet, A. Levitt, G. Stoltz, “Ab initio canonical sampling based on variational inference”, Phys. Rev. B, 106, L161110 (2022) DOI
  7. A. Castellano, R. Béjaud, P. Richard, O. Nadeau, C. Duval, G. Geneste, G. Antonius, J. Bouchet, A. Levitt, G. Stoltz, F. Bottin “Machine Learning Assisted Canonical Sampling (MLACS)”, arXiv:2412.15370 DOI (2024) (2024)

Researchers involved

R. Béjaud, F. Bottin, F. Brieuc, G. Geneste, G. Kluth, P. Lafourcade, J.-B. Maillet, M. Torrent