Representations in atomistic ML

I have spent some time wondering about the fundamental questions about crafting inputs to atomistic ML models, which are also closely tied to their architectures. More specifically, in terms of descriptors, what are the ingredients central to their mathematical representations? How can we incorporate equivariance with the physical symmetries underlying structures in 3D Euclidean space? How to ensure that two structures unrelated by symmetries are mapped to different descriptors? How can we capture long-range (Coulomb) interactions into machine learning models while keeping a local description of atomic environments? Another line of research has been on identifying the similarities and differences between models that rely on these descriptors as inputs and models that work directly with input structures (such as graph neural networks).

Related publications

2024

Expanding density-correlation machine learning representations for anisotropic coarse-grained particles

Arthur Lin, Kevin K Huguenin-Dumittan, Yong-Cheol Cho, Jigyasa Nigam, and Rose K Cersonsky

The Journal of Chemical Physics, 2024

Bib

@article{lin2024expanding,
  title = {Expanding density-correlation machine learning representations for anisotropic coarse-grained particles},
  author = {Lin, Arthur and Huguenin-Dumittan, Kevin K and Cho, Yong-Cheol and Nigam, Jigyasa and Cersonsky, Rose K},
  journal = {The Journal of Chemical Physics},
  volume = {161},
  number = {7},
  year = {2024},
  publisher = {AIP Publishing},
  reps = {true}
}

Completeness of atomic structure representations

Jigyasa Nigam, Sergey N Pozdnyakov, Kevin K Huguenin-Dumittan, and Michele Ceriotti

APL Machine Learning, 2024

Bib

@article{nigam2024completeness,
  title = {Completeness of atomic structure representations},
  author = {Nigam, Jigyasa and Pozdnyakov, Sergey N and Huguenin-Dumittan, Kevin K and Ceriotti, Michele},
  journal = {APL Machine Learning},
  volume = {2},
  number = {1},
  year = {2024},
  publisher = {AIP Publishing},
  reps = {true}
}

2022

Unified theory of atom-centered representations and message-passing machine-learning schemes

Jigyasa Nigam, Sergey Pozdnyakov, Guillaume Fraux, and Michele Ceriotti

The Journal of Chemical Physics, 2022

Bib

@article{nigam2022unified,
  title = {Unified theory of atom-centered representations and message-passing machine-learning schemes},
  author = {Nigam, Jigyasa and Pozdnyakov, Sergey and Fraux, Guillaume and Ceriotti, Michele},
  journal = {The Journal of Chemical Physics},
  volume = {156},
  number = {20},
  pages = {204115},
  year = {2022},
  publisher = {AIP Publishing LLC},
  reps = {true}
}

2021

Multi-scale approach for the prediction of atomic scale properties

Andrea Grisafi, Jigyasa Nigam, and Michele Ceriotti

Chemical Science, 2021

Bib

@article{grisafi2021multi,
  title = {Multi-scale approach for the prediction of atomic scale properties},
  author = {Grisafi, Andrea and Nigam, Jigyasa and Ceriotti, Michele},
  journal = {Chemical Science},
  volume = {12},
  number = {6},
  pages = {2078--2090},
  year = {2021},
  publisher = {Royal Society of Chemistry},
  reps = {true},
}

2020

Recursive evaluation and iterative contraction of N-body equivariant features

Jigyasa Nigam, Sergey Pozdnyakov, and Michele Ceriotti

The Journal of Chemical Physics, 2020

Bib

@article{nigam2020recursive,
  title = {Recursive evaluation and iterative contraction of N-body equivariant features},
  author = {Nigam, Jigyasa and Pozdnyakov, Sergey and Ceriotti, Michele},
  journal = {The Journal of Chemical Physics},
  volume = {153},
  number = {12},
  pages = {121101},
  year = {2020},
  publisher = {AIP Publishing LLC},
  reps = {true},
}