The Kitchin Research Group

In this work, we show that we can use large graph neural networks to predict transition metal complex energies. We developed an improved dataset at a higher level of theory, and tested models ranging from GemNet-T (best) to SchNet (worst). The model performance saturates with the size of neutral structures, and improves with increasing size of charged structures. Finally, we showed that a pre-trained model from OC20 was even better than training from scratch. This indicates a degree of transferability from heterogeneous catalyst models to homogeneous molecular catalysts.

@article{garrison-2023-apply-large,
  author =       {Garrison, Aaron G. and Heras-Domingo, Javier and Kitchin, John
                  R. and dos Passos Gomes, Gabriel and Ulissi, Zachary W. and
                  Blau, Samuel M.},
  title =        {Applying Large Graph Neural Networks To Predict Transition
                  Metal Complex Energies Using the tmQM\_wB97MV Data Set},
  journal =      {Journal of Chemical Information and Modeling},
  volume =       0,
  number =       0,
  pages =        {null},
  year =         2023,
  doi =          {10.1021/acs.jcim.3c01226},
  URL =          {https://doi.org/10.1021/acs.jcim.3c01226},
  eprint =       {https://doi.org/10.1021/acs.jcim.3c01226},
  note =         {PMID: 38049389},
}

Copyright (C) 2023 by John Kitchin. See the License for information about copying.

org-mode source

Org-mode version = 9.7-pre

The electron density is one of the most important quantities we use DFT to calculate. It is the foundation of how we compute energy, forces, and many other properties. DFT is expensive though, so in this work we show that we can build a graph neural network that can be used to predict electron densities directly from the atomic coordinates of a system. We show that the predicted densities can also be used to estimate dipole moments and Bader charges.

@article{sunshine-2023-chemic-proper,
  author =       {Sunshine, Ethan M. and Shuaibi, Muhammed and Ulissi, Zachary
                  W. and Kitchin, John R.},
  title =        {Chemical Properties From Graph Neural Network-Predicted
                  Electron Densities},
  journal =      {The Journal of Physical Chemistry C},
  volume =       0,
  number =       0,
  pages =        {null},
  year =         2023,
  doi =          {10.1021/acs.jpcc.3c06157},
  url =          {https://doi.org/10.1021/acs.jpcc.3c06157},
  eprint =       {https://doi.org/10.1021/acs.jpcc.3c06157},
}

Copyright (C) 2023 by John Kitchin. See the License for information about copying.

org-mode source

Org-mode version = 9.7-pre

In this work we show that prediction errors from graph neural networks for related atomistic systems tend to be correlated, and as a result the differences in energy are more accurate than the absolute energies. This is similar to what is observed in DFT calculations where systematic errors also cancel in differences. We show this quantitatively through differences of systems with systematically different levels of similarity. This article is part of the Special Collection: 2023 JCP Emerging Investigators Special Collection.

@article{ock-2023-beyon-indep,
  author =       {Janghoon Ock and Tian Tian and John Kitchin and Zachary
                  Ulissi},
  title =        {Beyond Independent Error Assumptions in Large {GNN} Atomistic
                  Models},
  journal =      {The Journal of Chemical Physics},
  volume =       158,
  number =       21,
  pages =        {nil},
  year =         2023,
  doi =          {10.1063/5.0151159},
  url =          {http://dx.doi.org/10.1063/5.0151159},
  DATE_ADDED =   {Sun Sep 24 14:53:46 2023},
}

Copyright (C) 2023 by John Kitchin. See the License for information about copying.

org-mode source

Org-mode version = 9.7-pre

We have a new publication out! In this work we show how to use classification algorithms to find boundaries in science and engineering applications. These applications come up all over the place, for example you may want to know what compositions phase separate, and which ones are single phase, or what conditions lead to degradation and which ones don't. You might want to know which operating parameters have desirable properties, and which don't. In this work we show how to efficiently find the boundaries between these regions using active learning and a classifier. We show that this approach is generally better (more accurate and fewer experiments) than doing a dense grid search.

@article{bhat-2023-sequen-sampl,
  author =       {Maya Bhat and John R. Kitchin},
  title =        {Sequential Sampling Methods for Finding Classification
                  Boundaries in Engineering Applications},
  journal =      {Industrial \& Engineering Chemistry Research},
  volume =       {n/a},
  number =       {n/a},
  pages =        {n/a},
  year =         2023,
  doi =          {10.1021/acs.iecr.3c02362},
  url =          {http://dx.doi.org/10.1021/acs.iecr.3c02362}
}

Check out the Youtube video summary here:

Copyright (C) 2023 by John Kitchin. See the License for information about copying.

org-mode source

Org-mode version = 9.7-pre

Solving inverse problems, where we know what outputs we want from a model and seek the inputs that provide them, is a difficult task. A conventional approach to this problem is to use a nonlinear program (NLP) solver to iteratively find the inputs for a specific output. If you seek a desired output space, then you must solve the NLP many times to map out the corresponding input space. This is often expensive, and tedious to perform. In this work, we demonstrate a new approach to solving this problem that avoids the NLP formulation, and is faster. The idea is simple; we compute a system of differential equations that maps how the input space changes with the output space. Then from a single known point we can integrate a path in the output space to automatically trace the corresponding path in the input space! We compute the system of differential equations using automatic differentiation, and the implicit derivative theorem. We show two examples of this using a steady state continuously stirred tank reactor, which is a set of nonlinear algebraic equations that define the output space from input variables, and another plug flow reactor where the output space is defined by a set of differential equations that must be numerically integrated. In both cases we use automatic differentiation to define the system of ODEs that relate outputs and inputs, and show that the path integration method developed here is as accurate and faster than even the best NLP approach. The idea in this paper is general and applicable to many other systems, not just chemical reactors.

@article{alves-2023,
  author =       {Alves, Victor and Kitchin, John R. and Lima, Fernando V.},
  title =        {An inverse mapping approach for process systems engineering
                  using automatic differentiation and the implicit function
                  theorem},
  journal =      {AIChE Journal},
  year =         2023,
  volume =       {n/a},
  number =       {n/a},
  pages =        {e18119},
  keywords =     {automatic differentiation, implicit function theorem, inverse
                  mapping, inverse problems, process systems engineering},
  doi =          {10.1002/aic.18119},
  url =
                  {https://aiche.onlinelibrary.wiley.com/doi/abs/10.1002/aic.18119}
}

Copyright (C) 2023 by John Kitchin. See the License for information about copying.

org-mode source

Org-mode version = 9.5.5