Synthetic data enable experiments in atomistic machine learning

Gardner, John L. A.; Beaulieu, Zoe Faure; Deringer, Volker L.

doi:10.1039/d2dd00137c

Cited by 5 publications

(14 citation statements)

References 91 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Although this data set is extensive and achieves substantial coverage of the reaction space, the number of occurrences a particular catalyst is used varies widely which may impact the reliability of the generality metrics derived from this literature curated data set (see above discussion). Consequently, and to explore different substrate–catalyst combinations more comprehensively, we investigated several robust non-linear machine learning (ML) regression techniques for correlating the enantioselectivity outcomes represented as ΔΔ G ‡ to the structure of the reaction components. , The resulting models could then be deployed to create a virtual data set by predicting the enantioselectivity for every combination of imine, nucleophile, and catalyst contained in the experimental database.…”

Section: Resultsmentioning

confidence: 99%

A Data-Driven Workflow for Assigning and Predicting Generality in Asymmetric Catalysis

et al. 2023

View full text Add to dashboard Cite

The development of chiral catalysts that can provide high enantioselectivities across a wide assortment of substrates or reaction range is a priority for many catalyst design efforts. While several approaches are available to aid in the identification of general catalyst systems, there is currently no simple procedure for directly measuring how general a given catalyst could be. Herein, we present a catalyst-agnostic workflow centered on unsupervised machine learning that enables the rapid assessment and quantification of catalyst generality. The workflow uses curated literature data sets and reaction descriptors to visualize and cluster chemical space coverage. This reaction network can then be applied to derive a catalyst generality metric through designer equations and interfaced with other regression techniques for general catalyst prediction. As validating case studies, we have successfully applied this method to identify-through-quantification the most general catalyst chemotype for an organocatalytic asymmetric Mannich reaction and predicted the most general chiral phosphoric acid catalyst for the addition of nucleophiles to imines. The mechanistic basis for catalyst generality can then be gleaned from the calculated values by deconstructing the contributions of chemical space and enantiomeric excess to the overall result. Finally, our generality techniques permitted the development of mechanistically informative catalyst screening sets that allow experimentalists to rationally select catalysts that have the highest probability of achieving a good result in the first round of reaction development. Overall, our findings represent a framework for interrogating catalyst generality, and this strategy should be relevant to other catalytic systems widely applied in asymmetric synthesis.

show abstract

Section: Resultsmentioning

confidence: 99%

A Data-Driven Workflow for Assigning and Predicting Generality in Asymmetric Catalysis

et al. 2023

View full text Add to dashboard Cite

show abstract

“…In conclusion, our work has shown that local-environment energies in ZIFs can be “machine-learned” using cg structural representations, with less than a factor of 2 loss of accuracy compared to established, fully atomistic approaches. In doing so, we showed that local energies from an empirical force field for ZIFs 14 can be readily available “synthetic” regression targets – extending prior work in the field of atomistic ML 15,16 to the construction of cg models. Chemically, our results provide direct and quantitative support for the long-standing idea that there exists a mapping between ZIFs and zeolites (Fig.…”

mentioning

confidence: 76%

“…The hypothetical ZIF dataset, including technical details of how it was constructed, is available at . In addition to per-cell (total) energies, this model by construction yields per-atom (local) energies – allowing us to build a “synthetic” dataset with which the properties of ML models can be studied, following ideas in ref. 15 and 16.…”

mentioning

confidence: 99%

Coarse-grained versus fully atomistic machine learning for zeolitic imidazolate frameworks

Faure Beaulieu,

Nicholas,

Gardner

et al. 2023

Chem. Commun.

Self Cite

View full text Add to dashboard Cite

show abstract

“…More recently, researchers have been actively exploiting these locally predicted values to interpret the local stability of chemical environments in complex phases, 46−50 guide structural optimization, 51 and even use them as synthetic data for the pretraining of large neural network (NN) models. 52 While the practical benefits of local decomposition for atomistic ML are clear, one must be mindful of how reliable, or "robust", the resulting local predictions are. Since only the global quantity is rigorously defined, its decomposition into local contributions can take place in numerous different ways.…”

Section: Introductionmentioning

confidence: 99%

“…In the case of ML models of the electronic density of states, − a plausible correlation could be found between different local structural motifs and how they “contribute” to the total density of states. More recently, researchers have been actively exploiting these locally predicted values to interpret the local stability of chemical environments in complex phases, − guide structural optimization, and even use them as synthetic data for the pretraining of large neural network (NN) models …”

Section: Introductionmentioning

confidence: 99%

Robustness of Local Predictions in Atomistic Machine Learning Models

Chong,

Grasselli,

Ben Mahmoud

et al. 2023

J. Chem. Theory Comput.

Self Cite

View full text Add to dashboard Cite

Machine learning (ML) models for molecules and materials commonly rely on a decomposition of the global target quantity into local, atom-centered contributions. This approach is convenient from a computational perspective, enabling large-scale ML-driven simulations with a linear-scaling cost and also allows for the identification and posthoc interpretation of contributions from individual chemical environments and motifs to complicated macroscopic properties. However, even though practical justifications exist for the local decomposition, only the global quantity is rigorously defined. Thus, when the atom-centered contributions are used, their sensitivity to the training strategy or the model architecture should be carefully considered. To this end, we introduce a quantitative metric, which we call the local prediction rigidity (LPR), that allows one to assess how robust the locally decomposed predictions of ML models are. We investigate the dependence of the LPR on the aspects of model training, particularly the composition of training data set, for a range of different problems from simple toy models to real chemical systems. We present strategies to systematically enhance the LPR, which can be used to improve the robustness, interpretability, and transferability of atomistic ML models.

show abstract

Synthetic data enable experiments in atomistic machine learning

Abstract: Machine-learning models are increasingly used to predict properties of atoms in chemical systems. There have been major advances in developing descriptors and regression frameworks for this task, typically starting from...

Cited by 5 publications

References 91 publications

A Data-Driven Workflow for Assigning and Predicting Generality in Asymmetric Catalysis

A Data-Driven Workflow for Assigning and Predicting Generality in Asymmetric Catalysis

Coarse-grained versus fully atomistic machine learning for zeolitic imidazolate frameworks

Robustness of Local Predictions in Atomistic Machine Learning Models

Contact Info

Product

Resources

About