A Data-Driven Construction of the Periodic Table of the Elements

Michael J. Willatt, Félix Musil, Michele CeriottiarXiv 2018 · arXiv

Abstract

Machine-learning of atomic-scale properties amounts to extracting correlations between structure, composition and the quantity that one wants to predict. Representing the input structure in a way that best reflects such correlations makes it possible to improve the accuracy of the model for a given amount of reference data. When using a description of the structures that is transparent and well-principled, optimizing the representation might reveal insights into the chemistry of the data set. Here we show how one can generalize the SOAP kernel to introduce a distance-dependent weight that accounts for the multi-scale nature of the interactions, and a description of correlations between chemical species. We show that this improves substantially the performance of ML models of molecular and materials stability, while making it easier to work with complex, multi-component systems and to extend SOAP to coarse-grained intermolecular potentials. The element correlations that give the best performing model show striking similarities with the conventional periodic table of the elements, providing an inspiring example of how machine learning can rediscover, and generalize, intuitive concepts that constitute the foundations of chemistry.

Introduction

Predicting the properties of molecules and materials through machine learning (ML) requires sophisticated representations of local atomic environments. A prominent method is the Smooth Overlap of Atomic Positions (SOAP) kernel. However, standard SOAP faces two significant challenges: the "Distance Problem," where it fails to prioritize short-range interactions effectively, and the "Alchemical Problem," where it treats every chemical element as a distinct, unrelated entity. In systems with high elemental diversity, this leads to a combinatorial explosion of features, necessitating vast amounts of training data to learn the relationships between every possible pair of elements.

Methodology

To address these limitations, the authors introduced two primary enhancements to the SOAP framework:

  1. Radially-scaled Kernels: A weight function $u(r)$ scales the importance of atoms based on distance, prioritizing short-range interactions while maintaining long-range context.
  2. Alchemical Kernels: Elements are projected into a low-dimensional "alchemical space." Each element is represented by a vector of weights, and similarity is calculated as the overlap of these vectors.
  3. Feature Optimization: Instead of manual selection, optimal radial scaling and alchemical weights are determined via data-driven cross-validation.

Results and Discussion

Performance on Complex Systems

The researchers evaluated their approach on the Elpasolite dataset, which features crystalline structures across 39 different main-group elements. By optimizing the alchemical kernels, the model achieved superior performance compared to previous state-of-the-art methods.

Learning curves for Elpasolite crystals comparing different alchemical dimensions ($d_J$) against standard benchmarks.

As shown in Figure 1, even with a drastic reduction in feature dimensionality (using $d_J = 2$ or $4$), the model maintains high accuracy. This dimensionality reduction—up to 99%—allows for more efficient learning in elementally diverse systems.

Rediscovering the Periodic Table

The most striking outcome of the alchemical optimization is the autonomous reconstruction of chemical relationships. When the model optimized weights for the 39 elements in the Elpasolite set, the resulting 2D projection revealed clusters that mirror the traditional Periodic Table.

Data-driven reconstruction of the periodic table through alchemical kernel optimization.

In Figure 2, elements from the same groups (e.g., alkali metals, halogens) naturally cluster together. This demonstrates that the ML model, purely by optimizing for predictive accuracy of material stability, "rediscovered" the fundamental chemical principles that govern elemental behavior.

Molecular Stability and Radial Scaling

On the QM9 dataset of organic molecules, the introduction of radial scaling improved accuracy by approximately 25%. The final model reached a Mean Absolute Error (MAE) of 0.14 kcal/mol, achieving the rigorous standard of "chemical accuracy."

Conclusion

By integrating physical intuition—such as distance-dependent weighting and elemental similarity—into the SOAP framework, the authors have created a more robust and interpretable ML model. This approach not only pushes the boundaries of predictive accuracy for multi-component systems but also provides a transparent window into the underlying chemistry, as evidenced by the data-driven emergence of the Periodic Table.

Summary updated

Abstract

Machine-learning of atomic-scale properties amounts to extracting correlations between structure, composition and the quantity that one wants to predict. Representing the input structure in a way that best reflects such correlations makes it possible to improve the accuracy of the model for a given amount of reference data. When using a description of the structures that is transparent and well-principled, optimizing the representation might reveal insights into the chemistry of the data set. Here we show how one can generalize the SOAP kernel to introduce a distance-dependent weight that accounts for the multi-scale nature of the interactions, and a description of correlations between chemical species. We show that this improves substantially the performance of ML models of molecular and materials stability, while making it easier to work with complex, multi-component systems and to extend SOAP to coarse-grained intermolecular potentials. The element correlations that give the best performing model show striking similarities with the conventional periodic table of the elements, providing an inspiring example of how machine learning can rediscover, and generalize, intuitive concepts that constitute the foundations of chemistry.
BibTeX
@misc{willatt2018data,
  author = {Michael J. Willatt and Félix Musil and Michele Ceriotti},
  title = {{A Data-Driven Construction of the Periodic Table of the Elements}},
  year = {2018},
  eprint = {1807.00236},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.00236},
}
Figure