Recreation of the periodic table with an unsupervised machine learning algorithm

Minoru Kusaba, Chang Liu, Yukinori Koyama, Kiyoyuki Terakura, Ryo Yoshida

TL;DR

This paper introduces a novel unsupervised machine learning algorithm, the Periodic Table Generator (PTG), that can automatically recreate tabular representations of chemical elements based on their physicochemical properties. The PTG, built upon Generative Topographic Mapping (GTM), demonstrates the potential of AI to discover and visualize underlying periodic trends, offering a new perspective on element descriptors for materials science applications.

What Problem Does This Paper Address?

The periodic table, a cornerstone of chemistry, is a remarkable human achievement in compressing complex element properties into an intuitive 2D grid. While numerous variations exist, they are all human-designed. This paper asks: can machine learning, specifically unsupervised learning, independently "recreate" or "discover" such a periodic table using only the raw physicochemical data of elements? Existing attempts using machine learning (like Self-Organizing Maps) have struggled to fully capture the nuanced periodicities that Mendeleev's table elegantly represents, often missing crucial between-group and between-family similarities. The challenge lies in mapping high-dimensional element data into a low-dimensional, structured tabular format that reflects underlying chemical periodicity.

Methods

The core of this work is the Periodic Table Generator (PTG), an unsupervised machine learning algorithm. The PTG is based on Generative Topographic Mapping with latent variable dependent length-scale and variance (GTM-LDLV). This approach allows for flexible representation of complex data landscapes.

The PTG employs a three-step coarse-to-fine strategy to ensure accurate, one-to-one assignments of elements to nodes in the latent space, avoiding undesirable matching and computational issues:

  1. Initial GTM-LDLV Training: A small number of nodes (K < N, where N is the number of elements) are used to train the GTM-LDLV model.
  2. Node Expansion: The number of nodes is increased (K > N), and new nodes are generated by interpolating from the existing GTM-LDLV model. This expands the latent space.
  3. Fine-tuning with One-to-One Assignment: The GTM-LDLV model is fine-tuned with the expanded node set, but crucially, with a constraint for a one-to-one mapping between elements and nodes. This step uses a transportation problem formulation (solved with lpSolve) to find the optimal assignment.

This iterative process allows the algorithm to first capture broad trends and then refine the structure for precise element placement.

Datasets

  • Element Features: A dataset of 39 physicochemical features (including melting point, electronegativity, etc.) for 54 elements (atomic number 1-54, Hydrogen to Xenon). This dataset was derived from the XenonPy library. Features were standardized to have zero mean and unit variance.
  • Compound Formation Energies: A dataset of formation energies per atom for 12,373 inorganic compounds, obtained from the Materials Project database. This dataset was used for evaluating the predictive power of the generated periodic tables as element descriptors.

Key Insights and Ideas

  • AI can "discover" periodic trends: The PTG successfully generates tabular arrangements of elements that reflect underlying periodic laws, demonstrating that machine learning can autonomously identify and represent these fundamental chemical relationships.
  • Flexibility in layout: The PTG is not limited to a 2D grid; it can generate various layouts (e.g., square, conical) based on the data and desired structure, offering new ways to visualize element relationships.
  • Periodic tables as powerful descriptors: The study highlights that the coordinates of elements within a generated periodic table can serve as effective "element descriptors." These descriptors can then be used to predict material properties, suggesting a data-driven approach to materials discovery.
  • Conical layout excels for certain properties: The 3D conical layout, in particular, showed superior performance in predicting formation energies compared to the standard periodic table, PCA, or t-SNE, suggesting it better captures the relevant element features for this task.
  • Visualizing feature landscapes: Mapping individual element features (like atomic radius or melting point) onto the generated tables provides intuitive "property landscapes," revealing how these properties are compressed and distributed in the latent space.

Key Results

  • The PTG successfully generated a square table similar to the standard periodic table, but also a novel 3D conical table.
  • The square table clustered elements by group (e.g., alkali metals, halogens) and separated metals from nonmetals.
  • The conical table arranged elements radially by group and showed a more continuous representation of s-block elements, including Helium.
  • In predicting formation energies of inorganic compounds, the conical PTG achieved the best performance with a Mean Absolute Error (MAE) of 0.464 eV/atom and Root Mean Square Error (RMSE) of 0.643 eV/atom.
  • This performance significantly outperformed the standard periodic table (MAE: 0.549, RMSE: 0.734), PCA (MAE: 0.631, RMSE: 0.830), and t-SNE (MAE: 0.667, RMSE: 0.859).
  • Analysis of the prediction results revealed that the conical table better represented transition elements in the fourth period and Aluminium, which were enriched in compounds predicted accurately by the conical descriptor.

Limitations and Future Work

  • Limited element set: The study focused on elements with atomic numbers 1-54. Future work could extend this to include heavier elements.
  • Specific property prediction: The evaluation of element descriptors was primarily based on predicting formation energies. The authors suggest incorporating more diverse material properties to further validate and refine the descriptor quality.
  • Layout exploration: While the PTG can generate various layouts, the study explored only a few (square, conical, and others in supplementary figures). Exploring a wider range of potentially promising layouts is a future direction.
  • Interpretability of complex features: While some feature landscapes were visualized, a deeper understanding of how all 39 features are compressed and represented in the latent space could be explored further.
  • Comparison to human-designed tables: While the PTG aims to recreate periodic trends, directly comparing its "interpretability" to human-designed tables across various chemical contexts remains an area for further investigation.

Summary updated

Abstract

In 1869, the first draft of the periodic table was published by Russian chemist Dmitri Mendeleev. In terms of data science, his achievement can be viewed as a successful example of feature embedding based on human cognition: chemical properties of all known elements at that time were compressed onto the two-dimensional grid system for a tabular display. In this study, we seek to answer the question of whether machine learning can reproduce or recreate the periodic table by using observed physicochemical properties of the elements. To achieve this goal, we developed a periodic table generator (PTG). The PTG is an unsupervised machine learning algorithm based on the generative topographic mapping, which can automate the translation of high-dimensional data into a tabular form with varying layouts on-demand. The PTG autonomously produced various arrangements of chemical symbols, which organized a two-dimensional array such as Mendeleev's periodic table or three-dimensional spiral table according to the underlying periodicity in the given data. We further showed what the PTG learned from the element data and how the element features, such as melting point and electronegativity, are compressed to the lower-dimensional latent spaces.
BibTeX
@inproceedings{kusaba2021recreation,
  author = {Minoru Kusaba and Chang Liu and Yukinori Koyama and Kiyoyuki Terakura and Ryo Yoshida},
  title = {{Recreation of the periodic table with an unsupervised machine learning algorithm}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-81850-z},
}
Figure