Genetic descriptor search algorithm for predicting hydrogen adsorption free energy of 2D material

Jaehwan Lee, Seokwon Shin, Jaeho Lee, Young‐Kyu Han, Woojin Lee, Youngdoo SonScientific 2023 · DOI

TL;DR

This paper introduces a novel Genetic Descriptor Search (GDS) algorithm that efficiently identifies key material descriptors for predicting hydrogen adsorption free energy in 2D materials. By leveraging a genetic algorithm and symbolic regression, GDS significantly outperforms existing feature selection methods, offering a computationally efficient and interpretable approach for materials discovery.

What Problem Does This Paper Address?

The hydrogen evolution reaction (HER) is a crucial process for clean energy technologies, and transition metal dichalcogenides (TMDs) are promising candidates for electrocatalysts. The efficiency of these catalysts is heavily influenced by the hydrogen adsorption free energy (AGH) on their surface, which dictates how strongly they bind hydrogen. Accurately predicting AGH is essential for designing better catalysts. However, traditional methods like Density Functional Theory (DFT) are computationally expensive. While machine learning (ML) approaches can accelerate predictions, they often require vast amounts of training data and can be prone to overfitting or poor generalization without sufficient samples. Existing descriptor search methods, like SISSO, can be computationally intensive due to the large search space for descriptors. This paper aims to bridge this gap by developing a more efficient and effective descriptor search algorithm that requires less computational power and data.

Methods

The core of this work is the Genetic Descriptor Search (GDS) algorithm. GDS combines symbolic regression with a genetic algorithm to efficiently explore the space of potential material descriptors.

  • Symbolic Regression: This technique is used to generate mathematical expressions (represented as expression trees) that relate material features to the target property (AGH). Unlike traditional regression, it doesn't assume a fixed model structure, allowing it to discover novel relationships.
  • Genetic Algorithm (GA): GDS employs a GA to evolve a population of these expression trees. The GA iteratively refines the descriptors by mimicking natural evolution:
  • Initialization: An initial population of expression trees is created using primary material features and a set of mathematical operators.
  • Fitness Score: Each tree (descriptor set) is evaluated based on its ability to predict the residual of the AGH (the part not explained by previously found descriptors). This score also penalizes tree complexity (depth) to encourage simpler, more interpretable descriptors.
  • Tournament Selection: A robust selection method where a subset of trees competes, and the fittest ones are chosen to "reproduce."
    • Genetic Operators: Crossover, subtree mutation, point mutation, and hoist mutation are used to create new, potentially improved descriptors (offspring) from the selected parents. Hoist mutation is particularly noted for reducing tree complexity.
  • Iterative Refinement: The process is repeated. In each iteration, GDS finds a new descriptor that best explains the remaining error in AGH prediction, building upon the descriptors found in previous steps. This iterative approach ensures that each new descriptor adds unique predictive power.

Datasets

The study utilized a dataset of 70 Transition Metal Dichalcogenides (TMDs). This dataset was previously used in a related study.

  • Primary Features: 27 primary features were extracted for each TMD. These include properties like atomic radii, electron numbers, bond lengths, and quantum numbers. Some features were obtained from publicly available databases, while others were derived from structural information or calculations.
  • Target Property: The hydrogen adsorption free energy (AGH) for each TMD was the target property to be predicted.

Key Insights and Ideas

  • Descriptor Discovery without Domain Expertise: A significant finding is that GDS can identify descriptors that are not only predictive but also align with existing chemical knowledge and literature, even without explicit prior domain input. This suggests the algorithm can uncover fundamental relationships.
  • Efficiency through Residual Learning: Instead of trying to predict the entire AGH at once, GDS iteratively predicts the residual error. This allows each newly discovered descriptor to focus on explaining the remaining unexplained variance, leading to more efficient descriptor discovery.
  • Balancing Prediction and Interpretability: The fitness function's penalty for tree depth is crucial. It ensures that while predictive power is maximized, the resulting descriptors remain relatively simple and interpretable, avoiding overly complex "black box" expressions.
  • Robustness through Ensemble Potential: The inherent randomness in the genetic algorithm allows for the generation of diverse descriptor sets. Averaging models built on these different sets (ensemble) leads to more robust and stable performance, reducing variance.

Key Results

  • Superior Predictive Performance: GDS significantly outperformed several established feature extraction methods (PCA variants, manifold learning) and the SISSO descriptor search algorithm in predicting AGH.
  • Quantitative Improvement: When using up to 10 descriptors, GDS consistently achieved higher R-squared values compared to all other methods. For instance, the GDS-based models showed a steady increase in performance with more descriptors, whereas other methods plateaued or showed less improvement.
  • Qualitative Validation (t-SNE): Visualizing the data in a reduced dimensional space using t-SNE showed that the descriptors generated by GDS effectively clustered the TMDs and aligned them with their AGH values, indicating a meaningful representation of the underlying physics. Raw primary features, in contrast, showed little discernible relationship with AGH.
  • Interpretability of Descriptors: Analysis of the generated descriptors revealed that they often incorporated physically meaningful variables (e.g., covalent radius, average valence electron number, distance to outermost electron) that are known to influence AGH, aligning with findings in existing literature.

Limitations and Future Work

  • Computational Cost of GA: While GDS is more efficient than exhaustive search, the genetic algorithm itself can still be computationally intensive, especially with large population sizes or many generations.
  • Randomness and Reproducibility: The inherent randomness of the GA means that descriptor search results can vary across trials. The paper addresses this by sampling multiple sets of descriptors for analysis.
  • Future Work: The authors suggest that incorporating domain expertise into the GDS process (e.g., by restricting primary features or operators) could further improve performance. They also highlight that GDS is not task-specific and can be applied to predict other material properties and for different material classes.

Summary updated

Abstract

Transition metal dichalcogenides (TMDs) have emerged as a promising alternative to noble metals in the field of electrocatalysts for the hydrogen evolution reaction. However, previous attempts using machine learning to predict TMD properties, such as catalytic activity, have been shown to have limitations in their dependence on large amounts of training data and massive computations. Herein, we propose a genetic descriptor search that efficiently identifies a set of descriptors through a genetic algorithm, without requiring intensive calculations. We conducted both quantitative and qualitative experiments on a total of 70 TMDs to predict hydrogen adsorption free energy ([Formula: see text]) with the generated descriptors. The results demonstrate that the proposed method significantly outperformed the feature extraction methods that are currently widely used in machine learning applications.
BibTeX
@inproceedings{lee2023genetic,
  author = {Jaehwan Lee and Seokwon Shin and Jaeho Lee and Young‐Kyu Han and Woojin Lee and Youngdoo Son},
  title = {{Genetic descriptor search algorithm for predicting hydrogen adsorption free energy of 2D material}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-39696-0},
}
Figure