Evolutionary Design of Interpretable Spatial Descriptors for Structure-Property Relationship Discovery

Aditya Raghavan, Sergei V KalininarXiv 2026 · arXiv

In the world of materials science, the holy grail is understanding the structure-property relationship: how the arrangement of atoms or domains in a material dictates its physical behavior. While machine learning (ML) has become a powerful tool for predicting these properties from images, it often acts as a "black box." We might get a good prediction, but we don't always know why the model looked at a specific corner of an image or how much context it actually needed.

A new paper from researchers at the University of Tennessee, Knoxville, titled "Evolutionary Design of Interpretable Spatial Descriptors for Structure-Property Relationship Discovery," proposes a clever way to open this box. By using genetic algorithms to "evolve" the way models look at images, they’ve created a framework that balances predictive power with physical interpretability.

TL;DR

The researchers developed a framework that uses genetic algorithms to automatically design optimal, interpretable "spatial descriptors"—customized maps that tell a machine learning model which parts of a microscopy image to focus on and how to summarize that data. This approach allows for high-accuracy property predictions using significantly fewer data points, revealing exactly which spatial features drive a material's physical response.

What Problem Does This Paper Address?

When predicting local properties (like a spectrum or a hysteresis loop) from a microscopy image, researchers face a "Goldilocks" problem with the image neighborhood:

  1. Too small: You miss the surrounding context (like nearby grain boundaries or interfaces) that influences the local physics.
  2. Too large/finely sampled: You overwhelm the model with redundant or noisy information, leading to overfitting and a lack of clarity.

Standard Deep Learning models (like CNNs) handle this internally, but their decision-making process is opaque. This paper seeks to discover the smallest, most informative representation of an image neighborhood that is both highly predictive and human-readable.

Methods

The authors treat the design of an image descriptor as an optimization problem solved by a Genetic Algorithm (GA).

  1. Benchmarking: They first tested five standard regressors (ranging from simple Ridge regression to Deep Convolutional Neural Networks) to establish a baseline.
  2. Spatial Attribution: They used SHAP (SHapley Additive exPlanations) maps to visualize which pixels a trained model actually prioritized.
  3. Evolutionary Construction: This is the core innovation. The GA evolves "chromosomes" that define a spatial partition of the image (e.g., Cartesian grids or Polar sectors). For each region in the partition, the GA decides:
  4. Which operation to use: Should it take the Mean, Minimum, or Maximum intensity?
  5. Whether to discard it: Is this region even necessary for the prediction?
  6. Multi-Objective Optimization: The GA searches for a Pareto front—a set of solutions that represent the best possible trade-offs between minimizing prediction error and minimizing the number of features (dimensionality).
Figure 2: SHAP attribution maps showing where a DCNN focuses to predict ferroelectric switching.

Datasets

The framework was validated on two distinct, publicly available microscopy datasets:

  • STEM–EELS: Scanning Transmission Electron Microscopy paired with Electron Energy-Loss Spectroscopy. This involves atomic-scale images where the goal is to predict local chemical/electronic spectra.
  • PFM–BEPS: Piezoresponse Force Microscopy with Band-Excitation Piezoresponse Spectroscopy. This involves domain-scale images of ferroelectric materials where the goal is to predict local electromechanical hysteresis loops.

Key Insights and Ideas

  • Non-Uniform Detail: The GA-designed descriptors often retain high resolution near the center (the measurement point) and use broader, "pooled" summaries for the surrounding context. This confirms that while context matters, we don't need to see every pixel of it.
  • Non-Linearity via Pooling: Interestingly, assigning min or max pooling operations to regions introduces non-linearity into the descriptor itself. This allows simple, fast linear models (like Ridge regression) to achieve performance previously reserved for complex neural networks.
  • Geometry Matters: The authors explored Cartesian, Polar, and even Hexagonal partitions. They found that the "best" geometry depends on the underlying physics—for example, polar descriptors are naturally suited for capturing radial influences around a probe.
Figure 4: The Pareto front showing the trade-off between descriptor complexity and prediction error.

Key Results

  • Efficiency: The evolved descriptors achieved comparable or better accuracy than uniform image compression while using a fraction of the features. For instance, in the PFM-BEPS task, descriptors with only 4–18 values outperformed Ridge regression using all 169 pixels of a patch.
  • Interpretability: Because each feature in the descriptor has an explicit spatial support (a specific box or slice on the image) and a specific statistic (like "the maximum brightness in this corner"), the resulting model is a direct map of physical influence.
  • Performance: The GA-optimized Cartesian descriptors for STEM-EELS showed a rapid improvement in validation error as the first few regional summaries were added, eventually outperforming standard uniform pooling benchmarks (see Figure 5).
Figure 5: Evolution of Cartesian descriptors, showing how the GA finds a "sweet spot" of complexity.

Limitations and Future Work

  • Search Budget: Genetic algorithms are computationally expensive, requiring the regressor to be refitted thousands of times during the search.
  • Dataset Specificity: While the framework is general, the specific optimal descriptors found are tied to the specific datasets and instruments used.
  • Beyond Images: The authors suggest that this approach could be extended to molecular structures and crystalline solids. Instead of partitioning pixels, the GA would evolve ways to group atoms or coordination shells into interpretable descriptors, helping us discover the fundamental building blocks of material properties.

For a deeper dive into the math and to explore the code, check out the GitHub repository associated with the paper.

Summary updated

Abstract

Structure property prediction is a classical task in machine learning, ranging from molecular activity and materials properties to local functional responses measured in microscopy. For image-based regression, two choices define the information available to the predictor: the size of the image neighborhood and the spatial resolution retained within it. A small neighborhood can omit relevant surrounding structure, whereas a large, finely sampled neighborhood can introduce redundant or irrelevant information. Here, we develop a framework for constructing spatial descriptors for prediction of local spectra and hysteresis loops. We benchmark regression models, use SHAP maps to examine their spatial attribution, and explore the effect of the patch size and compression to establish the relationship between predictive accuracy and descriptor dimensionality. We then use a genetic algorithm to optimize spatial partitions and the assignment of mean, minimum, and maximum pooling operations. Each retained region contributes an explicit measurement of the image, allowing the search to balance prediction error against the number of descriptor values. Cartesian and polar constructions provide complementary ways to organize this information. The resulting descriptors make it possible to examine where spatial detail is needed, where regional summaries are sufficient, and how these choices depend on the predictor and the measured response. The same approach can be used for molecules and solids once the genetic algorithm rules for these are established.
BibTeX
@misc{raghavan2026evolutionary,
  author = {Aditya Raghavan and Sergei V Kalinin},
  title = {{Evolutionary Design of Interpretable Spatial Descriptors for Structure-Property Relationship Discovery}},
  year = {2026},
  eprint = {2611.06811},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.06811},
}
Figure