A new approach for indexing powder diffraction data based on whole-profile fitting and global optimization using a genetic algorithm

Benson M. Kariuki, S. A. Belmonte, M. I. McMahon, Roy L. Johnston, Kenneth D. Harris, R. J. NelmesJournal of Synchrotron Radiation 1999 · DOI

This 1999 paper by Kariuki et al. introduces a robust method for "indexing" powder diffraction data—the process of determining a crystal's unit-cell dimensions—by moving away from traditional peak-picking. Instead, the authors use a Genetic Algorithm (GA) to navigate the complex landscape of whole-profile fits, allowing for successful indexing even in the presence of significant peak overlap and impurities.

TL;DR

The authors replace fragile, peak-based indexing methods with a global optimization strategy that fits the entire diffraction profile using a Genetic Algorithm. This approach effectively "ignores" the noise of impurities and handles overlapping peaks that typically cause traditional indexing software to fail.

What Problem Does This Paper Address?

In crystallography, "indexing" is the critical first step of structure determination. It involves finding the lattice parameters $\{a, b, c, \alpha, \beta, \gamma\}$ that define the size and shape of the crystal's repeating unit cell.

Historically, indexing relied on identifying the exact positions of a few dozen sharp peaks. However, this "peak-picking" approach is highly sensitive to:

  1. Peak Overlap: In low-symmetry or large unit cells, peaks crowd together, making it impossible to resolve individual maxima.
  2. Impurities: Extra peaks from a secondary material can lead traditional algorithms down a dead-end path.
  3. Experimental Errors: Small shifts in peak positions due to sample displacement or zero-point errors can ruin the calculation.

The authors sought a method that uses the entire dataset (the whole profile) rather than just a few selected points, making the process more resilient to these common experimental headaches.

Methods

The researchers combined two powerful concepts: Whole-Profile Fitting and Genetic Algorithms (GA).

  • Whole-Profile Fitting (Le Bail Method): Instead of picking peaks, the algorithm takes a trial set of lattice parameters and performs a "Le Bail fit." This partitions the observed intensity among the possible reflections allowed by those parameters. The quality of the fit is measured by the weighted profile R-factor ($R_{wp}$). A lower $R_{wp}$ means the trial lattice parameters are more likely to be correct.
  • Genetic Algorithm (GA): To find the global minimum of $R_{wp}$ in the multi-dimensional parameter space, the authors used an evolutionary approach.
  • Population: A set of random trial lattice parameters.
  • Mating & Mutation: Successful "parent" parameters (those with low $R_{wp}$) are combined and slightly altered to create "offspring."
    • Natural Selection: Only the "fittest" (best-fitting) parameters survive to the next generation.
  • Regional Weighting: To ensure the GA didn't ignore weak but important peaks, the authors split the diffraction pattern into regions and calculated $R_{wp}$ for each, preventing high-intensity peaks from dominating the optimization.

Datasets

The primary evaluation was performed on experimental powder X-ray diffraction data of GaAs-II (an orthorhombic high-pressure phase).

  • The Challenge: This specific dataset contained noticeable impurity peaks.
  • Comparison: The authors noted that "all of the commonly available indexing programs" (like ITO, TREOR, and DICVOL) failed to index this pattern because they couldn't distinguish the GaAs-II peaks from the impurity peaks.

Key Insights and Ideas

  • Information Density: By using the whole profile, the method utilizes all $d$-spacing information, including the shapes and widths of peaks, which traditional methods discard.
  • The "Schemata" Effect: The authors highlight that certain lattice parameters (like the $a$-axis in a tetragonal system) might correctly fit a subset of peaks even if other parameters are wrong. The GA is particularly good at identifying and "propagating" these partially correct solutions (schemata) toward the global optimum.
  • Robustness to Impurities: Because an impurity phase only contributes a constant background-like error to the $R_{wp}$ of the correct majority phase, the global minimum for the majority phase remains findable. Peak-picking methods, by contrast, treat impurity peaks as "truth," leading to total failure.

Key Results

  • Successful Indexing: The GA successfully identified the correct orthorhombic unit cell for GaAs-II ($a \approx 4.97, b \approx 4.78, c \approx 5.27$ Å) despite the presence of impurities.
  • Convergence: The $R_{wp}$ dropped sharply within the first 100 generations and stabilized by generation 500, demonstrating that the GA could efficiently navigate the search space without an exhaustive grid search.
  • Visual Confirmation: The "valleys" in the $R_{wp}$ hypersurface (shown in Figure 2) clearly lead to the correct solution, validating the use of $R_{wp}$ as a reliable fitness landscape for global optimization.

Limitations and Future Work

  • Peak Shape Modeling: The method requires a reasonable initial estimate of peak shapes and widths. If the peak model is poor, the $R_{wp}$ might not accurately reflect the "fitness" of the lattice parameters.
  • Computational Cost: While more efficient than a grid search, running a Le Bail fit for every member of a population over hundreds of generations was computationally intensive for 1999 hardware (though trivial by modern standards).
  • Future Integration: The authors suggested that this indexing approach could be integrated directly with "direct-space" structure solution methods, potentially creating a "one-pot" pipeline from raw diffraction data to a finished crystal structure.

Summary updated

Abstract

This paper describes a new technique for indexing powder diffraction data. The lattice parameters (unit-cell dimensions) { a , b , c ,α,β,γ} define the parameter space of the problem, and the aim is to find the optimal lattice parameters for a given experimental powder diffraction pattern. Conventional methods for indexing consider the measured positions of a limited number of peak maxima, whereas this new approach considers the whole powder diffraction profile. This new strategy offers several advantages, which are discussed fully. In this approach, the quality of a given set of lattice parameters is determined from the profile R -factor, R wp , obtained following a Le Bail-type fit of the intensity profile. To find the correct lattice parameters ( i.e. the global minimum in R wp ), a genetic algorithm has been used to explore the R wp ( a , b , c ,α,β,γ) hypersurface. (Other methods for global minimization, such as Monte Carlo and simulated annealing, may also be effective.) Initially, a number of trial sets of lattice parameters are generated at random, and this `population' is then allowed to evolve subject to well defined evolutionary procedures for mating, mutation and natural selection (the fitness of each member of the population is determined from its value of R wp ). Examples are presented to demonstrate the success and underline the potential of this new approach for indexing powder diffraction data.
BibTeX
@article{kariuki1999new,
  author = {Benson M. Kariuki and S. A. Belmonte and M. I. McMahon and Roy L. Johnston and Kenneth D. Harris and R. J. Nelmes},
  title = {{A new approach for indexing powder diffraction data based on whole-profile fitting and global optimization using a genetic algorithm}},
  journal = {Journal of Synchrotron Radiation},
  year = {1999},
  doi = {10.1107/s0909049598017932},
}
Figure