Characterising particulate random media from near-surface backscattering: A machine learning approach to predict particle size and concentration

Artur L. Gower, Robert M. Gower, Jonathan Deakin, William J. Parnell, I. David AbrahamsEurophysics 2018 · DOI

TL;DR

This paper introduces a machine learning approach to characterize particulate media using only near-surface backscattered waves. They demonstrate that by analyzing the first two moments of the backscattered signal, both particle concentration and size can be accurately predicted, even for challenging scenarios with strong scattering.

What Problem Does This Paper Address?

Measuring the properties of particulate media, such as particle size and concentration, is crucial across many scientific disciplines. While non-invasive wave-based methods are desirable, accurately characterizing these materials, especially at high concentrations where multiple scattering becomes significant, remains a challenge. Traditional methods often struggle with broad ranges of particle sizes and concentrations, and can require complex setups or extensive data. This paper tackles the problem of efficiently and accurately determining particle size and concentration using a simplified measurement setup: a single source/receiver measuring backscattered waves.

Methods

The authors employ a supervised machine learning approach, specifically kernel ridge regression, to build predictive models. They leverage the moments of the backscattered wave field as input features. The core innovation lies in:

  • Numerical Simulation: Developing a robust numerical method (based on the multipole method) to accurately simulate backscattering from an infinite half-space of particles. This allows for the creation of a high-quality, precise dataset for training.
  • Feature Engineering: Investigating which moments of the backscattered signal are most informative. They find that the first moment (mean amplitude) is sufficient for predicting concentration, while the second moment (average intensity) is crucial for accurately determining particle radius.
  • Learning Curves: Using learning curves to rigorously assess the predictive power of their models and to understand the relationship between the input features (moments) and the target properties (radius, concentration). This helps identify the ideal frequency range for measurements.

Datasets

The primary dataset used in this study is synthetically generated using the authors' novel numerical simulation approach. This dataset consists of backscattered wave moments from 205 different simulated media, each with varying particle radii and concentrations. They generated a large number of backscattering simulations (83,000 in total) to ensure high precision and cover a broad range of parameters. This simulated dataset serves as both the training and testing ground for their machine learning models.

Key Insights and Ideas

  • Moment Matters: The choice of statistical moments from the backscattered signal is critical. The mean amplitude alone is surprisingly effective for concentration, but insufficient for radius. The second moment unlocks accurate radius prediction.
  • Frequency Range Optimization: The study identifies an "ideal" frequency range (specifically, wavenumbers $0 < k < 0.8$) for accurately measuring particles within a certain size range ($0 < a < 2$). This is a valuable practical insight for experimental design.
  • Simulation as a Foundation: The ability to generate highly accurate simulated data is a key enabler for training robust machine learning models in this domain, overcoming the challenges of experimental data acquisition.
  • Near-Surface Advantage: By focusing on early arriving backscattered waves (time $t < 98$), the authors effectively simulate and analyze near-surface scattering, simplifying the computational domain while retaining essential information.

Key Results

  • Concentration Prediction: The machine learning model trained on the first moment (mean amplitude) achieved an impressive R² of 0.96 for predicting particle concentration.
  • Radius Prediction: When the second moment (average intensity) was included in the training data, the model's accuracy for predicting particle radius significantly improved, reaching an R² of 0.93. Without the second moment, the radius prediction accuracy was a mere R² of 0.53.
  • Frequency Range: The ideal range for measuring particles with radii $0 < a < 2$ was found to be wavenumbers $0 < k < 0.8$.
  • Robustness: The models demonstrated robustness, with concentration prediction remaining accurate even with reduced training data quality (e.g., fewer training media or larger standard errors).

Limitations and Future Work

  • Dimensionality: The study is limited to a two-dimensional setting. Extending the approach to three dimensions is a natural next step.
  • Particle Type: The current work focuses on particles with Dirichlet boundary conditions. Future work could explore other particle types and boundary conditions.
  • Uncertainty Quantification: While the paper mentions the potential for using Gaussian process regression to quantify uncertainties, this was not the focus of the current study. Quantifying the uncertainty in predictions is a crucial area for future research.
  • Real-World Data: The models are currently trained on simulated data. Adapting and validating these models with real experimental data is essential for practical deployment.
  • Computational Efficiency: For larger simulations or more complex scenarios, exploring more efficient simulation techniques like multilevel Monte Carlo or fast multipole methods is suggested.

Summary updated

Abstract

To what extent can particulate random media be characterised using direct wave backscattering from a single receiver/source? Here, in a two-dimensional setting, we show using a machine learning approach that both the particle radius and concentration can be accurately measured when the boundary condition on the particles is of Dirichlet type. Although the methods we introduce could be applied to any particle type. In general backscattering is challenging to interpret for a wide range of particle concentrations, because multiple scattering cannot be ignored, except in the very dilute range. Across the concentration range from 1% to 20% we find that the mean backscattered wave field is sufficient to accurately determine the concentration of particles. However, to accurately determine the particle radius, the second moment, or average intensity, of the backscattering is necessary. We are also able to determine what is the ideal frequency range to measure a broad range of particles sizes. To get rigorous results with supervised machine learning requires a large, highly precise, dataset of backscattered waves from an infinite half-space filled with particles. We are able to create this dataset by introducing a numerical approach which accurately approximates the backscattering from an infinite half-space.
BibTeX
@inproceedings{gower2018characterising,
  author = {Artur L. Gower and Robert M. Gower and Jonathan Deakin and William J. Parnell and I. David Abrahams},
  title = {{Characterising particulate random media from near-surface backscattering: A machine learning approach to predict particle size and concentration}},
  booktitle = {Europhysics},
  year = {2018},
  doi = {10.1209/0295-5075/122/54001},
}
Figure