Utilizing Bayesian Optimization for Efficient Dispersion Curve Feature Acquisition [Slides]

Amber Day, Sinead A. Williamson, Natalie Klein2024 · DOI

This presentation from Los Alamos National Laboratory and the University of Texas at Austin outlines a sophisticated approach to a classic "needle in a haystack" problem in spectroscopy: how do you find critical features in a data curve without measuring every single point?

By applying Bayesian Optimization and Active Learning to Fast Field Cycling (FFC) dispersion curves, the researchers demonstrate a path toward faster, more efficient detection of substances like synthetic opioids.

TL;DR

Researchers are using Bayesian Active Learning to intelligently sample dispersion curves, specifically for identifying synthetic opioids like fentanyl citrate. By employing physics-informed Gaussian Process models and the Active Learning Cohn (ALC) algorithm, they can locate key spectral features while minimizing the total number of expensive measurements required.

What Problem Does This Paper Address?

In the field of nuclear magnetic resonance (NMR) and quadrupole resonance (NQR), "dispersion curves" represent how a substance's relaxation time ($T_1$) changes across different frequencies. For complex molecules like fentanyl citrate, these curves contain specific "features"—peaks or deviations caused by nitrogen atoms—that act as a chemical fingerprint.

The challenge is that acquiring these curves is time-consuming and computationally expensive. Measuring the entire frequency spectrum at high resolution is impractical for rapid screening. Researchers need a way to "smart-sample": identifying where the interesting features are likely to be and focusing their measurement efforts there, while maintaining a general understanding of the overall curve.

Methods

The authors propose a framework based on Bayesian Active Learning using Gaussian Processes (GPs). Key components include:

  • Active Learning Cohn (ALC): Instead of sampling randomly, the system uses the ALC acquisition function. This algorithm chooses the next measurement point ($x_{n+1}$) that will most significantly reduce the average variance (uncertainty) across the entire input space.
  • Physics-Informed Mean Functions: Standard GPs often assume a mean of zero, but dispersion curves have a known "vanilla" shape (a smooth, monotonically increasing trend). The researchers tested several basis functions to model this trend, including:
  • Polynomial Basis Functions.
  • Gaussian Basis Functions.
  • 2-Layered Gaussian Basis Functions, which proved most effective at capturing the underlying signal.
  • Detrending: By modeling the smooth background trend and subtracting it, the researchers can isolate the residuals—the specific "features" or deviations that identify the target substance.
  • Software: The implementation utilizes GPyTorch, a highly efficient library for Gaussian Processes.

Datasets

The study utilizes two primary types of data:

  1. Synthetic Dispersion Curves: Simulated curves featuring one or two nitrogen-related peaks to test the algorithm's ability to "find" known features.
  2. Experimental Data: A real-world $^1H T_1$ dispersion curve for fentanyl citrate. This dataset is characterized by high noise and specific features located between 3.0 and 4.0 MHz.

Key Insights and Ideas

  • The "Vanilla" vs. The Feature: The core insight is that the dispersion curve is a composite of a predictable, smooth background and unpredictable, high-value features. Success depends on accurately modeling the "vanilla" curve so that the features stand out as residuals.
  • Mean Functions Matter: Simple polynomials are insufficient for these physical curves. The transition to multi-layered Gaussian basis functions allowed the model to fit the "true signal" much more closely (achieving an MSE of approximately 0.0002 in simulations).
  • Variance as a Compass: By using the ALC algorithm to minimize integrated variance, the model naturally gravitates toward areas where the data is most "surprising" or uncertain, which is exactly where the chemical features tend to hide.

Key Results

  • Efficient Subsampling: The researchers demonstrated that they could reconstruct the essential shape and features of the fentanyl citrate curve using a fraction of the total data points ("Subsampled data").
  • Feature Localization: The model successfully identified the frequency ranges (e.g., 3.0–3.2 MHz and 3.3–3.7 MHz) where fentanyl citrate exhibits its most distinct deviations.
  • Model Accuracy: The 2-layered Gaussian basis function provided a significantly better fit for the trend than standard polynomial methods, allowing for more precise "detrending" and feature extraction.

Limitations and Future Work

While the results are promising, the presentation notes several areas for continued development:

  • Cost-Aware Sampling: One of the stated goals is to incorporate the "higher cost of sampling at higher frequencies." While the current model reduces the number of samples, it hasn't yet fully integrated a cost-benefit function that accounts for the physical difficulty of high-frequency measurements.
  • Value Weighting: The authors suggest that features at higher frequencies are "more valuable to identify." Future iterations of the acquisition function could be weighted to prioritize these high-value zones even more aggressively.
  • Physics Integration: Further refining the "physics-informed" aspect of the mean function could lead to even more robust models that require even fewer data points to reach a high-confidence identification.

Summary updated

Abstract

Locate features and densely sample near those features. This helps us calculate T<sub>1</sub> which is the magnitude of the deviation from the otherwise smooth monotonically increasing function. Get a generally good idea of the rest of the curve. Potentially incorporate the higher cost of sampling at higher frequencies and the fact that the features at higher frequencies are more valuable to identify. Create a useful physics informed mean function.
BibTeX
@misc{day2024utilizing,
  author = {Amber Day and Sinead A. Williamson and Natalie Klein},
  title = {{Utilizing Bayesian Optimization for Efficient Dispersion Curve Feature Acquisition [Slides]}},
  year = {2024},
  doi = {10.2172/2335729},
}
Figure