Exploiting a Domain Model in an Expert Spectral Analysis Program

David R. BarstowAAAI 1980

Abstract

This paper introduces GAMMA, an expert system designed to identify the elemental composition of unknown substances by interpreting gamma-ray activation spectra. Unlike the "shallow" rule-based AI common in the late 1970s, GAMMA leverages a "deep" six-level domain model of nuclear physics to predict, match, and optimize interpretations. By modeling the physical processes of isotope decay and gamma-ray emission, GAMMA achieves performance comparable to human experts in nuclear physics.

Introduction

Nuclear physicists utilize "activation analysis" to determine the elemental composition of unknown samples. By bombarding a sample with neutrons, stable isotopes are converted into unstable ones that emit gamma rays as they decay. Each isotope possesses a unique "fingerprint" of energy peaks and intensities. However, interpreting these spectra is notoriously difficult due to several factors:

  1. Peak Overlap: Different isotopes often emit gamma rays at very similar energy levels.
  2. Noise and Background: Detectors are imperfect, and background radiation introduces "ghost" peaks.
  3. Dynamic Decay: Isotope concentrations change over time based on their half-lives, meaning the spectrum is non-static.
Figure 1: A complex gamma-ray spectrum illustrating the challenges of peak identification and background noise.

The GAMMA system, developed by David R. Barstow, automates this interpretation process by moving beyond simple lookup tables toward a system that understands the underlying physical processes.

Methods: The Deep Domain Model

GAMMA’s core innovation is its six-level domain model, which represents the physical journey from a raw element to a detected signal: elements in the original sample, their isotopes, isotopes after bombardment, decays, gamma-ray emissions, and finally, detections.

Predict-and-Match Strategy

GAMMA operates "downward" through this model. Given a hypothesis (e.g., "The sample contains Sodium"), it predicts the resulting detections at Level 6 and compares them to the observed data. To handle measurement errors and detector resolution, GAMMA employs a trapezoidal evaluation function.

Figure 2: The trapezoidal rating function used to evaluate the match between predicted energy (E) and observed peaks.

As shown in Figure 2, a peak is considered a "perfect match" (rating 1.0) within a narrow energy range ($\delta_1$). The rating tapers off linearly through a buffer zone ($\delta_2$) until it reaches a penalty level (-1.0) for peaks far from the predicted energy. This allows the system to be robust against minor calibration drifts.

Optimization and Multi-Spectral Analysis

Because multiple elements can contribute to the same peak, GAMMA uses a hill-climbing algorithm to find a set of <element, concentration> pairs that maximizes a global interpretation measure. Furthermore, GAMMA can analyze a sequence of spectra taken at different times. This allows it to use half-life information as a primary discriminator—short-lived isotopes appear early and vanish, while long-lived ones persist.

Results and Performance

GAMMA does not use a modern "training set" but rather a LISP-based knowledge base derived from published nuclear physics data (e.g., Erdmann’s Neutron Activation Tables). Evaluation against known substances, such as salt samples containing Sodium (Na), Chlorine (Cl), and Sulfur (S), demonstrated that GAMMA’s interpretations are generally quite good and compare favorably with those of human experts.

The system is highly accurate at identifying specific isotopes. While slightly less robust at the elemental level due to simplifying assumptions in the transition from isotopes to elements, it rarely produces implausible results. Its use of temporal data (multiple spectra) makes it resilient to sensor noise and software bugs.

Discussion

The development of GAMMA highlights a major architectural divide in early AI: "deep" vs. "shallow" models. While contemporary systems like MYCIN used heuristic "IF-THEN" rules, GAMMA’s success demonstrates that deep models are essential for physical sciences where a computational theory of the domain exists. By modeling the generative process, GAMMA can account for negative evidence—penalizing hypotheses that predict peaks which are absent in the data.

Future work involves improving the hill-climbing optimization to avoid local maxima and refining the model's assumptions about nuclear transitions. Despite its complexity, Barstow notes that GAMMA’s "deep" model is still relatively shallow compared to subatomic physics, suggesting that model depth is always relative to the specific analytical task.

Summary updated

BibTeX
@inproceedings{barstow1980exploiting,
  author = {David R. Barstow},
  title = {{Exploiting a Domain Model in an Expert Spectral Analysis Program}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}

Loading related content...

Figure