Integrating Clustering and Classification for Estimating Process Variables in Materials Science

Aparna S. Varde, Elke A. Rundensteiner, Carolina Ruiz, David C. Brown, Mohammed Maniruzzaman, Richard D. Sisson Jr.AAAI 2006

Abstract

Estimating complex experimental outcomes in materials science is often hindered by the "semantic gap" between mathematical similarity and domain-specific relevance. This paper introduces AutoDomainMine, a framework that automates the analysis of heat treatment experiments by integrating clustering and classification. By learning domain-specific distance metrics and generating optimized cluster representatives, the system achieves up to 91% accuracy in predicting heat transfer curves from input conditions, significantly outperforming traditional similarity-search methods.

Introduction

In materials science, understanding how materials react to processes like quenching (rapid cooling) is critical. These results are typically represented as heat transfer curves—graphs plotting heat transfer ($h$) against temperature ($T$). However, physical experiments are resource-intensive. While computational estimation is a logical alternative, existing mathematical models often fail to capture the nuances that human experts prioritize, leading to a "semantic gap."

Methodology

The AutoDomainMine framework mimics the expert workflow of grouping similar results and identifying their causal factors through a multi-stage process:

  1. Clustering and Classification: The system clusters existing experimental graphs using $k$-means. It then employs decision tree algorithms (such as ID3 or J4.8) to learn "clustering criteria," mapping specific input conditions (e.g., quenchant type, agitation) to resulting graph clusters.
  2. LearnMet: To bridge the semantic gap, the authors developed LearnMet. This component uses a weighted sum of various distance metrics, iteratively adjusting weights until computational clusters align with expert-provided ground truth.
  3. DesRept: To represent clusters effectively, the DesRept method uses the Minimum Description Length (MDL) principle to build "combined representatives." These include superimposed graphs and sorted condition tables, ensuring that the estimated output retains critical information for decision support.
The AutoDomainMine framework workflow, integrating clustering of experimental graphs with decision tree classification to learn domain-specific criteria.

Experimental Setup and Data

The research focuses on the Heat Treating of Materials domain, utilizing data from the Center for Heat Treating Excellence (CHTE). The model was trained on historical experimental data and validated against a distinct test set of real-world experiments to ensure generalizability.

Key Results and Discussion

The system was evaluated based on its ability to estimate graphs within a 10% error threshold of actual experimental data. The integration of specialized components showed clear incremental improvements:

  • Baseline (Similarity Search): 65% accuracy.
  • Basic AutoDomainMine: 75% accuracy.
  • With LearnMet: 86% accuracy.
  • With DesRept (Final Model): 91% accuracy.

These results demonstrate that automating "expert intuition"—specifically how scientists prioritize certain graph features—is key to high-fidelity estimation. By treating cluster representation as an information compression problem (MDL), the framework ensures that the most relevant physical behaviors are preserved.

Conclusion

AutoDomainMine successfully bridges the gap between raw data and domain expertise in materials science. While the current implementation is tailored to heat treating, the underlying principles of learned distance metrics and MDL-based representation are applicable to broader scientific graphical data. Future work will focus on refining the weight adjustment heuristics to further reduce the need for initial expert intervention.

Summary updated

BibTeX
@inproceedings{varde2006integrating,
  author = {Aparna S. Varde and Elke A. Rundensteiner and Carolina Ruiz and David C. Brown and Mohammed Maniruzzaman and Richard D. Sisson Jr.},
  title = {{Integrating Clustering and Classification for Estimating Process Variables in Materials Science}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure