About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
AI-Enabled Materials Processing: Integrating Accelerated Experimental Workflows and Processing-Aware Machine Learning
|
| Presentation Title |
Resolving hidden subgroups in heterogeneous alloy datasets |
| Author(s) |
Jin-Young Lee, Albert Linda, Anurag Bajpai, Dierk Raabe |
| On-Site Speaker (Planned) |
Jin-Young Lee |
| Abstract Scope |
Aggregating heterogeneous materials datasets from diverse literature often introduces hidden biases and misleading trends, such as Simpson’s paradox, which can hinder reliable AI-driven alloy design. To overcome this data-centric challenge, this study presents a causal effect-based clustering framework designed to group material data sharing similar sensitivities regarding how tensile properties change with variations in individual processing and compositional variables. The heterogeneous processing-property dataset was extracted using Large Language Models (LLMs) from scientific literature. The causal sensitivities were systematically quantified utilizing a causal forest double machine learning approach. Subsequently, the optimal number of clusters (i.e., subgroups) and variable grouping were rigorously determined through a genetic algorithm. For the identified subgroups, we comprehensively analyzed the physical and processing factors driving the separation of each distinct cluster, mapping them directly to their dominant processing routes and resulting microstructural domains. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Iron and Steel, Characterization |