TY - JOUR
T1 - Transfer Learning Across Materials Classes: Conceptual Boundaries of Reusability
AU - Lucas Meyer
AU - Stefan Braun
AU - Anna Schmid
JF - Journal of Computational and Data-Driven Materials Engineering
JO - J. Comput. Data-Driven Mater. Eng.
SN - 3149-9368
Y1 - 2024
VL - 3
IS - 2
DO - 10.68159/r828506645
SP - 118
N2 - In the evolving landscape of computational and data-driven materials engineering, transfer learning has emerged as a pivotal strategy to address data scarcity and enhance predictive capabilities across diverse materials systems. This approach leverages pre-trained models from one materials class to inform modeling in another, capitalizing on shared representational structures within high-dimensional chemical and physical spaces. However, the conceptual boundaries of reusability remain underexplored, particularly in terms of how representational invariances and domain shifts influence cross-class applicability. This manuscript introduces a novel conceptual framework, termed the Reusability Boundary Architecture (RBA), which delineates the systemic interactions between data representations, model architectures, and discovery workflows in transfer learning paradigms. By integrating insights from materials informatics, graph neural networks, and uncertainty quantification, the RBA elucidates the epistemic trade-offs inherent in transferring knowledge across materials classes, such as from inorganic crystals to organic polymers or metallic alloys to ceramics. The framework emphasizes computational steering logics that dynamically adjust for feature misalignment and contextual divergences, fostering more robust integration of simulation and experimental pipelines. Implications for the field include enhanced design of multimodal datasets, refined autonomous discovery systems, and improved inverse materials engineering, ultimately accelerating innovation in sustainable materials development without relying on empirical validations. This work provides a theoretical foundation for navigating the reusability frontiers in computational materials science, promoting interdisciplinary synergies between machine learning and domain-specific knowledge.
UR - https://iamrp.net/r828506645
ER -