Graph-Structured Hyperdimensional Computing for Data-Efficient and Explainable Process-Structure-Property Prediction

Researchers developed PSP-HDC, a graph-structured hyperdimensional computing framework for predicting material properties in 3D microstructure fabrication with sparse, heterogeneous data. The approach achieves 91% accuracy while providing inherent explainability—a critical advantage over conventional machine learning models that struggle with limited datasets and poor generalization.

PSP-HDC addresses a fundamental challenge in materials science: predicting process-structure-property relationships when training data is scarce and noisy. Traditional machine learning models fail in this regime because they require large, homogeneous datasets and tend to learn spurious correlations that don't transfer across manufacturing conditions. The proposed hyperdimensional computing framework encodes domain knowledge as a directed graph prior, constraining the model to learn physically meaningful relationships rather than statistical artifacts.

The technical innovation lies in combining three elements: a learnable encoder that maps heterogeneous parameters to hypervectors, graph-aligned composition operations that respect known PSP dependencies, and associative memory retrieval for prediction and attribution. This architecture inherently produces explanations—practitioners can understand which parameters drove decisions at multiple levels of granularity. The 91% accuracy on sheet-resistance prediction demonstrates practical viability, while the 89.6% generalization score under process-fold validation suggests robustness to manufacturing variations.

For the materials and manufacturing sector, this work bridges a critical gap. Early-stage process development relies on mechanistic understanding because data is limited; yet mechanistic models require expert calibration. PSP-HDC offers a middle path: it leverages domain structure without requiring fully calibrated submodels, making it deployable when conventional ML and physics-based approaches both struggle. The explainability component addresses regulatory and quality concerns in manufacturing, where decisions must be justified.

Looking forward, adoption hinges on validation across additional material systems and manufacturing platforms. If PSP-HDC generalizes beyond 3D photoreduction, it could accelerate innovation cycles in semiconductor processing, additive manufacturing, and materials discovery—domains where data scarcity chronically limits AI deployment.

• →Hyperdimensional computing framework achieves 91% accuracy on material property prediction with sparse, heterogeneous manufacturing data.
• →Graph-structured representation encodes domain knowledge to prevent spurious correlations and improve generalization across process conditions.
• →Intrinsic explainability at parameter and group levels addresses manufacturing and regulatory requirements for decision justification.
• →Approach viable during early process development when mechanistic submodels and large datasets are unavailable.
• →Validation on 3D photoreduction demonstrates proof-of-concept; broader manufacturing applications remain to be tested.