Many tabular datasets exhibit recurring statistical structure within specific application domains (omics, raman spectroscopy data). These shared characteristics suggest that knowledge about effective modeling may transfer across datasets. Here we propose the development of domain-aware ensemble portfolios: curated sets of modeling configurations learned from domain specific priors. Ensemble methods currently achieve state-of-the-art performance on tabular prediction tasks, and ensemble selection offers a principled mechanism for combining diverse models with limited additional computational cost. However, existing approaches often rely on large, domain-agnostic configuration spaces. This project examines whether domain-informed portfolios can achieve better compute-performance trade-off than domain agnostic solutions.