Diversity-Aware Pretraining in Materials Learning via Task Similarity

Large-scale datasets underpin recent advances in machine learning; however, in materials science and chemistry, data acquisition remains expensive, making performance in low-data regimes a central challenge. Pretraining and transfer learning are effective strategies in this setting, yet their success critically depends on the choice of pretraining tasks. Poorly selected tasks can yield marginal gains or induce negative transfer, while principled criteria for assembling pretraining datasets remain underexplored. In this work, we leverage task similarity metrics to move beyond selecting a single source task and instead construct diverse, representative pretraining task subsets. Using similarity-derived structure among tasks, we show how pretraining datasets can be assembled to balance relevance and diversity, maximizing knowledge transfer under fixed data budgets. Experiments on the QM9 benchmark demonstrate that models pretrained on such diversity-aware task subsets achieve performance comparable to that of substantially larger pretraining datasets assembled without regard to task relationships. These results identify task diversity as a key factor governing transfer efficiency and provide a practical strategy for scaling general-purpose models in materials science under high labeling costs.