When Does Multi-Task Learning Fail? Quantifying Data Imbalance and Task Independence in Metal Alloy Property Prediction

Abstract

Multi-task learning (MTL) is widely adopted in materials informatics under the assumption that related properties share leverageable physical principles. This study critically examines this premise by simultaneously predicting electrical resistivity, Vickers hardness, and amorphous-forming ability using a dataset of 54,028 metal alloys.1 Contrary to expectations, we observe a striking dichotomy: MTL significantly degrades regression accuracy (e.g., hardness 2$R^2$ drops from 3$0.832$ to 4$0.694$) while improving classification performance (amorphous F1 increases from 5$0.703$ to 6$0.744$).7 Analysis of learned task graphs reveals negligible inter-task correlations, attributing regression failure to negative transfer driven by severe data imbalance (52,388 vs. 800 samples). To mitigate this, we evaluate Deep Imbalanced Regression techniques. PCGrad recovers hardness performance ($R^2 \rightarrow 0.855$) by resolving gradient conflicts, while LDS+GradNorm achieves the best overall multi-task balance. Our findings suggest that alloy properties often behave independently, necessitating specific strategies: independent models for maximum regression precision, PCGrad for minority tasks, and LDS+GradNorm when balanced joint prediction is required.