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

TL;DR

The paper addresses when multi-task learning is beneficial for metal alloy property prediction. Using a union dataset of 54,028 samples across resistivity, hardness, and amorphous-forming ability, it compares independent models, standard MTL, and structured MTL with learned task graphs. It finds a striking division: MTL degrades regression performance due to severe data imbalance and near-zero inter-task transfer, yet improves amorphous-forming classification through regularization and favorable optimization dynamics. Learned task relations reveal actual task independence, guiding practical recommendations: use independent models for precise property prediction, but leverage MTL for screening tasks where recall matters; always validate MTL assumptions and account for data imbalance. The work thus provides a nuanced roadmap for applying MTL in materials discovery pipelines and highlights avenues for future improvement, including gradient balancing, curriculum learning, and soft parameter sharing.

Abstract

Multi-task learning (MTL) assumes related material properties share underlying physics that can be leveraged for better predictions. We test this by simultaneously predicting electrical resistivity, Vickers hardness, and amorphous-forming ability using 54,028 alloy samples. We compare single-task models against standard and structured MTL. Results reveal a striking dichotomy: MTL significantly degrades regression performance (resistivity $R^2$: 0.897 $\to$ 0.844; hardness $R^2$: 0.832 $\to$ 0.694, $p < 0.01$) but improves classification (amorphous F1: 0.703 $\to$ 0.744, $p < 0.05$; recall +17%). Analysis shows near-zero inter-task weights, indicating property independence. Regression failure is attributed to negative transfer caused by severe data imbalance (52k vs. 800 samples). We recommend independent models for precise regression, while reserving MTL for classification tasks where recall is critical.

Figures (7)

• Figure 1: Learned task relation graph. All inter-task weights are near zero, indicating task independence.
• Figure 2: Training curves showing validation loss over epochs. MTL models exhibit higher loss for regression tasks despite longer training.
• Figure 3: Error distributions for each property. Independent models show tighter distributions for regression tasks.
• Figure 4: Confusion matrices for amorphous classification. MTL improves recall (fewer false negatives) at slight cost to precision.
• Figure 5: Cost of data imbalance. Hardness $R^2$ decreases as the resistivity dataset grows, demonstrating that data imbalance directly causes negative transfer. Error bars show $\pm$1 std across 5 seeds.