Multi-Task Learning for Metal Alloy Property Prediction: An Empirical Study of Negative Transfer and Mitigation Strategies

Abstract

Multi-task learning (MTL) in materials science relies on the assumption that physically related properties share learnable representations. We challenge this assumption using a 54,028-sample metal alloy dataset exhibiting extreme task-level imbalance. Our results reveal a striking dichotomy: MTL significantly degrades regression performance for resistivity and hardness but improves classification recall for amorphous-forming ability. We trace this divergence to mismatched functional forms--such as resistivity's polynomial dependence versus hardness's complex interactions--which cause severe gradient misalignment during optimization. Evaluating Deep Imbalanced Regression techniques, we find that projecting conflicting gradients (PCGrad) recovers single-task performance, while combining label distribution smoothing with gradient normalization achieves the best overall balance. Consequently, we propose a strategic framework: utilize independent models for high-precision characterization, but employ MTL for high-throughput screening where recall is paramount. These findings support a "materials property clustering" hypothesis, suggesting that distinct physical mechanisms require specialized optimization strategies to overcome negative transfer.

arXiv commentWithdrawn by the author due to critical issues identified in the experimental methodology and dataset structure: the reported negative transfer and optimization penalties were found to be artifacts of data leakage and lack of headroom over trivial lookup baselines, and the main findings failed to generalize to an external benchmark (JARVIS-DFT)