This paper has been withdrawn by Sungwoo Kang
[Submitted on 28 Dec 2025 (v1), last revised 23 Oct 2026 (this version, v4)]
No PDF available, click to view other formats
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.
Submission history
From: Sungwoo Kang [view email]
[v1] Sun, 28 Dec 2025 01:52:37 UTC (75 KB)
[v2] Tue, 6 Jan 2026 09:28:29 UTC (91 KB)
[v3] Sun, 1 Feb 2026 15:29:06 UTC (143 KB)
[v4] Fri, 23 Oct 2026 03:44:04 UTC (1 KB) (withdrawn)
