aiqbench analysis
This paper tests a common assumption in materials-science ML: that related properties like resistivity, hardness, and glass-forming ability can share representations in one multi-task model. On a 54,028-sample alloy dataset, multi-task learning hurt regression for resistivity and hardness but improved recall for amorphous-forming ability, and the authors tie this to gradient conflicts. For ML practitioners, the useful takeaways are that task grouping can cause negative transfer and that gradient-level fixes like PCGrad, or label smoothing with gradient normalization, can recover performance, though the proposed "materials property clustering" hypothesis still needs broader validation.
Multi-Task Learning for Metal Alloy Property Prediction: An Empirical Study of Negative Transfer and Mitigation Strategies
arXiv:2512.22740v4 Announce Type: replace 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.
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.
Machine Learning (cs.LG)
arXiv:2512.22740 [cs.LG]
(or arXiv:2512.22740v4 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2512.22740 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 238 words
Related coverage
Hierarchical Latent Prediction for Language Models
arXiv · Aug 7, 2026
Generative artificial intelligence for reliable mechanistic reasoning for corrosion
arXiv · Sep 2, 2026
Variational Inference for Evidential Deep Learning
arXiv · Jul 20, 2026
Amortized Predictability-aware Training Framework for Time Series Forecasting and Classification
arXiv · Jun 9, 2026