ZENN: A thermodynamics-inspired computational framework for heterogeneous data-driven modeling.

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Traditional entropy methods struggle with integrating heterogeneous datasets.
• Growing data complexity necessitates advanced AI algorithms.
• Existing models face challenges in capturing intrinsic data disparities.

Purpose of the Study:

• To introduce a zentropy-enhanced neural network (ZENN) for effective learning from heterogeneous data.
• To extend zentropy theory into data science for improved AI.
• To enhance generalization and robustness in machine learning tasks.

Main Methods:

• Developed a zentropy-enhanced neural network (ZENN) architecture.
• Integrated intrinsic entropy and energy learning components.
• Introduced a learnable temperature variable to model data heterogeneity.

Main Results:

• ZENN demonstrated superior generalization and robustness on image and text classification tasks (CIFAR-10/100, BBC News, AG News).
• Successfully reconstructed the Helmholtz energy landscape of Fe3Pt, capturing material behaviors like negative thermal expansion.
• Outperformed state-of-the-art models in classification and energy landscape reconstruction.

Conclusions:

• ZENN provides a versatile and robust framework for data-driven machine learning.
• The zentropy-grounded approach is effective for scientific problems with complex, heterogeneous datasets.
• ZENN advances AI capabilities in handling data uncertainty and physical disorder.