Physics-supervised deep learning-based optimization (PSDLO) with accuracy and efficiency.

Area of Science:

• Computational Science
• Engineering Optimization
• Artificial Intelligence

Background:

• Optimization algorithms are crucial for scientific and engineering advancements.
• Current methods often combine evolutionary algorithms with deep learning, but face limitations.
• Achieving both accuracy and efficiency in optimization remains a significant challenge.

Purpose of the Study:

• To address the limitations of combined evolutionary and deep-learning optimization methods.
• To introduce and validate a novel physics-supervised deep-learning optimization algorithm (PSDLO).
• To demonstrate the capability of PSDLO in solving complex, multi-featured optimization problems.

Main Methods:

• Development of a physics-supervised deep-learning optimization algorithm (PSDLO).
• Supervision of deep learning model outputs using physics-based principles.
• Intervention within the evolutionary process to enhance optimization outcomes.
• Validation across three distinct physics-based case studies with varying data availability.

Main Results:

• PSDLO successfully achieved simultaneous accuracy and efficiency in optimization tasks.
• The algorithm demonstrated effectiveness even with insufficient datasets.
• The intrinsic limitations of purely evolutionary methods in conjunction with deep learning were highlighted.
• PSDLO proved capable of handling complex problems with numerous features.

Conclusions:

• PSDLO offers a superior approach to optimization compared to simple combinations of evolutionary and deep-learning methods.
• The physics-supervised deep learning framework provides a new perspective for tackling complex scientific and engineering optimization challenges.
• PSDLO shows significant potential for real-world engineering applications requiring accurate and efficient solutions.