Autonomous design of noise-mitigating structures using deep reinforcement learning.

Area of Science:

• Computational engineering
• Acoustic metamaterials
• Machine learning

Background:

• Designing effective noise-mitigating structures often requires extensive simulation and prior knowledge.
• Conventional deep learning methods for structural design typically need labeled data, limiting their applicability.
• Autonomous design offers a potential solution for discovering novel acoustic solutions.

Purpose of the Study:

• To explore the application of deep reinforcement learning (DRL) for the autonomous design of noise-mitigating structures.
• To investigate the efficacy of deep Q-networks and double deep Q-networks in achieving broadband noise mitigation.
• To demonstrate the generalizability of DRL algorithms across different acoustic environments (reflection and transmission).

Main Methods:

• Utilizing deep Q-networks (DQN) and double deep Q-networks (DDQN) for material distribution optimization.
• Employing pixel-based inputs for DDQN to learn noise mitigation strategies without prior knowledge.
• Implementing unified hyperparameters and network architectures for both reflection and transmission problems.
• Comparing the performance of DRL algorithms against a genetic algorithm.

Main Results:

• DDQN successfully learned material distributions for broadband noise mitigation without requiring labeled data.
• The DRL approach demonstrated generalizability across different acoustic environments (transmission and reflection).
• Comparison with a genetic algorithm indicated potential for generalized design in complex scenarios, though DRL showed a tendency towards local maxima.

Conclusions:

• Autonomous design using DRL, particularly DDQN, offers a powerful, knowledge-free approach for creating noise-mitigating structures.
• The method shows promise for generalized learning in acoustics, adaptable to various shapes and environments.
• Further research can explore hyperparameter optimization and mitigation of local maxima prediction for enhanced performance.