Optimizing thermodynamic trajectories using evolutionary and gradient-based reinforcement learning.

Area of Science:

• Thermodynamics
• Machine Learning
• Artificial Intelligence

Background:

• Optimizing thermodynamic cycles is crucial for energy efficiency.
• Traditional methods may not explore all possible efficient trajectories.
• Reinforcement learning (RL) offers a novel approach to complex optimization problems.

Purpose of the Study:

• To investigate the use of neural-network-based reinforcement learning for identifying maximally efficient thermodynamic trajectories.
• To compare gradient and gradient-free RL methods in this context.
• To explore RL's capability in discovering new thermodynamic cycles.

Main Methods:

• Utilized a model heat engine for simulations.
• Employed an evolutionary learning algorithm to evolve neural networks.
• Trained networks to maximize efficiency over thermodynamic trajectories.
• Incorporated both gradient and gradient-free reinforcement learning techniques.

Main Results:

• Neural networks successfully learned to perform efficient thermodynamic cycles like Carnot, Stirling, and Otto.
• An evolutionary approach achieved the optimal Carnot cycle.
• Gradient-based RL identified the Stirling cycle.
• A novel thermodynamic cycle was discovered when an irreversible process was included.

Conclusions:

• Neural-network-based reinforcement learning is a viable tool for identifying optimal thermodynamic processes.
• RL strategies from game playing can be effectively applied to solve complex physical problems.
• This method has the potential to discover new, highly efficient thermodynamic cycles.