An AI just invented dozens of 'impossible' heat engines that beat 100 years of human engineering.
April 17, 2026
Original Paper
Automated co-design of high-performance thermodynamic cycles via graph-based hierarchical reinforcement learning
arXiv · 2604.13133
The Takeaway
For over a century, we have relied on the same basic designs for refrigerators and power plants. This AI was given the rules of thermodynamics and discovered 39 brand-new cycles that are more efficient than anything humans ever built. It found complex configurations that human engineers simply never considered because they were too non-intuitive. This means we could suddenly see a huge jump in the efficiency of everything from home AC units to massive power grids. It proves that even in 'solved' fields like thermodynamics, AI can find better ways to power our world.
From the abstract
Thermodynamic cycles are pivotal in determining the efficacy of energy conversion systems. Traditional design methodologies, which rely on expert knowledge or exhaustive enumeration, are inefficient and lack scalability, thereby constraining the discovery of high-performance cycles. In this study, we introduce a graph-based hierarchical reinforcement learning approach for the co-design of structure parameters in thermodynamic cycles. These cycles are encoded as graphs, with components and connec