computer chip glowing green

Applications 

  • Training low-energy computational devices for machine learning applications

Benefits

  • Thermodynamic computers use much less energy than classical or digital computers
  • The method allows for training thermodynamic computers to perform machine-learning operations on a clock (i.e., at a pre-defined time)
  • Applicable to diverse thermodynamic hardware

Background

The escalating energy consumption of conventional digital and quantum computers, particularly for demanding tasks like machine learning, presents a critical challenge. Thermodynamic computing offers a promising alternative by harnessing thermal noise as a computational resource rather than suppressing it. However, existing approaches have faced significant limitations. Most existing algorithms are designed to work in equilibrium. This requires devices to attain thermal equilibrium, which can be a slow process. It is more convenient to do calculations out of equilibrium, on a clock, but existing methods for training thermodynamic computers out of equilibrium rely on computationally expensive genetic algorithms, restricting the development of thermodynamic algorithms for modern machine learning tasks.

Technology Overview

Scientists at Berkeley Lab have developed a gradient-descent method for training thermodynamic computers to perform neural-network-like computations by harnessing thermal noise. Key features of this technology include:

  • Efficient Out-of-Equilibrium Operation: Training occurs at finite observation times, eliminating the need for slow thermal equilibration.
  • Teacher–Student Framework: A digital neural network generates a guide for a “student” thermodynamic computer, whose parameters are updated using functional analytic gradients.
  • High Accuracy with Low Energy Cost: Demonstrated on image recognition simulations, the method achieved near-neural-network performance (98% accuracy for 3-class and 92% for 10-class classification) while requiring up to eight orders of magnitude less energy than comparable digital implementations.
  • Hardware-Ready: Once trained in simulation, parameters can be exported to physical thermodynamic computing hardware for practical deployment.

Development Stage

Basic principles validated in simulation (TRL 1)

Principal Investigators

Stephen Whitelam

IP Status

Patent pending

Opportunities

Available for licensing or collaborative research

For More Information

Whitelam, S. (2025). Training thermodynamic computers by gradient descent. arXiv preprint arXiv:2509.15324. https://doi.org/10.48550/arXiv.2509.15324