Lawrence Berkeley National Lab Proposes Design and Training Framework for Thermodynamic Computing Mimicking Neural Networks

Researchers at Lawrence Berkeley National Laboratory have proposed a design and training framework for a thermodynamic computer that mimics a neural network. The work is described in a paper published in Nature Communications. It stems from a collaboration between the Molecular Foundry and the National Energy Research Scientific Computing Center, both Department of Energy user facilities at the lab. The framework supports nonlinear computations without the need to wait for thermodynamic equilibrium. Stephen Whitelam, a staff scientist at the Molecular Foundry, and Corneel Casert of NERSC developed the approach through digital simulations. When system components are nonlinear, the computer can perform calculations at specified times regardless of equilibrium status. Casert built a training framework that relies on genetic algorithms and massively parallel evolutionary simulations. The simulations ran on 96 GPUs of the Perlmutter supercomputer at NERSC. They evaluated billions of noisy dynamical trajectories per generation and totaled more than a trillion runs. Whitelam said a nonlinear thermodynamic circuit can behave like a neuron in a neural network. The method broadens thermodynamic computing to complex problems of the type handled by neural networks. The researchers said further work is needed to implement the designs in hardware and to create additional algorithms for nonequilibrium and nonlinear systems.