Thermodynamic Networks: Harnessing Non-Equilibrium Steady States for Computation
Researchers have developed a new framework called thermodynamic networks that uses physics-based computation through non-equilibrium steady states. These networks leverage the exchange of conserved quantities between reservoirs to encode computational problem solutions. The presence of Negative Differential Conductance (NDC) is identified as crucial for computational expressivity, enabling universal function approximation, while its absence limits computations to monotonic functions. The approach has been demonstrated on quantum dot and enzymatic reaction networks, achieving high performance on benchmarks like sine function approximation and MNIST digit classification. AI
IMPACT Introduces a novel physics-based computation method that could lead to new AI hardware architectures.