Thermodynamic Networks Use Non-Equilibrium States for Universal Computation

A team of physicists from the University of Geneva, the University of Barcelona, and the University of Vienna have proposed a radical new computing paradigm: thermodynamic networks. Instead of silicon transistors, these networks use collections of finite-size reservoirs that exchange conserved quantities—like electric charge or molecule counts—as they relax to a non-equilibrium steady state. The steady state itself encodes the solution to a computational problem. The critical insight is that Negative Differential Conductance (NDC) is the physical property that determines expressivity. Without NDC, the network is limited to monotonic functions. With NDC, it can approximate any function universally, matching the expressivity of classical neural networks.

The team validated their framework on two very different physical platforms: quantum dot arrays and enzymatic reaction networks. Both can be engineered to exhibit NDC. In benchmarks, the networks performed sine function approximation and MNIST digit classification with high accuracy, proving that thermodynamic computation is not just theoretical. This work bridges quantum physics, statistical mechanics, and machine learning, offering a path toward computers that compute using the natural tendency of systems to equilibrate—potentially with far lower energy costs than traditional digital electronics.