In traditional computing, we solve optimization problems by simulating a search. We write algorithms like gradient descent that take discrete steps across a mathematical landscape, consuming clock cycles and energy to find a minimum.
A growing body of research suggests a more direct path: instead of simulating the math, we can build physical systems whose natural laws are the optimization. By mapping a cost function to a physical quantity like power dissipation or magnetic energy, we can let the system "relax" into a solution. In this paradigm, the answer to a complex problem is simply the state the hardware reaches when it stops moving.
The Physics of the Minimum
The intuition behind this approach dates back to early neuromorphic engineering. Christof Koch and Carver Mead demonstrated in Computing Motion Using Resistive Networks that analog resistive networks governed by Kirchhoff’s Laws naturally minimize power dissipation. If you map the constraints of a computer vision problem (like optical flow) onto the resistors and voltages of a circuit, the steady-state voltage distribution becomes the mathematical solution.
This isn't just a clever trick; it's a fundamental shift in how we view "computation." In a digital system, every step of an optimization requires active switching of transistors. In a physical relaxation system, the hardware is a "physical optimizer" John G. Harris that settles into a state of minimum energy An Analog VLSI Chip for Thin-Plate Surface Interpolation. The "math" happens at the speed of light or the charging of a capacitor, often reaching solutions in near-constant time regardless of problem size Neural Network Routing for Random Multistage Interconnection Networks.
Ising Machines and Combinatorial Search
One of the most successful applications of this idea is the Ising Machine. Many of the world's hardest problems—from protein folding to 5G signal decoding—can be mapped to the Ising model, a mathematical description of interacting magnetic spins that want to align in a way that minimizes total energy.
Researchers like Antik Mallick have shown that we can build these machines using standard CMOS technology. By using bistable latches (circuits with two stable states) to represent "spins" and programmable resistors to represent their interactions, the hardware becomes a dynamical system that physically "anneals" CMOS-compatible ising machines built using bistable latches coupled through ferroelectric transistor arrays.
Because these systems exploit intrinsic noise and electrical equilibrium rather than iterative logic, they can achieve staggering performance gains. For example, the NP-GL framework developed by Chunshu Wu and Ang Li achieves a 7,000x speedup and 100,000x energy reduction over GPUs by allowing a circuit to reach electrical equilibrium to solve graph learning problems Extending Power of Nature from Binary to Real-Valued Graph Learning in Real World.
Learning as Equilibrium
If inference is a relaxation to equilibrium, can learning be framed the same way? Equilibrium Propagation (EqProp), introduced by Benjamin Scellier and Yoshua Bengio, suggests the answer is yes.
In EqProp, a neural network is treated as a physical system with an energy function. To train it, you "nudge" the output layer toward a target state. This nudge creates a physical perturbation that propagates through the system, naturally calculating the gradients needed for learning without the need for a separate "backpropagation" pass Equilibrium Propagation: Bridging the Gap Between Energy-Based Models and Backpropagation.
This approach allows for "agnostic" learning: a physical system can optimize its own parameters even if we don't know the exact equations governing its internal physics Agnostic Physics-Driven Deep Learning. The parameters simply co-evolve with the system states toward a global energy minimum.
The Thermodynamic Advantage
The move toward "Thermodynamic Computing" Thomas M. Conte represents a final break from the von Neumann architecture. In this view, the "power wall" of modern chips isn't a bug, but a feature we aren't using. By harnessing thermodynamic fluctuations and natural annealing, we can create systems that self-organize to solve problems Thermodynamic Computing.
Recent work on InstaTrain and DS-LLM by Chuan Liu and Ruibing Song applies this to Large Language Models. By mapping LLM operations to the physical steady-state of electronic circuits, they bypass discrete digital logic entirely. The result is training that is thousands of times faster because it happens at the speed of the hardware's physical response DS-LLM: Leveraging Dynamical Systems to Enhance Both Training and Inference of Large Language Models.
Ultimately, these researchers are arguing that the most efficient way to solve an optimization problem is to stop calculating and start evolving. When the mathematical loss function is equivalent to physical heat dissipation Learning with springs and sticks, the universe itself becomes the computer.