Researchers at Massachusetts Institute of Technology have demonstrated a surprising new way to compute—by using heat instead of electricity.
In a proof-of-concept study published in Physical Review Applied, the team showed that microscopic silicon structures can carry out mathematical calculations with more than 99 percent accuracy by guiding the flow of heat already present inside electronic devices.
In most electronics, heat is an unwanted byproduct.
Chips warm up as they work, and engineers spend enormous effort trying to remove that excess heat to prevent damage and wasted energy. The MIT researchers flipped that idea on its head.
Rather than fighting heat, they asked: what if heat itself could carry information and do useful work?
Their answer is a new form of “thermal computing.” In this approach, input data are encoded as temperatures.
Instead of sending electrical signals through wires, the system lets heat spread through a carefully designed piece of silicon.
The way that heat flows—how fast it moves and where it concentrates—naturally performs a calculation. At the end of the structure, the result is read out as a change in thermal power at a fixed-temperature boundary.
To test the idea, the team focused on matrix-vector multiplication, a core mathematical operation used by machine-learning systems, including large language models.
Almost everything these models do, from recognizing images to generating text, relies on repeated matrix calculations. Using simulations, the researchers showed that their heat-based structures could perform these operations with impressive accuracy.
Each computing element is a tiny piece of silicon, roughly the size of a dust particle, filled with carefully placed pores.
The shapes are far too complex to design by hand. Instead, the researchers used a technique called inverse design.
Rather than starting with a shape and seeing what it does, they start with the mathematical function they want and let a computer algorithm discover the best structure to make it happen.
The algorithm works by adjusting a grid of pixels that represent the silicon’s geometry.
Over many iterations, it tweaks the pattern until the way heat diffuses through the structure matches the desired calculation. The result is a physical object whose shape “stores” the math it performs.
One challenge comes from the basic laws of heat flow. Heat naturally moves from hotter areas to cooler ones, which makes it difficult to represent negative numbers.
The researchers solved this by splitting calculations into positive and negative parts, using separate structures for each and combining the results afterward. They also adjusted the thickness of the silicon, since thicker regions conduct heat more strongly, allowing for a wider range of calculations.
For now, the team tested only small matrices with two or three columns. While simple, these calculations are still useful for tasks such as monitoring heat flow in microelectronics or detecting unwanted hot spots on chips. In many cases, the structures achieved better than 99 percent accuracy.
Scaling this idea up to full-scale artificial intelligence is still a major challenge. Large neural networks would require millions of these structures working together, and accuracy drops as calculations become more complex or heat must travel longer distances. The system’s speed, or bandwidth, would also need significant improvement.
Even so, the researchers see immediate potential. Because the method uses waste heat, it could provide energy-free ways to sense temperature changes, detect dangerous hot spots, or monitor thermal stress inside electronics—without adding extra sensors or power consumption.
Looking ahead, the team aims to build chains of thermal computing elements so that the output of one feeds into the next, mimicking how machine-learning models operate.
They also hope to design programmable structures that can be reconfigured for different calculations, bringing the idea of computing with heat one step closer to practical use.
