Computers fight heat; a team at MIT has instead put it to work. Led by undergraduate physics student Caio Silva with senior author Giuseppe Romano of the Institute for Soldier Nanotechnologies, the researchers designed microscopic silicon structures in which flowing heat itself performs calculations, reporting the work in Physical Review Applied. Input data are encoded as temperatures across the device, and the answer is read out as the power collected at the far end.
Each structure is roughly the size of a dust particle, and its porous geometry is produced by inverse design: the team specifies the desired mathematical function first, then lets an algorithm iteratively sculpt the silicon until heat flow through it implements that function. Because heat moves only from hot to cold, negative numbers pose a challenge, which the researchers solved by splitting each matrix into positive and negative parts. In simulations on small matrices of two or three columns, the structures achieved over 99 percent accuracy in many cases.
The near-term promise is energy-free sensing inside microelectronics, where the devices could detect unwanted heat sources and temperature gradients without extra power or bulky instrumentation. The authors are candid about the limits: accuracy declines for more complex matrices and bandwidth is constrained, so thermal deep learning remains distant. As a demonstration that waste heat can carry computation at all, however, the result opens an unusual design space.