MIT engineers convert waste heat into computing power with a new silicon structure.

Machine Learning


Artistic rendering showing a thermal analog computing device

ⓘ Massachusetts Institute of Technology

Artistic rendering showing a thermal analog computing device

A team of researchers at MIT has created a microscopy device that performs machine learning operations using only the heat already present in the electronics, achieving more than 99% accuracy.

MIT engineers have turned waste heat, a common nuisance in electronics, into a computational resource. In a study published in the journal Physical Review Applied, researchers have uncovered microscopic silicon structures that can perform mathematical calculations using heat instead of electricity.

The research team, consisting of undergraduate student Caio Silva and research scientist Giuseppe Romano, created these structures using a technique called inverse design. By inputting the required features into a software system, the algorithm generated a complex silicon shape filled with pores, about the size of a dust particle. These structures direct heat flow and perform matrix-vector multiplication, the basic mathematics behind machine learning models such as large-scale language models (LLMs), with over 99% accuracy in simulations.

In most cases, performing calculations on electronic devices wastes heat. We often want to remove as much heat as possible. But here we took the opposite approach, using heat as a form of information itself to show that thermal computing is possible. — Caio Silvafirst author of the paper.

To overcome the physical limitation that heat only flows from hot to cold, the team split the target matrix into positive and negative components and processed them through separate structures. We also adjusted the thickness of the silicone to more precisely control heat transfer.

Although this technology faces hurdles in bandwidth and scaling for complex deep learning tasks, it has immediate potential in thermal management. This structure can autonomously detect overheating and temperature gradients in electronic devices without the need for external power sources or digital sensors. The team now aims to develop programmable structures capable of continuous operation.



Source link