Analog Thermal Computing Structures that Count on Heat

Related Vendor

Wuxi InfiMotion Technology Co., Ltd.

Engineers at MIT have developed structures that can calculate with heat. By using excess heat instead of electricity, microscopically small silicon structures could enable more energy-efficient heat sensors and signal processing.

Researchers at MIT have developed silicon structures that can perform calculations in an electronic device using excess heat instead of electricity. These tiny structures could one day enable more energy-efficient calculations.

In this calculation method, the input data is encoded as a series of temperatures using the waste heat already present in an appliance. The flow and distribution of heat through a specially designed material forms the basis for the calculation. The output is then represented by the energy collected at the other end, which is thermostatted to a fixed temperature.

The researchers used these structures to perform matrix-vector multiplications with an accuracy of more than 99 percent. Matrix multiplication is the basic mathematical technique that machine learning models such as LLMs use to process information and make predictions.

While researchers still have many challenges to overcome to scale this computational method for modern deep learning models, the technique could be used to detect heat sources and measure temperature changes in electronic devices without consuming additional energy. This would also eliminate the need for multiple temperature sensors taking up space on a chip."When you perform calculations in an electronic device, heat is usually generated as a waste product. You often want to get rid of as much heat as possible. Here, however, we have taken the opposite approach by using heat as a form of information itself and have shown that computing with heat is possible," says Caio Silva, a student in the Department of Physics and lead author of a publication on the new computing paradigm.

Silva is supported in the article by Giuseppe Romano, a senior author and scientist at the MIT Institute for Soldier Nanotechnologies. The research results are published today in Physical Review Applied.

Turn Up the Heat

This work was made possible by a software system previously developed by the researchers, which allows them to automatically design a material that conducts heat in a certain way.

Using a technique called "inverse design", this system turns the traditional approach to engineering on its head. Researchers first define the desired functionality, then the system uses powerful algorithms to iteratively design the best geometry for the task.

Using this system, they designed complex silicon structures, each about the size of a grain of dust, that can perform calculations using thermal conduction. This is a form of analog computing in which data is encoded and signals are processed with continuous values instead of digital bits that are either 0 or 1.

The researchers enter the specifications of a number matrix, which represents a specific calculation, into their software system. Using a grid, the system designs a series of rectangular silicon structures filled with tiny pores. The system continuously adjusts each pixel in the grid until it achieves the desired mathematical function.

The heat diffuses through the silicon in a way that performs matrix multiplication, with the geometry of the structure encoding the coefficients. "These structures are far too complicated for us to develop using our intuition alone. We have to teach a computer to design them for us. This makes inverse design a very powerful technique," says Romano.

However, the researchers encountered a problem. Due to the laws of heat conduction, according to which heat flows from warm to cold areas, these structures can only encode positive coefficients. They solved this problem by splitting the target matrix into its positive and negative components and representing them with separately optimized silicon structures that encode positive entries. By subtracting the results in a later phase, they can calculate negative matrix values.

They can also adjust the thickness of the structures, allowing them to realize a greater variety of matrices. Thicker structures have a higher thermal conductivity. "Finding the right topology for a particular matrix is a challenge. We solved this problem by developing an optimization algorithm that ensures that the developed topology is as close as possible to the desired matrix without containing strange parts," explains Silva.

Microelectronic Applications

The researchers tested the structures using simulations on simple matrices with two or three columns. Although these small matrices are simple, they are relevant for important applications such as fusion sensors and diagnostics in microelectronics. In many cases, the structures performed calculations with an accuracy of over 99 percent.

However, there is still a long way to go before this technology can be used for large-scale applications such as deep learning, as millions of structures would have to be linked together. The more complex the matrices become, the less accurate the structures become, especially if the distance between the input and output connections is large. In addition, the devices have a limited bandwidth, which would have to be significantly expanded if they were to be used for deep learning.

However, as the structures are based on excess heat, they could be used directly for tasks such as thermal management and the detection of heat sources or temperature gradients in microelectronics. "This information is crucial. Temperature gradients can lead to thermal expansion and damage a circuit or even cause an entire device to fail. If we have a localized heat source where we don't want a heat source, that means we have a problem. With these structures, we could detect such heat sources directly and simply connect them without the need for digital components," says Romano.

Building on this proof of concept, the researchers want to develop structures that can perform sequential operations, with the output of one structure becoming the input for the next. In this way, machine learning models perform calculations. They also plan to develop programmable structures that allow them to encode different matrices without having to start from scratch with a new structure each time.(sg)

Additive Manufacturing in production processes

New Methods, Breakthroughs and Applications