Rethinking AI from the Hardware Up

Mark Hersam aims to make computing more energy efficient with brain-inspired electronics.

Artificial intelligence is advancing at extraordinary speed, and so is its appetite for energy. In 2013, U.S. data centers consumed an estimated 4.4 percent of the nation’s electricity, and Berkeley Lab projects that that share could reach approximately 12 percent by 2030, driven substantially by the rapid growth of AI computing. At Northwestern University, Mark Hersam and his team are working to change that trajectory. They are developing computing hardware with the potential to reduce the energy needs of AI data centers by orders of magnitude. The ultimate goal is to realize the benefits of AI while making the technology more sustainable.

Hersam’s approach involves rethinking the fundamental design of computing hardware. A leading nanotechnology researcher, he develops new materials and electronic devices that process information more like the human brain. This field, known as neuromorphic computing, could make AI dramatically more energy efficient by integrating memory and computation, adapting hardware to specific tasks, and avoiding unnecessary calculations.

“As a nanotechnologist for the last 25 years, I strongly believe that, by understanding matter at the shortest of length scales, we can find solutions to the world's largest problems,” he said in a 2025 TEDx talk. “What we're asking in our lab is, can we use nanotechnology to produce an alternative hardware platform for AI that will be much more energy efficient than a conventional digital computer?”

Hersam is the Walter P. Murphy Professor and chair of materials science and engineering at Northwestern’s McCormick School of Engineering and the director of Northwestern’s Materials Research Science and Engineering Center. His group’s work spans low-dimensional nanomaterials, nanoelectronics, energy storage, quantum information science, biosensors, and scalable manufacturing.

Across this broad portfolio, a consistent theme is the use of new materials to achieve functions that conventional silicon electronics cannot easily provide.

Moving memory and computation together

One major source of computing’s energy consumption is the constant movement of data. In conventional computers, processors perform calculations while memory components store information. Because the two functions occur in physically separate locations, computers shuttle data back and forth continuously. For data-intensive AI, that movement can consume substantial energy and slow down processing.

The brain operates differently. Its neurons and synapses store and process information within the same interconnected system. Hersam’s group is creating electronic devices that emulate this arrangement, performing memory and computation together.

A key example is the memtransistor, a device that combines the functions of a transistor and a memory element. By using atomically thin and nanoscale materials, the Hersam Research Group builds memtransistors that can adjust their electrical conductivity in response to past electrical signals, much like biological neurons can change through experience. This ability allows a small number of highly adaptable devices to do work that otherwise would require many conventional transistors.

In a 2023 study published in Nature Electronics, Hersam and his collaborators also created reconfigurable Gaussian transistors using a single-atom-thick layer of molybdenum disulfide (MoS₂) and a network of semiconducting single-walled carbon nanotubes. These devices can be dynamically adjusted to efficiently carry out machine-learning classification tasks.

Just two of these Gaussian transistors performed work that would typically require dozens of conventional transistors, dramatically reducing energy use. In a demonstration using electrocardiogram data, the system identified six types of irregular heartbeats with nearly 95 percent accuracy. Because the device processed the information locally, it also avoided sending sensitive health data to cloud servers.

This research illustrates the broader promise of Hersam’s approach. Rather than merely shrinking conventional computer components, his team designs new materials and devices that accomplish more with less hardware—and therefore less energy.

Learning from an overlooked part of the brain

Hersam’s group has continued to push neuromorphic hardware beyond conventional AI classification. In 2026, his team drew inspiration from the cerebellum, a brain region that helps the body react quickly to unexpected events. Rather than giving equal attention to every incoming signal, the cerebellum learns what is expected and reserves its resources for detecting change.

Conventional AI systems, by contrast, continuously analyze streams of incoming data, even when little or nothing has changed. This makes always-on systems such as health monitors, autonomous vehicles, robots, and cybersecurity tools unnecessarily energy intensive.

In a Nature Communications study, Hersam and collaborators developed a cerebellum-inspired memtransistor that reproduces complementary excitatory and inhibitory responses found in biological neural circuits. Working together, those responses allow the device to avoid unnecessary computation in response to expected information while rapidly flagging something new.

In a proof-of-concept demonstration, the system detected abnormal heart rhythms with more than 98 percent accuracy and within one-fifth of a heartbeat. It required approximately 10,000 times fewer computational operations than conventional AI approaches.

The advance demonstrates that energy-efficient computing involves more than making each calculation consume less power. Systems also can save energy by deciding which calculations are actually necessary.

Building more realistic artificial neurons

Another branch of Hersam’s research aims to reproduce the rich, varied signals generated by biological neurons. Many artificial neurons produce only simple electrical pulses. Engineers therefore need large networks of devices to emulate the complex behavior of the brain—adding hardware and increasing energy consumption.

Hersam’s team instead developed printable artificial neurons capable of producing multiple biologically realistic firing patterns. The devices are made from electronic inks containing nanoscale flakes of molybdenum disulfide, which functions as a semiconductor, and graphene, which conducts electricity. Using aerosol jet printing, the researchers deposited the materials onto flexible polymer surfaces.
The resulting devices can generate single spikes, continuous firing, and spiking bursts. In experiments, their signals were sufficiently biologically realistic to activate living neurons in slices of mouse brain tissue.

Reported in Nature Nanotechnology, the work could contribute to future neuroprosthetics and brain-machine interfaces. It also offers a possible foundation for computing systems that achieve complex behavior with fewer components and less energy.

A materials strategy for sustainable AI

Making AI more sustainable will require cleaner energy and more efficient data-center infrastructure, but it also will require fundamentally more efficient computation. Hersam’s work shows how materials science can contribute at that foundational level, enabling greater computing power with less energy consumption.

By developing new nanomaterials, combining memory and processing, and reproducing specialized functions of the brain, Hersam’s group is establishing alternative hardware architectures that could ultimately reduce dependence on centralized, energy-intensive computing.
His leadership also reflects an unusually interdisciplinary model. The research connects materials science, electrical engineering, computer science, chemistry, neuroscience, and medicine. It moves between fundamental questions—such as how nanoscale materials transport charge—and practical goals, including wearable health technologies, autonomous systems, neuroprosthetics, and low-power AI.

As artificial intelligence becomes embedded in more aspects of daily life, Hersam is helping establish a new principle for its development: greater intelligence should not automatically require greater energy consumption. By learning from the brain and inventing hardware capable of doing more with fewer devices and fewer operations, his group is opening a path toward AI systems that are not only more capable, but also more sustainable.

Selected Publications

Hadke, S. S., Klingler, C. N., Brown, S. T., Holla, M., Zhuang, X., Li, L., Utama, M. I. B., Diaz-Arauzo, S., Chapagain, A., Li, S., Lee, J. H., Raman, I. M., Sangwan, V. K., & Hersam, M. C. (2026). Printed MoS₂ memristive nanosheet networks for spiking neurons with multi-order complexity. Nature Nanotechnology, 21, 672–679. https://doi.org/10.1038/s41565-026-02149-6

Kang, M.-A., Brown, S. T., Jayasinghe, N., Holla, M. R., Pham, T. T., Zeng, T. T., Wu, R., Trdinich, Z. J., Zhuang, X., Dravid, V. P., Raman, I. M., Trivedi, A. R., Sangwan, V. K., & Hersam, M. C. (2026). Cerebellum-inspired memtransistors enable emergent differentiation for hardware-efficient novelty detection. Nature Communications. Advance online publication. https://doi.org/10.1038/s41467-026-75212-4

Yan, X., Qian, J. H., Ma, J., Zhang, A., Liu, S. E., Bland, M. P., Liu, K. J., Wang, X., Sangwan, V. K., Wang, H., & Hersam, M. C. (2023). Reconfigurable mixed-kernel heterojunction transistors for personalized support vector machine classification. Nature Electronics, 6, 862–869. https://doi.org/10.1038/s41928-023-01042-7

Hadke, S., Kang, M.-A., Sangwan, V. K., & Hersam, M. C. (2025). Two-dimensional materials for brain-inspired computing hardware. Chemical Reviews, 125(2), 835–932. https://doi.org/10.1021/acs.chemrev.4c00631

Yan, X., Qian, J. H., Sangwan, V. K., & Hersam, M. C. (2022). Progress and challenges for memtransistors in neuromorphic circuits and systems. Advanced Materials, 34(48), 2108025. https://doi.org/10.1002/adma.202108025