Computing with Heat Using Biocompatible Mott Neurons

Kyung Min Kim, Gwangmin Kim, Jae Hyun In, Younghyun Lee, Hakseung Rhee, Woojoon Park, Hanchan Song, Juseong Park, Jae Bum Jeon, Timothy D. Brown, A. Alec Talin, Suhas KumarResearch Square 2023 · DOI

Abstract

Heat dissipation is a natural consequence of operating any electronic system. In nearly all of computing, such heat is minimized by design and discarded via cooling, while some post-digital electronics (such as phase-change memories) utilize only the static electrical outcome of such heat dissipation within a single device to represent information. Thus, neither the naturally produced heat energy nor its dynamics are directly used for computing. Here we demonstrate electronic NbO<sub>x</sub> Mott neurons integrated on a biocompatible flexible organic substrate, designed to exploit both the thermal dynamics of the Mott transition and the dynamics of thermal interactions with the organic substrate, to exhibit 18 bio-mimetic neuronal behaviors and frequency-based nociception (sensing of noxious stimuli) within a single component. Further, multiple Mott neurons can spatiotemporally communicate purely via their thermal dynamics, which we use for graph optimization by consuming over 10<sup>7</sup>× lower energy compared to the best digital processors. Thus, we show that naturally produced heat in electronic systems and its dynamics can be used directly to encode and process information, thereby enabling a higher-complexity energy efficient and radically novel mixed-physics computing primitive.

Introduction: Beyond the Single-Physics Paradigm

Modern computing is fundamentally an exercise in heat management. As transistors switch, they dissipate energy as heat, which is traditionally viewed as a parasitic loss that must be mitigated through complex cooling systems. This "single-physics" approach—relying almost exclusively on electrical charge—stands in stark contrast to biological brains. Biological systems are "mixed-physics" processors, utilizing a sophisticated interplay of electrical, chemical, and thermal signals to achieve high-order dynamics with extreme energy efficiency.

The Biocompatible Mott Neuron

To bridge this gap, researchers have developed a novel electronic primitive: the NbOₓ (Niobium Oxide) Mott neuron integrated on a flexible, biocompatible polyimide (PI) substrate. The choice of substrate is critical. While traditional silicon substrates are highly thermally conductive and quickly whisk heat away, polyimide acts as a thermal insulator and "heat reservoir."

By exploiting the thermal dynamics of the Mott transition alongside the thermal properties of the organic substrate, the researchers created a system with "third-order" dynamics. While standard memristive neurons typically track two variables (voltage and device temperature), this system adds a third: the substrate temperature ($T_{sub}$). This additional degree of freedom allows a single, compact device to replicate complex neuronal behaviors that would otherwise require hundreds of transistors or bulky external capacitors.

Bio-mimetic Capabilities and Performance

The PI-based Mott neuron demonstrated remarkable versatility, successfully replicating all 18 biological neuronal behaviors defined by the Izhikevich model. These include fundamental patterns like tonic spiking and more complex "third-order" behaviors such as phasic bursting and subthreshold oscillations.

Figure 1: Device architecture, 18 bio-mimetic neuronal behaviors, and graph optimization performance.

Beyond standard spiking, the devices exhibit frequency-based nociception—the ability to sense and respond to "painful" or noxious stimuli. By utilizing residual heat to lower firing thresholds, the neurons mimic biological phenomena like hyperalgesia (increased pain sensitivity) and allodynia (pain response to non-painful stimuli), making them ideal candidates for advanced prosthetics or wearable health monitors.

Thermal Communication and Graph Optimization

Perhaps the most radical departure from traditional design is the use of "thermal crosstalk." In standard chip design, heat leaking from one component to another is a defect to be avoided. Here, it is a feature. Multiple Mott neurons can communicate spatiotemporally through the diffusion of heat across the polyimide substrate, allowing them to "talk" to neighbors without extra wiring.

This thermal communication was put to the test in solving the Max-Cut problem, a classic NP-hard graph optimization challenge. By mapping the graph onto a network of thermally coupled neurons, the system reached solutions with over $10^7\times$ lower energy consumption than state-of-the-art digital processors. This massive efficiency gain stems from the fact that the computation is performed by the "natural" physics of the system rather than by millions of switching logic gates.

Limitations and Future Outlook

While the results are groundbreaking, challenges remain. Thermal communication is currently effective over short distances, making it best suited for "sparse" networks. Scaling the system while maintaining high oscillation frequencies will require further miniaturization to the nanometer scale. Additionally, future iterations may incorporate electrochemical tuning to create more dynamic and programmable thermal interconnects.

By turning waste heat into a computational resource, this work paves the way for a new class of biocompatible, energy-efficient, and mixed-physics computing systems that more closely resemble the elegance of the human brain.

Summary updated

Abstract

<title>Abstract</title> Heat dissipation is a natural consequence of operating any electronic system. In nearly all of computing, such heat is minimized by design and discarded via cooling, while some post-digital electronics (such as phase-change memories) utilize only the static electrical outcome of such heat dissipation within a single device to represent information. Thus, neither the naturally produced heat energy nor its dynamics are directly used for computing. Here we demonstrate electronic NbO<sub>x</sub> Mott neurons integrated on a biocompatible flexible organic substrate, designed to exploit both the thermal dynamics of the Mott transition and the dynamics of thermal interactions with the organic substrate, to exhibit 18 bio-mimetic neuronal behaviors and frequency-based nociception (sensing of noxious stimuli) within a single component. Further, multiple Mott neurons can spatiotemporally communicate purely via their thermal dynamics, which we use for graph optimization by consuming over 10<sup>7</sup>× lower energy compared to the best digital processors. Thus, we show that naturally produced heat in electronic systems and its dynamics can be used directly to encode and process information, thereby enabling a higher-complexity energy efficient and radically novel mixed-physics computing primitive.
BibTeX
@inproceedings{kim2023computing,
  author = {Kyung Min Kim and Gwangmin Kim and Jae Hyun In and Younghyun Lee and Hakseung Rhee and Woojoon Park and Hanchan Song and Juseong Park and Jae Bum Jeon and Timothy D. Brown and A. Alec Talin and Suhas Kumar},
  title = {{Computing with Heat Using Biocompatible Mott Neurons}},
  booktitle = {Research Square},
  year = {2023},
  doi = {10.21203/rs.3.rs-3134569/v1},
}
Figure