Source-linked AI summary
Neural Dust: An Ultrasonic, Low Power Solution for Chronic Brain-Machine Interfaces
Dongjin Seo, Jose M. Carmena, Jan M. Rabaey, Elad Alon, Michel M. Maharbiz
TL;DR
The paper addresses the lack of a lifetime-viable, tetherless, high-density neural interface by analyzing low-power CMOS systems using ultrasonic power delivery and backscatter communication. It proposes free-floating neural dust nodes with a sub-cranial interrogator and reports substantially greater simulated ultrasonic power delivery than EM at the same scale, while identifying front-end, transducer-integration, and scaling constraints.
Problem
A tetherless, high-density, chronic brain interface remains a major challenge, and many energy modalities are unsuitable for neural-dust size scales.
Method
The paper analyzes system design trade-offs and size, power, and bandwidth scaling limits for low-power CMOS coupled with ultrasonic power delivery and backscatter communication.
Results
Approximately 500 µW can be received by a 100 µm node embedded 2 mm in brain tissue using a 1 mm^2 interrogator, over 10^7 more than the modeled EM result at the same scale.
Takeaways & Limitations
The proposed neural-dust architecture provides a path toward massive scaling of chronic brain recordings using independent miniature nodes and a sub-cranial interrogator.
Takeaways & Limitations
The analysis identifies front-end operation under decreasing power and SNR, miniature transducer–CMOS integration, and proper encapsulation as major realization challenges.
Abstract
from arXiv · showhide
A major hurdle in brain-machine interfaces (BMI) is the lack of an implantable neural interface system that remains viable for a lifetime. This paper explores the fundamental system design trade-offs and ultimate size, power, and bandwidth scaling limits of neural recording systems built from low-power CMOS circuitry coupled with ultrasonic power delivery and backscatter communication. In particular, we propose an ultra-miniature as well as extremely compliant system that enables massive scaling in the number of neural recordings from the brain while providing a path towards truly chronic BMI. These goals are achieved via two fundamental technology innovations: 1) thousands of 10 - 100 μm scale, free-floating, independent sensor nodes, or neural dust, that detect and report local extracellular electrophysiological data, and 2) a sub-cranial interrogator that establishes power and communication links with the neural dust.
System Concept
The paper proposes neural dust: miniature, free-floating sensor nodes powered and queried ultrasonically by a sub-cranial interrogator. This architecture targets chronic, high-density neural recording while addressing size, power, tissue attenuation, and communication constraints.
- Low-power CMOS with ultrasonic harvesting and backscatter communication provides a toolset for scalable, chronic extracellular recording systems.
- Thousands of 10–100 µm free-floating nodes detect local extracellular electrophysiology, while a sub-cranial interrogator provides power and spatially and frequency-selective communication.The interrogator is placed beneath the skull and dura mater to reduce ultrasound attenuation by bone.
- Implant size and power are primary constraints because implants larger than one or two cortical cell diameters can provoke tissue responses that degrade performance over months.
- EM scaling is further constrained by exponentially increasing tissue loss with frequency and dramatically declining mutual coupling as chiplet dimensions shrink.A modeled 100 µm neural dust has an approximately 10 GHz resonant frequency, while Fig. 2 reports 20 dB loss at 10 GHz through 2 mm of brain tissue.
- For 100 µm neural dust embedded 2 mm in brain tissue, EM transmission incurs 64 dB loss and delivers approximately 40 pW from a 1 mm^2, 100 µW transmitter.The paper concludes that EM power transmission is inappropriate for powering tens-of-micrometers-scale neural dust implants.
- Ultrasound offers short wavelength and low attenuation; at 10 MHz in brain tissue its wavelength is 150 µm, compared with 5 mm for 10 GHz EM.The shorter wavelength supports far-field operation and broader spatial coverage for interrogating multiple nodes.
System design and constraints: Power Delivery
The paper models ultrasonic power delivery with finite-element-informed KLM and cascaded two-port link models, while accounting for severe scaling constraints in miniature neural dust. Simulations indicate substantial power transfer at 100 µm, but efficiency becomes increasingly dependent on dust cross-sectional area below 100 µm.
- Interrogator geometry: With 2 mm transmission distance and 10 MHz ultrasound, the external interrogator’s maximum dimension should be approximately 1 mm to place the receiver near one Rayleigh distance.The preferred placement minimizes beam spreading and maximizes received power.
- Modeling constraints: The KLM model is limited for cube-like neural dust because it assumes predominantly one-dimensional thickness vibration in transducers with aspect ratio greater than 10.Cubic neural dust has a 1:1:1 aspect ratio, so coupled resonant modes alter the model’s validity and frequency predictions.
- Modeling approach: Finite-element simulations characterize cubic neural-dust transducers and provide parameters for a KLM-based ultrasonic link model.The model captures resonance shifts and higher resonances before integration into the complete link analysis.
- Link modeling: The complete link model cascades the interrogator, lossy brain tissue, and neural dust, using ABCD two-port parameters to analyze power-transfer efficiency.The analysis includes impedance matching at the interrogator input and models the ultrasonic channel through tissue.
- Power-delivery results: For a 100 µm node embedded 2 mm into the brain, ultrasonic transmission achieves 7 % efficiency (-11.6 dB) and approximately 500 µW received power with a 1 mm2 interrogator.The model assumes input-side impedance control because perfect output conjugate matching would require approximately 100 µH of infeasible inductance.
- Power-delivery results: Approximately 3.5 µW can be recovered by a 20 µm dust node through ultrasound, although designing a CMOS neural front-end within a 20 µm x 20 µm footprint remains extremely challenging.For nodes smaller than 100 µm, link efficiency drops at least quadratically as dust dimension decreases.
System design and constraints: Sensing / Communication
Neural dust must capture and communicate extracellular signals despite shrinking electrode separation, limited power, and tissue constraints. The proposed scaling analysis identifies 50 µm operation under stated assumptions and motivates wireless backscatter communication for tiny chronic nodes.
- Sensing: Extracellular recording requires closely placed recording and common electrodes within the same tiny dust footprint, limiting differential signal as separation shrinks.The maximum differential signal is constrained by electrode spacing on free-floating nodes.
- Sensing: As electrode spacing decreases, neural potential amplitude falls quadratically while front-end noise requirements increase within the available power budget.At 100 µm electrode separation, the expected action-potential amplitude is 10 µV.
- Scaling limit: 50 µm is the minimum operating neural-dust size for SNR = 3 under BaTiO3, FDA-approved ultrasound power limits, and no matching layers.Below 50 µm, captured power is insufficient to distinguish neural activity from noise under these assumptions.
- Communication: Backscattering reports neural data by modulating reflection of the incident carrier, avoiding the batteries and substantial energy storage required by active transmitters.This approach is attractive for ultra-miniature chronic implants with stringent size and power constraints.
- Communication: Active nodes require rectifiers, regulators, amplifiers, converters, and modulators, but CMOS front-end footprint and packaging create major scaling roadblocks.The smallest published CMOS neural front-end, excluding rectifiers and modulators, occupies approximately 100 µm2 of silicon.
System design and constraints: Passive node
The passive neural-dust architecture eliminates active front-end circuitry by using a MOSFET to modulate piezoelectric load impedance with neural voltage. This impedance variation changes ultrasonic reflectivity, encoding neural activity in the backscattered signal.
- Architecture: A MOSFET modulates the piezoelectric transducer impedance as a function of neural signals, eliminating the need for active front-ends.The approach uses the neural potential to control the electrical load presented by the transducer.
- Architecture: A single FET connects across the piezoelectric transducer and modulates IDS according to VGS derived from the neural signal.The proposed circuit uses the FET’s drain and source at the transducer terminals and the gate-to-source voltage for sensing.
- Biasing: An alternative biasing scheme lets the neural signal modulate VGS during both halves of the harvested ultrasonic sinusoid.Resistors place the neural potential between the gate and both alternating source/drain terminals.
- Biasing: Operating the FET in deep sub-threshold maximizes gm/IDS and allows a nominal VGS bias of 0 V.In this region, the FET behaves like a bipolar junction transistor, simplifying bias circuitry.
- Performance: 250 ppm sensitivity is obtained for a 10 µV neural signal after accounting for FET non-idealities.A perfect BJT would provide 400 ppm sensitivity, while FET non-idealities reduce it by a factor of 1.5–2.
System design and constraints: Interrogator
The interrogator must power, address, localize, and sensitively read many implanted nodes, while node transistor sizing is bounded by noise, available power, and physical area. The analysis reports passive operation scaling to roughly 20 µm.
- Interrogator requirements: The interrogator must extract node information, provide power, perform localization and addressing, and support BMI-relevant recording density.Sufficient receiver sensitivity is required to resolve neural spikes and distinguish among multiple sensor nodes.
- System constraints: Electromagnetic delivery is limited by tissue attenuation, geometry-dependent coupling, and regulated power density, motivating ultrasound for miniature implants.The analysis assumes neural-dust size and power, rather than interrogator capability, are the primary scaling bottlenecks.
- Optimization: For 100 µm and 20 µm dust nodes, FET widths of 75 µm and 16 µm maximize ΔPin, respectively.The optimum transistor width is relatively insensitive to threshold variability and electrode DC offsets because the maximum-reflection region is flat.
- Performance: At 100 µm, a FET sensor can generate a 16.6 ppm input-power change and approximately 120 nW (-39 dBm) of backscattered power.The estimate assumes a 1 mm2 interrogator aperture outputting 7.2 mW within the stated safety limit.
- Scaling limit: At 20 µm, the FET-size upper bound from available power falls below the lower bound imposed by noise, indicating passive scaling to roughly 20 µm.Transistor width is constrained simultaneously by thermal noise, node power, and the neural-dust area limit.
Re-design of neural dust node
Neural-dust scalability is constrained primarily by electrode separation and front-end noise rather than node size alone. The proposed compliant tails increase electrode separation while preserving miniature node dimensions, though interference makes interrogator sensitivity requirements more stringent.
- Re-design of neural dust node: The scaling of active and passive neural-dust nodes is limited by front-end noise requirements determined by achievable differential electrode signals.Decoupling implant size from achievable SNR could improve scalability.
- Re-design of neural dust node: Adding 1–5 µm-wide tails can position electrodes more than 50–100 µm from the neural-dust base, increasing separation without enlarging the node footprint.The design uses a flexible, ultra-compliant substrate populated with recording sites.
- Re-design of neural dust node: With 100 µm electrode separation on a 20 µm node, the noise limit permits 0.4 µm transistor width and 2.3e-3 ppm optimal sensitivity.The corresponding backscattered power is 16.6 pW (-77.8 dBm), within feasibility for a traditional CMOS receiver.
- Re-design of neural dust node: The analysis indicates that optimal transistor width for maximal sensitivity is small compared with the available neural-dust footprint, although impedance affects reflection coefficient and backscatter efficacy.The conclusion is not dependent on the exact technology used in the prior analysis.
- Re-design of neural dust node: Additional ultrasonic reflections from brain structures such as vasculature are not included in the analysis, making the interrogator sensitivity requirement more stringent than predicted.These reflections may cause intersymbol interference; active systems can use equalization or error correction, while passive systems may require filtering or pulse-based signaling.
Discussion and Conclusions
The analysis identifies three major realization challenges for ultra-small ultrasound-based neural recording systems: low-power, low-SNR front ends; integration and encapsulation; and sensitive low-power sub-cranial transceivers.
- Discussion and Conclusions: The first challenge is designing front ends that operate under decreasing available power and decreasing SNR with scale.Potential remedies include CMOS process and circuit innovation plus thinned, multi-substrate integration.
- Discussion and Conclusions: The second challenge is integrating extremely small piezoelectric transducers with CMOS electronics in a properly encapsulated package.The design must protect the implant while exposing two recording electrodes, and adding tails creates further fabrication challenges.
- Discussion and Conclusions: The third challenge is designing and implementing suitably sensitive sub-cranial transceivers that operate at low power.