Ecovative's Mycelium Chips Perform Machine Learning Tasks in DARPA-Funded Study

• Researchers at Ecovative have demonstrated that mycelium infused with a conductive polymer can perform machine learning tasks, publishing their findings in Scientific Reports in June 2026.

• The chips exploit physical reservoir computing, a technique that uses the natural nonlinear dynamics of a physical material to process time-varying information without requiring a conventional silicon processor.

• With production yields exceeding three million chips per growth cycle and a biodegradable construction, the approach could offer a low-cost, sustainable alternative to conventional analogue computing hardware.

---

Fungal Networks as Computing Hardware

The idea that a mushroom's root-like structure could double as a computing substrate sounds implausible until you consider what mycelium actually does. The hyphal networks that fungi use to colonize a substrate are intricate, branching, and electrically interactive when treated correctly. A team at Ecovative, the Green Island, New York company better known for its mycelium-based packaging and food products, has now demonstrated that those networks can be engineered into functional analogue computing chips.

Neuromorphic Mycelium Chip mounted on a carrier board. Credits: Telhan et al.

The research, published in Scientific Reports on 10 June 2026 and funded by the Defense Advanced Research Projects Agency (DARPA) under grant DARPA-SCA-24-01, introduces what the authors describe as a neuromorphic computing substrate: a material that processes information in a manner loosely analogous to how biological neural tissue does, through nonlinear electrical dynamics rather than binary on-off switching.

The "Design-Grow-Compute" Workflow

The chips are produced by infusing mycelium with PEDOT:PSS, a conductive polymer widely used in flexible electronics. The infusion renders the fungal network electrically active, transforming its hyphae into a mesh of resistors, capacitors, and non-linear elements. The team developed what they call a "design-grow-compute" workflow, which integrates morphological modelling, parametric growth protocols under controlled environmental conditions, and vacuum-assisted polymer infusion to achieve consistent, repeatable chip architectures.

Credits: Telhan et al.

The physical reservoir computing framework the team employs is worth explaining. In conventional machine learning, a neural network is trained by adjusting the connections between computational nodes. Physical reservoir computing sidesteps that process by using the inherent complexity of a physical material as a fixed computational layer. Time-varying input signals are fed into the material, which transforms them into high-dimensional nonlinear responses; a simple readout layer then interprets those responses to complete a task. The mycelium network, with its branching geometry and charge-transport properties, serves as that fixed physical layer.

To benchmark the chips, the authors used NARMA-10, a standard sequence-prediction test in the reservoir computing literature that requires a system to model nonlinear temporal dependencies across ten time steps. The chips demonstrated robust nonlinearity and temporal dynamics sufficient to perform the task. Benchmarking also included linear memory capacity, fading-memory relaxation measurements, and assessments of device-to-device variability across independently fabricated chips.

A critical design variable turned out to be morphological complexity: the degree of branching and density of the hyphal network influences both charge transport and memory capacity. That relationship gives engineers a new axis of control, adjusting growth conditions to tune computational properties rather than redesigning silicon circuitry. The prototype chips interface with a custom carrier board that handles analogue signal conditioning and readout.

Scalability, Cost, and Honest Trade-Offs

The economic and environmental case rests on two figures the authors report. First, production yields exceeding three million chips per growth cycle, achievable using existing mushroom farming infrastructure. Second, a biodegradable construction that, unlike conventional silicon or memristor-based hardware, does not require rare materials or generate persistent electronic waste.

Credits: Telhan et al.

The authors position their platform against three alternative reservoir computing approaches: memristor arrays, photonic systems, and living-cell-based reservoirs. The mycelium chip, being non-living after processing, avoids the biological containment requirements of living-cell systems while remaining far cheaper and more sustainable than photonic or memristor architectures, at least at the proof-of-concept stage.

The team is candid about trade-offs. The paper acknowledges performance compromises relative to more mature computing substrates, framing them as justified by the sustainability advantages and cost reductions the platform offers. The manuscript is also being published in an unedited early-access form ahead of final peer review, meaning further editorial scrutiny may affect specific claims or data points.

This work sits within a broader pattern of interest in mycelium as a functional electronic material. Earlier research demonstrated that shiitake mushroom mycelium could be used to construct a working memristor, a type of electronic component that retains information based on past electrical history. Ecovative's reservoir computing demonstration extends that direction considerably, moving from single components toward complete computational architectures.

What the Ecovative study does not yet establish is how the chips perform outside laboratory conditions, whether production consistency holds at true industrial scale, or how the approach competes with established analogue computing hardware on tasks beyond NARMA-10. Those questions mark the boundary between a compelling proof of concept and a commercially deployable technology.