Digital Twins for Fungal Computing: Viable XOR Regimes, Parameter Inference, and Waveform-Guided Rediscovery

This paper presents a digital-twin workflow for fungal computing that identifies viable XOR parameter regimes, infers key biophysical parameters from electrical data using machine learning, and validates a waveform-guided refinement process that significantly improves model accuracy and reduces parameter error in simulated mycelial networks.

The Gist

Imagine you have a living, breathing computer made of mushroom roots (called mycelium). These roots are amazing: they can send electrical signals, remember past events, and even perform simple math like an "XOR" gate (a logic operation that says "yes" only if one input is true, but not both).

The problem? Every mushroom is different. Just like no two human fingerprints are identical, no two mushroom networks are exactly alike. One might be great at math, while its neighbor is terrible at it. This makes building reliable mushroom computers very hard.

This paper introduces a solution: A "Digital Twin" for mushrooms.

Think of a Digital Twin like a video game simulation of a real mushroom. Before you touch the actual, messy, living fungus, you create a perfect virtual copy on your computer. You can poke, prod, and test this virtual copy thousands of times to figure out how to make it work, then apply those lessons to the real thing.

Here is how the researchers did it, broken down into simple steps:

1. Finding the "Sweet Spot" (The Viable Regime)

Imagine you are trying to tune a radio to find a clear station. If you just spin the dial randomly, you might never find it.

• What they did: The researchers simulated 160 different virtual mushrooms. They tweaked the "knobs" (biological settings like speed and sensitivity) and the "antenna placement" (where they put the electrical probes).
• The Result: They found that only a specific range of settings works for doing math. It's like finding that the radio only works clearly between 98.5 and 100.0 MHz. They mapped out this "sweet spot" so future researchers know exactly what kind of mushroom to look for.

2. The Detective Work (Parameter Inference)

Now, imagine you have a real mushroom, but you can't see inside its brain. You can only zap it with electricity and listen to the sparks.

• The Challenge: Can you guess the mushroom's internal settings just by watching how it reacts to zaps?
• The Solution: They used a computer program (Machine Learning) as a detective. They fed it data from 400 virtual mushrooms, showing it three types of "tests":
  1. The Step Test: A long, steady zap (like holding a button down).
  2. The Double Tap: Two quick zaps close together (to see if it remembers the first one).
  3. The Triangle Sweep: A voltage that slowly goes up and down (to check for memory effects).
• The Result: The detective was surprisingly good! It could accurately guess the mushroom's "speed" and "sensitivity" settings about 80-90% of the time. However, it struggled to guess the exact "resistance" (how hard it is for electricity to flow), kind of like trying to guess the weight of a box just by shaking it.

3. The "Fine-Tuning" (Rediscovery)

The detective gave a good guess, but it wasn't perfect.

• The Fix: The researchers took the detective's guess and used it as a starting point. Then, they ran a "fine-tuning" process where they adjusted the virtual mushroom's settings until its electrical reactions matched the real mushroom's reactions perfectly.
• The Result: This two-step process (Guess + Fine-Tune) reduced the error by 96%. It's like a tailor making a suit: first, they guess your size based on a photo (Machine Learning), then they have you try it on and pin the fabric to get the perfect fit (Waveform Matching).

4. What Actually Matters? (Sensitivity)

The researchers asked: "If we get one of these settings wrong, does it ruin the computer?"

• The Discovery: They found a funny pattern. The settings that were hardest to guess (like the exact resistance) were actually the least important for the math to work.
• The Good News: The settings that were easiest to guess (like speed and sensitivity) were the most critical.
• The Metaphor: Imagine driving a car. If you get the color of the paint wrong, the car still drives fine (easy to guess, doesn't matter). But if you get the engine timing wrong, the car won't start (hard to guess, but matters a lot). Fortunately, the researchers found that their "guessing machine" was best at guessing the engine timing.

Why This Matters

This paper is a roadmap for building biological computers.

• Before: Scientists were guessing in the dark, trying random mushrooms and hoping for the best.
• Now: They have a Digital Twin workflow. They can simulate a mushroom, figure out exactly what settings it needs, and then go find a real mushroom that matches those settings.

In short: They built a virtual simulator that teaches us how to talk to mushroom brains, helping us turn nature's messy, living networks into reliable, living computers.

Technical Summary

1. Problem Statement

Fungal computing utilizes living mycelial networks as substrates for unconventional, adaptive information processing. While promising, the field faces a critical bottleneck: specimen-to-specimen variability. Differences in network topology, excitability, recovery times, and adaptive conductance mean that an electrode configuration successful on one fungal specimen often fails on another. This variability prevents the reproducible fabrication of logic gates (specifically the XOR gate, a non-linearly separable operation).

The paper addresses the need for a computational framework (a "digital twin") that can:

1. Identify which biophysical parameter regimes support XOR computation.
2. Infer latent biophysical parameters from limited electrical characterization data.
3. Refine these inferences to enable specimen-specific calibration without requiring full physical re-optimization.

2. Methodology

The authors propose a multi-stage computational pipeline using a Fungal Digital Twin model.

A. The Fungal Network Model

• Structure: The mycelium is modeled as a Random Geometric Graph (RGG) embedded in a 20×20 mm domain. Nodes represent hyphal junctions, and edges connect nodes within a 5 mm coupling radius.
• Dynamics:
  • Nodes: Governed by FitzHugh–Nagumo (FHN) equations to simulate excitable, spiking behavior (fast voltage $v$, slow recovery $w$).
  • Edges: Feature memristive conductance ($g_{ij}$) that depends on the history of voltage differences, modeled by a state variable $m_{ij}$ and an adaptation rate $\alpha$.
• Parameters: The model contains 8 latent biophysical parameters: $\tau_v, \tau_w, a, b$ (FHN dynamics), $v_{scale}$ (voltage scaling), $R_{on}, R_{off}$ (memristor resistance states), and $\alpha$ (adaptation rate).

B. Three-Stage Workflow

1. Viable Regime Discovery (Optimization):

   • A systematic study over 160 simulated specimens (varying network sizes from 20 to 120 nodes).
   • Bayesian Optimization was used to tune electrode geometry, stimulus voltage, pulse duration, and inter-input delay to maximize XOR performance.
   • A second stage optimized the 8 biophysical parameters while holding the geometry fixed to isolate intrinsic substrate properties.
   • Viability Definition: Specimens in the top quartile of XOR scores were classified as "viable."

2. Parameter Inference (Characterization):

   • A dataset of 400 simulated specimens was generated.
   • Each specimen was probed with three electrical protocols to extract 94 response features:
     • Step Response: 2V pulse (3000ms) to measure transient excitation/recovery.
     • Paired-Pulse: Two pulses with varying inter-pulse intervals to probe facilitation/depression.
     • Triangle Sweep: Linear voltage ramp to measure hysteresis (memristive behavior).
   • Machine Learning: Random Forest (RF) and Multilayer Perceptron (MLP) regressors were trained to predict the 8 latent parameters from the 94 features.

3. Rediscovery & Refinement:

   • A validation study on 15 optimized specimens tested a "Rediscovery" workflow.
   • Condition 1 (ML Only): Parameters predicted directly from characterization features.
   • Condition 2 (ML + Refine): ML predictions served as initialization for a local optimization (dual annealing) that minimized the waveform mismatch between the twin and the target specimen.

4. Sensitivity Analysis:

   • Perturbation analysis on 72 viable specimens to determine which parameters most critically impact XOR accuracy when varied.

3. Key Results

A. Viable XOR Subspace

• XOR viability is not random; it occupies a constrained region of the parameter space.
• Key Constraints: Viable specimens cluster around specific time scales ($\tau_v \approx 54$ ms, $\tau_w \approx 894$ ms) and excitability thresholds ($a \approx 0.67$).
• Multimodality: Four parameters ($\tau_v, b, v_{scale}, R_{off}$) showed bimodal distributions, suggesting multiple distinct dynamical regimes can support XOR, rather than a single "optimal" set.
• Network Size: Larger networks (100–120 nodes) showed a higher probability of yielding viable specimens, though no single topology was privileged.

B. Parameter Identifiability

• High Accuracy: The Random Forest model reliably recovered time constants and the excitability threshold:
  • $\tau_v$: $R^2 = 0.912$
  • $\tau_w$: $R^2 = 0.816$
  • $a$: $R^2 = 0.717$
• Low Accuracy: Memristive resistance states ($R_{on}, R_{off}$) and voltage scaling ($v_{scale}$) were poorly identifiable ($R^2 \le 0.02$) using the current protocols.
• Protocol Contribution:
  • Step Response was most informative for $\tau_v$.
  • Paired-Pulse was critical for the excitability threshold $a$.
  • Triangle Sweep combined with Step Response was best for $\tau_w$ and memristive adaptation $\alpha$.

C. Rediscovery Validation

• Waveform Matching: The "ML + Refine" approach significantly outperformed "ML Only."
  • Waveform Mismatch: Reduced from 1.070 to 0.042 (96.0% reduction, $p = 3.1 \times 10^{-5}$).
  • Core Parameter Error: Reduced from 16.6% to 8.8% ($p = 6.1 \times 10^{-5}$).
• Refinement Impact: The refinement stage successfully corrected errors in $\tau_w$, $a$, $R_{on}$, and $\alpha$, though it struggled with $v_{scale}$ (which was already poorly constrained).

D. Sensitivity Hierarchy

• Critical Parameters: The recovery time constant ($\tau_w$) and memristive adaptation rate ($\alpha$) were the most consequential for XOR accuracy. Small perturbations here caused the largest drops in performance.
• Tolerable Parameters: $v_{scale}$ and $R_{off}$ had negligible impact on XOR accuracy even with large perturbations.
• Alignment: The workflow recovers the most important parameters ($\tau_v, \tau_w, a$) most accurately, while the parameters it fails to recover ($v_{scale}, R_{off}$) are the least critical for the computation.

4. Key Contributions

1. Computational Framework: A complete pipeline for fungal digital twins that bridges substrate characterization, logic-gate viability prediction, and specimen-specific calibration.
2. Viable Regime Mapping: Identification of specific biophysical constraints required for XOR computation, moving fungal computing from trial-and-error to targeted design.
3. Identifiability Analysis: Demonstration that standard electrophysiological protocols can recover key dynamical parameters ($\tau, a$) but fail to resolve structural resistance parameters, highlighting a gap in current experimental design.
4. Hybrid Calibration Strategy: Proof that combining ML prediction with local waveform-matching refinement drastically improves twin fidelity, enabling small-scale specimen-specific optimization.
5. Sensitivity-Identifiability Alignment: The finding that the parameters most critical for computation are also the most identifiable, suggesting the current workflow is robust despite its limitations.

5. Significance and Future Directions

• Practical Impact: This work provides a roadmap for experimentalists to select viable fungal specimens and design electrode configurations without needing to "guess" the internal state of the living tissue.
• Protocol Design: The study suggests a specific experimental ordering: Step Response (baseline) $\to$ Paired-Pulse (for excitability) $\to$ Triangle Sweep (for hysteresis). Future protocols should target the memristive adaptation rate ($\alpha$), which is currently hard to predict but highly consequential.
• Limitations: The current study assumes fixed network topology (the twin copies the specimen's graph). Full rediscovery of topology from electrical data remains an open, harder inverse problem. Additionally, the study is currently simulation-based; experimental validation on real mycelium is the necessary next step.
• Broader Scope: While focused on XOR, the framework is generalizable to reservoir computing and other logic gates, offering a scalable infrastructure for the emerging field of biological computing.

In conclusion, the paper establishes that fungal digital twins are a viable tool for narrowing the search space of computational substrates, partially recovering their governing dynamics, and enabling specimen-specific refinement, marking a significant step toward reproducible biological computing.