Nature Preserve — Applied NPU Science Across 7 Domains

Date: April 30, 2026 Status: 7 domain application patterns documented. 28-model zoo (21 BrainChip + 4 physics + 2 NeuroBench + 1 hand-built). 5 standalone science demo binaries. Pure Rust conversion pipeline operational — no Python required at any stage. Domain: Applied neuromorphic inference across physics, biology, audio, vision, environmental, genomic, and industrial science Novelty: First documented application layer connecting a neuromorphic model zoo to multi-domain scientific use. First pure Rust pipeline from trained weights to deployed NPU model (.npy/.safetensors → quantize → FlatBuffer → Snappy → .fbz). First systematic bridge between curated AI models and domain-specific scientific workflows. Cross-Spring: 🔥♨️ hotSpring (lattice QCD steering, ESN readout) × 💧♨️ wetSpring (biological classifiers, bloom sentinel, spectral triage) × 🌬️♨️ airSpring (environmental monitoring, ET₀) × 🧠♨️ neuralSpring (quantization validation, dispatch cost models)

Abstract

The 🦀🧠 rustChip Zoo contains 29 curated neural network models that run on BrainChip Akida neuromorphic processors — parsed, converted, and validated entirely in Rust. But a zoo is a collection of specimens. The Nature Preserve is where those specimens live in their natural habitat: applied to real scientific problems, with real data, producing real decisions.

This paper documents 7 domain application patterns that bridge the zoo to science:

Each pattern follows the same shape: domain data → feature extraction → quantization → NPU inference → domain interpretation → output. The feature extraction and interpretation steps are domain-specific; everything between is generic and handled by 🦀🧠 rustChip’s crates.

1. The Pipeline Pattern

Every domain follows one pipeline:

Domain data (measurements, signals, sequences)
    │
    ▼
Feature extraction (domain-specific, in Rust)
    │
    ▼
Quantization (f32/f64 → int4/int8, symmetric per-layer)
    │
    ▼
NPU inference (rustChip: parse model, load via VFIO, run)
    │
    ▼
Domain interpretation (classify, predict, detect, steer)
    │
    ▼
Output (decision, measurement, alert, next-step control)

The feature extraction step is where domain expertise lives. The NPU inference step is where hardware acceleration lives. By separating them cleanly, the same hardware pipeline serves seven different sciences.

2. Physics: NPU as Simulation Steering Engine

Problem: Lattice QCD trajectories take 10–60 seconds on GPU. Between trajectories, the simulation must decide: accept or reject? continue thermalizing? which observable to measure next? These decisions must happen at sub-millisecond latency to avoid blocking the GPU.

Model: ESN readout (InputConv(50)→FC(128)→FC(1), int4). The echo state network reservoir runs on CPU or GPU; only the readout runs on NPU.

Evidence: 5,978 live AKD1000 calls over 24 hours in 🔥♨️ hotSpring Experiment 022. Phase classifier achieved 100% accuracy on confined vs deconfined SU(3) gauge configurations.

Extension: The reservoir-readout split works for any system where a temporal model must make fast decisions between expensive compute steps — molecular dynamics, weather prediction, financial simulation.

3. Biology: NPU as Pre-filter for Sequencing Pipelines

Problem: Biological pipelines process millions of reads. At multiple stages, small classification decisions gate expensive downstream computation. The NPU provides microsecond per-read classification that reduces downstream compute by 10–100×.

Models: ESN → int8 classifiers for quorum sensing (3-class), phylogenetic placement (5-class), genome binning (10-class), spectral triage (2-class), bloom sentinel (12-channel).

Evidence: 11 validate_npu_* binaries in 💧♨️ wetSpring, covering QS classification, phylo placement, genome binning, spectral screening, disorder classification, and bloom monitoring.

Extension: Any biological pipeline with a sparse-positive classification step benefits — nanopore quality scoring, PFAS environmental screening, pangenome navigation.

4. Audio: NPU for Always-On Acoustic Intelligence

Problem: Continuous audio streams require sub-100ms keyword detection and event classification at milliwatt power, without cloud connectivity.

Models: DS-CNN KWS (41 KB, 33 keywords), TENN Recurrent SC12 (70 KB, 12 commands), TENN Recurrent UORED (37 KB, 4 utterances). All parsed and validated by 🦀🧠 rustChip’s FBZ parser.

Extension: Custom wake words, multi-language detection, acoustic event classification (industrial sounds, environmental monitoring, accessibility).

5. Vision: NPU for Edge Perception

Problem: Visual perception at the edge — detect, segment, classify, identify — at frame rate, without cloud dependency.

Models: 14 BrainChip pretrained models spanning object detection (YOLO VOC, CenterNet), segmentation (UNet Portrait), classification (AkidaNet ImageNet, PlantVillage, GXNOR MNIST), face analysis (FaceID, UTK Face), gesture (DVS, Samsung), and 3D (PointNet++).

Extension: Agricultural monitoring (PlantVillage → crop-specific diseases), multi-camera fusion via multi-tenancy, privacy-preserving face analysis (embeddings on-device, no raw images transmitted).

6. Environmental: NPU for Continuous Ecosystem Monitoring

Problem: High-cadence environmental sensors generate continuous data streams that need real-time classification without cloud latency — on buoys, in fields, at remote watersheds.

Models: Streaming sensor 12ch (bloom classification), adaptive sentinel (domain-shift detection), ET₀ regressor (evapotranspiration estimation).

Evidence: 🌬️♨️ airSpring validate_npu_eco and validate_npu_high_cadence; 💧♨️ wetSpring validate_npu_bloom_sentinel and validate_npu_sentinel_stream.

Extension: Real-time irrigation control, satellite-derived feature augmentation, drought early warning.

7. Genomic: NPU as Throughput Multiplier

Problem: Genomic analysis pipelines process terabytes. At multiple stages, small classifiers gate expensive computation — the NPU pre-filters at microsecond latency.

Models: K-mer classifiers, genome binning models, spectral triage pre-filters, multi-head population genetics readouts.

Evidence: 💧♨️ wetSpring validate_npu_genome_binning, validate_npu_spectral_triage, validate_npu_spectral_screen; 🧠♨️ neuralSpring validate_introgression, validate_pangenome_selection, validate_upstream_kmer.

Extension: Nanopore real-time quality scoring, adaptive sampling, PFAS environmental genomics.

8. Industrial: NPU for Predictive Maintenance

Problem: Manufacturing and infrastructure produce high-volume sensor data. “Is this normal?” is a continuous classification task that must run on-site, in real time, without cloud dependency.

Models: Streaming sensor 12ch (multi-subsystem health), adaptive sentinel (domain-shift detection), DVS gesture models (operator action monitoring).

Extension: Frequency-domain vibration analysis, remaining useful life regression, multi-model fleets on edge nodes (one self-contained Rust binary per asset).

9. The Pure Rust Pipeline

All 7 domains share one conversion pipeline, implemented entirely in Rust:

1. Import weights from .npy (hand-rolled parser) or .safetensors
2. Quantize to int1/2/4/8 (symmetric max-abs, per-layer or per-channel)
3. Serialize to FlatBuffer (reverse-engineered Akida schema)
4. Compress with Snappy (matching the vendor’s .fbz format)
5. Write the .fbz model file

cargo run -p akida-cli -- convert \
  --weights trained.npy \
  --arch "InputConv(50,1,1) FC(128) FC(1)" \
  --output model.fbz --bits 4

No Python. No TensorFlow. No vendor SDK. The pipeline handles round-trip verification automatically.

10. How to Start

git clone https://github.com/syntheticChemistry/rustChip.git
cd rustChip
cargo build --workspace
cargo run -p akida-cli -- convert --weights "random:6400" \
  --arch "InputConv(50,1,1) FC(128) FC(1)" --output my_model.fbz --bits 4
cargo run -p akida-cli -- parse my_model.fbz

Full documentation: QUICKSTART.md, Nature Preserve, LEVERAGE.md.