-
Notifications
You must be signed in to change notification settings - Fork 0
Proof of Useful Work
Traditional Proof of Work (Bitcoin, Ethereum pre-merge) wastes energy on hash puzzles. Miners compute SHA-256 hashes billions of times, producing nothing useful.
Proof of Useful Work directs that computation toward real scientific problems. Instead of finding hash collisions, miners process real astronomical data with real ML models.
NASA Archive → FITS Download → Preprocessing → ML Inference → Quality Score → VOID Reward
- Download real data from public archives (MAST, SDSS, ZTF)
- Preprocess (filter, normalize, phase-fold)
- Run ML model on GPU (TransitCNN, Zoobot, AnomalyAE)
- Compute quality score from model metrics
- Submit results to MiningPool contract
- Receive VOID based on quality
Data: TESS 2-minute cadence light curves from MAST Model: AstroNetCNN (244K params, 89% accuracy) Output: Planet / False Positive / No Signal classification
What it does:
- Downloads FITS files from NASA MAST archive
- Extracts PDCSAP flux (detrended light curve)
- Quality filters bad cadences
- Outlier removal (5-sigma clip)
- Normalize and fill NaNs
- Phase-fold at orbital period
- Resample to fixed-length arrays (201 global, 81 local)
- Run multi-branch 1D CNN
- Output classification with confidence
Quality scoring:
- Based on model confidence (softmax probability)
- Higher confidence = higher quality
- Range: 50-100
Data: SDSS DR18 galaxy images Model: GalaxyClassifier (ConvNeXT architecture) Output: Spiral / Elliptical / Irregular / Merger / Unknown
What it does:
- Downloads galaxy cutouts from SDSS SkyServer
- Resize to 224x224 RGB
- Normalize to [0, 1]
- Run ConvNeXT classifier
- Output morphology class with probabilities
Quality scoring:
- Based on classification confidence
- Higher confidence = higher quality
- Range: 50-100
Data: ZTF alerts from Fink broker Model: AnomalyAE (autoencoder) Output: Anomaly scores + flags
What it does:
- Downloads alerts from ZTF/Fink API
- Extract features from alert stream
- Run autoencoder
- Compute reconstruction error
- Flag anomalies (> 2σ from mean)
Quality scoring:
- Based on anomaly score magnitude
- More anomalous = higher quality
- Range: 50-100
| Score | Status | Reward | Description |
|---|---|---|---|
| < 50 | REJECTED | 0 | Noise, invalid work |
| 50-69 | ACCEPTED | 1x | Base reward |
| 70-89 | GOOD | 1.2x | Bonus for high quality |
| 90-100 | EXCELLENT | 1.5x | Maximum reward |
- TESS light curves: Used by NASA to discover exoplanets
- SDSS images: Used by astronomers to study galaxy evolution
- ZTF alerts: Used to detect supernovae and other transients
- Results stored on IPFS with CID on-chain
- Anyone can verify predictions
- Models are pre-trained and reproducible
- Exoplanet candidates can be submitted to NASA
- Galaxy classifications contribute to Galaxy Zoo
- Anomaly detections help find rare transients
- Scale: 1000s of GPUs > 1 data center
- Cost: Miners pay for their own electricity
- Decentralization: No single point of failure
- Incentive alignment: Miners earn for useful work
- Open data: Results publicly accessible
| Aspect | Bitcoin PoW | Voidmap PoUW |
|---|---|---|
| Computation | SHA-256 hashing | ML inference |
| Output | Hash collisions | Scientific predictions |
| Usefulness | None | Exoplanet detection |
| Energy waste | High | Low (useful work) |
| Hardware | ASICs only | GPUs only |
| Centralization risk | High (ASIC farms) | Low (consumer GPUs) |
-
inputHash: Hash of input data (verifiable against archive) -
outputHash: Hash of output predictions -
modelHash: Hash of model weights -
ipfsCID: IPFS CID of full results
- Results saved locally at
~/.voidmap/results/ - Full predictions accessible via IPFS gateway
- Models are pre-trained and publicly available
- More tasks: Additional ML tasks (spectral classification, etc.)
- Better models: Larger, more accurate pre-trained models
- Real-time streaming: Process ZTF alerts in real-time
- Result distribution: Submit results to NASA/ESA archives
- Community models: Allow miners to contribute custom models