Claude’s Corner: Terranox AI — Finding Uranium With AI Instead of Intuition

---
description: Build a Terranox AI clone — multimodal geoscience ML for mineral prospectivity mapping and sequential drilling decisions
---

# Build Terranox AI: AI-Powered Mineral Prospectivity Mapping

## What You're Building
A geoscience AI platform that ingests heterogeneous geological data (geophysics, geochemistry, drill logs, geology maps), fuses it into a unified context, runs a trained prospectivity model to generate probability heatmaps, and uses sequential decision intelligence to recommend the optimal next exploration action. We will build a uranium-focused version but the architecture works for any critical mineral.

## Tech Stack
- **Frontend:** Next.js 14 (App Router), React, Mapbox GL JS (map visualization)
- **Backend:** Python FastAPI (ML serving), Next.js API routes (orchestration)
- **Database:** Supabase (Postgres + PostGIS extension, file storage)
- **ML/AI:** PyTorch (prospectivity model), scikit-learn (feature engineering), BoTorch (Bayesian optimization for sequential decisions)
- **Geospatial:** GDAL, rasterio, geopandas, shapely
- **Key Libraries:** torch, xgboost, lightgbm, botorch, geopandas, rasterio, GDAL, pydeck

## Step 1: Project Setup

```bash
# Next.js frontend
npx create-next-app@latest terranox-clone --typescript --tailwind --app
cd terranox-clone

# Python ML backend
mkdir ml-service && cd ml-service
python3 -m venv venv && source venv/bin/activate
pip install fastapi uvicorn torch torchvision numpy pandas geopandas rasterio shapely scikit-learn xgboost lightgbm botorch gpytorch requests python-dotenv
```

File structure:
```
terranox-clone/
  app/
    api/
      prospectivity/route.ts     # API route calling ML service
      next-action/route.ts       # Sequential decision API
      upload-data/route.ts       # Data ingestion endpoint
    project/
      [id]/page.tsx              # Main map + prospectivity view
    page.tsx                     # Dashboard
  src/
    components/
      Map/ProspectivityLayer.tsx # Heatmap overlay on Mapbox
      Map/DrillHoleMarkers.tsx   # Historical drill results
      Controls/ActionPanel.tsx   # Next-action recommendations
    lib/
      geo/raster-utils.ts
      supabase.ts
  ml-service/
    main.py                      # FastAPI app
    models/
      prospectivity_model.py     # PyTorch CNN + tabular fusion
      sequential_decision.py     # BoTorch Bayesian optimization
    data/
      ingestion.py               # Multi-source data normalization
      features.py                # Geoscience feature engineering
```

## Step 2: Core Data Models

```sql
-- Enable PostGIS for spatial queries
CREATE EXTENSION IF NOT EXISTS postgis;

CREATE TABLE projects (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  user_id UUID NOT NULL,
  name TEXT NOT NULL,
  target_mineral TEXT DEFAULT 'uranium',
  bounds GEOMETRY(POLYGON, 4326),
  status TEXT DEFAULT 'active',
  created_at TIMESTAMPTZ DEFAULT NOW()
);

CREATE TABLE drill_holes (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  project_id UUID REFERENCES projects(id),
  hole_id TEXT NOT NULL,
  location GEOMETRY(POINT, 4326),
  depth_m FLOAT,
  result TEXT CHECK (result IN ('mineralized', 'barren', 'anomalous')),
  grade_ppm FLOAT,
  lithology TEXT,
  geochemistry JSONB,
  geophysics JSONB,
  drilled_at DATE,
  created_at TIMESTAMPTZ DEFAULT NOW()
);

CREATE TABLE raster_layers (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  project_id UUID REFERENCES projects(id),
  layer_type TEXT NOT NULL,
  storage_path TEXT NOT NULL,
  resolution_m FLOAT,
  crs TEXT DEFAULT 'EPSG:4326',
  metadata JSONB,
  uploaded_at TIMESTAMPTZ DEFAULT NOW()
);

CREATE TABLE prospectivity_results (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  project_id UUID REFERENCES projects(id),
  model_version TEXT,
  output_raster_path TEXT,
  top_targets JSONB,
  model_metrics JSONB,
  generated_at TIMESTAMPTZ DEFAULT NOW()
);

CREATE TABLE action_recommendations (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  project_id UUID REFERENCES projects(id),
  action_type TEXT,
  location GEOMETRY(POINT, 4326),
  priority_score FLOAT,
  expected_info_gain FLOAT,
  reasoning TEXT,
  status TEXT DEFAULT 'pending',
  created_at TIMESTAMPTZ DEFAULT NOW()
);
```

## Step 3: Geoscience Data Ingestion Pipeline

```python
# ml-service/data/ingestion.py
import numpy as np
import pandas as pd
import geopandas as gpd
import rasterio

class GeoscienceDataIngester:
    def ingest_drill_logs(self, csv_path: str, crs: str = "EPSG:4326") -> gpd.GeoDataFrame:
        df = pd.read_csv(csv_path)
        col_map = {
            "latitude": "lat", "Latitude": "lat", "LAT": "lat",
            "longitude": "lon", "Longitude": "lon", "LON": "lon",
            "grade": "grade_ppm", "U3O8_ppm": "grade_ppm",
        }
        df = df.rename(columns={k: v for k, v in col_map.items() if k in df.columns})

        def normalize_result(r):
            r = str(r).lower()
            if any(x in r for x in ["mineralized", "hit", "ore"]):
                return "mineralized"
            elif any(x in r for x in ["anomalous", "trace"]):
                return "anomalous"
            return "barren"

        df["result"] = df["result"].apply(normalize_result)
        return gpd.GeoDataFrame(
            df, geometry=gpd.points_from_xy(df["lon"], df["lat"]), crs=crs
        )

    def ingest_raster(self, tif_path: str, layer_type: str) -> dict:
        with rasterio.open(tif_path) as src:
            data = src.read(1).astype(np.float32)
            valid = data[data != src.nodata] if src.nodata else data.flatten()
            if valid.std() > 0:
                data = (data - valid.mean()) / valid.std()
            return {
                "layer_type": layer_type,
                "data": data,
                "bounds": src.bounds,
                "crs": str(src.crs),
                "transform": src.transform,
            }

    def extract_point_features(self, point_gdf, raster_layers):
        features = []
        for _, row in point_gdf.iterrows():
            point_features = []
            for layer in raster_layers:
                try:
                    col, row_idx = ~layer["transform"] * (row.geometry.x, row.geometry.y)
                    val = layer["data"][int(row_idx), int(col)]
                    point_features.append(float(val))
                except (IndexError, TypeError):
                    point_features.append(0.0)
            # Pathfinder elements for uranium
            geochem = {}
            if "geochemistry" in row and isinstance(row["geochemistry"], dict):
                geochem = row["geochemistry"]
            for elem in ["U", "Th", "Mo", "V", "Se", "As", "Pb"]:
                point_features.append(float(geochem.get(elem, 0.0)))
            features.append(point_features)
        return np.array(features, dtype=np.float32)
```

## Step 4: Prospectivity Model (Multimodal Fusion)

```python
# ml-service/models/prospectivity_model.py
import torch
import torch.nn as nn
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score

class GeoscienceFusionModel(nn.Module):
    # CNN branch: spatial patch from stacked geophysics rasters
    # MLP branch: point-level geochemistry + tabular data
    # Fusion head: combined classifier
    def __init__(self, raster_channels: int, tabular_features: int):
        super().__init__()
        self.cnn = nn.Sequential(
            nn.Conv2d(raster_channels, 32, kernel_size=3, padding=1),
            nn.BatchNorm2d(32), nn.ReLU(), nn.MaxPool2d(2),
            nn.Conv2d(32, 64, kernel_size=3, padding=1),
            nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(2),
            nn.Conv2d(64, 128, kernel_size=3, padding=1),
            nn.ReLU(), nn.AdaptiveAvgPool2d((4, 4)),
        )
        cnn_out = 128 * 4 * 4
        self.mlp = nn.Sequential(
            nn.Linear(tabular_features, 64), nn.ReLU(), nn.Dropout(0.3),
            nn.Linear(64, 128), nn.ReLU(),
        )
        self.classifier = nn.Sequential(
            nn.Linear(cnn_out + 128, 256), nn.ReLU(), nn.Dropout(0.4),
            nn.Linear(256, 64), nn.ReLU(), nn.Linear(64, 1), nn.Sigmoid(),
        )

    def forward(self, raster_patch, tabular):
        cnn_features = self.cnn(raster_patch).flatten(1)
        mlp_features = self.mlp(tabular)
        return self.classifier(torch.cat([cnn_features, mlp_features], dim=1))


def train_prospectivity_model(X_raster, X_tabular, y_labels, n_epochs=50):
    model = GeoscienceFusionModel(X_raster.shape[1], X_tabular.shape[1])
    optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
    # Uranium deposits are rare -- weight positive class heavily
    pos_weight = torch.tensor([(y_labels == 0).sum() / max((y_labels == 1).sum(), 1)])
    criterion = nn.BCEWithLogitsLoss(pos_weight=pos_weight)

    X_r_tr, X_r_val, X_t_tr, X_t_val, y_tr, y_val = train_test_split(
        X_raster, X_tabular, y_labels, test_size=0.2, stratify=y_labels
    )
    for epoch in range(n_epochs):
        model.train()
        optimizer.zero_grad()
        preds = model(torch.FloatTensor(X_r_tr), torch.FloatTensor(X_t_tr)).squeeze()
        loss = criterion(preds, torch.FloatTensor(y_tr))
        loss.backward()
        optimizer.step()
        if epoch % 10 == 0:
            model.eval()
            with torch.no_grad():
                val_p = model(torch.FloatTensor(X_r_val), torch.FloatTensor(X_t_val)).squeeze().numpy()
            print(f"Epoch {epoch} | Loss: {loss.item():.4f} | AUC: {roc_auc_score(y_val, val_p):.4f}")
    return model


def generate_prospectivity_map(model, raster_stack, patch_size=32):
    model.eval()
    channels, height, width = raster_stack.shape
    heatmap = np.zeros((height, width))
    half = patch_size // 2
    with torch.no_grad():
        for row in range(half, height - half, 4):
            for col in range(half, width - half, 4):
                patch = raster_stack[:, row-half:row+half, col-half:col+half]
                prob = model(
                    torch.FloatTensor(patch).unsqueeze(0),
                    torch.zeros(1, model.mlp[0].in_features)
                ).item()
                heatmap[row-2:row+2, col-2:col+2] = prob
    return heatmap
```

## Step 5: Sequential Decision Intelligence (Bayesian Optimization)

```python
# ml-service/models/sequential_decision.py
import torch
from botorch.models import SingleTaskGP
from botorch.fit import fit_gpytorch_mll
from botorch.acquisition import qExpectedImprovement
from botorch.optim import optimize_acqf
from gpytorch.mlls import ExactMarginalLogLikelihood

class ExplorationOptimizer:
    # Bayesian optimization selects next drill location to maximize
    # expected improvement over current best observed grade.
    def __init__(self, bounds_lon: tuple, bounds_lat: tuple):
        self.bounds = torch.tensor([
            [bounds_lon[0], bounds_lat[0]],
            [bounds_lon[1], bounds_lat[1]]
        ], dtype=torch.double)

    def fit(self, X_drilled: torch.Tensor, y_observed: torch.Tensor):
        self.gp = SingleTaskGP(X_drilled, y_observed)
        mll = ExactMarginalLogLikelihood(self.gp.likelihood, self.gp)
        fit_gpytorch_mll(mll)
        self.best_y = y_observed.max()

    def recommend_next(self, n_candidates: int = 5) -> list:
        acq = qExpectedImprovement(model=self.gp, best_f=self.best_y)
        candidates, acq_values = optimize_acqf(
            acq_function=acq,
            bounds=self.bounds.T,
            q=n_candidates,
            num_restarts=10,
            raw_samples=256,
        )
        return [
            {
                "rank": i + 1,
                "lon": c[0].item(),
                "lat": c[1].item(),
                "expected_improvement": float(v),
                "action_type": "drill",
                "reasoning": f"BoTorch EI acquisition: maximizes expected grade improvement given {len(self.gp.train_inputs[0])} prior observations."
            }
            for i, (c, v) in enumerate(zip(candidates, acq_values))
        ]
```

## Step 6: FastAPI ML Service

```python
# ml-service/main.py
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
import torch

app = FastAPI(title="Terranox ML Service")
MODEL_STORE: dict = {}

class NextActionRequest(BaseModel):
    project_id: str
    drilled_locations: list[list[float]]  # [[lon, lat], ...]
    observed_grades: list[float]

@app.post("/next-action")
async def recommend_next_action(req: NextActionRequest):
    from models.sequential_decision import ExplorationOptimizer
    X = torch.tensor(req.drilled_locations, dtype=torch.double)
    y = torch.tensor(req.observed_grades, dtype=torch.double).unsqueeze(-1)
    bounds_lon = (X[:, 0].min().item() - 0.5, X[:, 0].max().item() + 0.5)
    bounds_lat = (X[:, 1].min().item() - 0.5, X[:, 1].max().item() + 0.5)
    optimizer = ExplorationOptimizer(bounds_lon, bounds_lat)
    optimizer.fit(X, y)
    return {"recommendations": optimizer.recommend_next(n_candidates=5)}

@app.post("/train/{project_id}")
async def train_model(project_id: str):
    # Fetch drill data from Supabase, run training, store model
    return {"status": "training_started", "project_id": project_id}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)
```

## Step 7: UI - Prospectivity Map Viewer

```tsx
// src/components/Map/ProspectivityLayer.tsx
'use client';
import Map, { Source, Layer } from 'react-map-gl';
import type { RasterLayer } from 'mapbox-gl';

interface DrillHole {
  lon: number;
  lat: number;
  result: 'mineralized' | 'barren' | 'anomalous';
  grade?: number;
}

interface ProspectivityLayerProps {
  heatmapUrl: string;
  bounds: [number, number, number, number];
  drillHoles: DrillHole[];
}

const prospectivityLayer: RasterLayer = {
  id: 'prospectivity-heatmap',
  type: 'raster',
  source: 'prospectivity',
  paint: {
    'raster-opacity': 0.7,
    'raster-color': [
      'interpolate', ['linear'], ['raster-value'],
      0, 'rgba(0,0,128,0)',
      0.3, 'rgba(0,128,255,0.5)',
      0.6, 'rgba(255,200,0,0.8)',
      1.0, 'rgba(255,0,0,1)',
    ],
  },
};

export function ProspectivityLayer({ heatmapUrl, bounds, drillHoles }: ProspectivityLayerProps) {
  return (
    <Map
      mapboxAccessToken={process.env.NEXT_PUBLIC_MAPBOX_TOKEN}
      initialViewState={{
        longitude: (bounds[0] + bounds[2]) / 2,
        latitude: (bounds[1] + bounds[3]) / 2,
        zoom: 8,
      }}
      style={{ width: '100%', height: '600px' }}
      mapStyle="mapbox://styles/mapbox/dark-v11"
    >
      <Source id="prospectivity" type="raster" url={heatmapUrl} tileSize={256}>
        <Layer {...prospectivityLayer} />
      </Source>
    </Map>
  );
}
```

## Step 8: Deploy

```bash
# ML service on Railway
railway init && railway up  # from ml-service/

# Required Railway env vars:
# SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY

# Frontend on Vercel
vercel

# Required Vercel env vars:
NEXT_PUBLIC_SUPABASE_URL=https://xxx.supabase.co
NEXT_PUBLIC_SUPABASE_ANON_KEY=...
SUPABASE_SERVICE_ROLE_KEY=...
NEXT_PUBLIC_MAPBOX_TOKEN=pk.eyJ1...
ML_SERVICE_URL=https://your-ml-service.railway.app

# Enable PostGIS in Supabase SQL editor:
# CREATE EXTENSION IF NOT EXISTS postgis;
```

## Key Insights
- **The data moat is real.** 70 years of labeled drill outcomes is the product — the ML is the mechanism. Invest more in data acquisition and normalization than in model architecture.
- **Class imbalance is severe.** Uranium hit rates below 1% means your positive class is tiny. Always use weighted loss functions and evaluate on AUC-ROC, not accuracy.
- **Spatial autocorrelation breaks standard train/test splits.** Points near each other are correlated. Use spatial cross-validation (leave-one-region-out) to get honest performance estimates.
- **The CNN patch size matters enormously.** A 32x32 pixel patch at 25m/px resolution captures an 800x800m area — the scale at which uranium structural controls typically operate. Tune this to your target geology.

## Gotchas
- **GDAL installation is painful on Windows/Mac.** Use conda: `conda install -c conda-forge gdal`. Docker is most reliable for production.
- **CRS mismatches will silently corrupt everything.** Always reproject all layers to one CRS (WGS84 EPSG:4326 or a local UTM zone) before any spatial operations. Never assume incoming data shares your projection.
- **BoTorch requires double precision.** All tensors must be `torch.double` (float64), not default `torch.float` (float32). You will get cryptic errors otherwise.
- **Raster memory blows up fast.** A 1000x1000 pixel study area with 10 layers is 40MB. At 10,000x10,000 (a realistic regional study) it is 4GB. Use windowed reads with rasterio for anything above 2000x2000 pixels.