NCKH – Pilot Scale Physics-Informed ML
Hybrid PINN (Physics-Informed Neural Network) Forecast Dashboard
AI model combining physical laws with neural networks for physics-aware aquaculture predictions
Air Temperature31,03°C
Air Humidity71,65%
Wind Speed13,82 km/h
Rainfall0,26 mm
ConditionsSunny intervals
Auto API
Air Temperature31,03 °C
Air Humidity71,65 %
Wind Speed13,82 km/h
Rainfall0,26 mm
ConditionsSunny intervals
ESP32 + BME280 + BH1750 + Rain gauge
28,33
7,97
6,7
4,08
31,03
71,65
13,82
0,26
Forecast Window:
━━ Observed   ╌╌ Hybrid PINN (Physics-Informed Neural Network)
Water Temperature (°C)
PH Parameter
Dissolved Oxygen (mg/L)
Turbidity (NTU)
Salinity (ppt)
TDS (mg/L)
ORP (mV)
25,1
7,52
7,51
2,56
31,03
71,65
13,82
0,26
Forecast Window:
━━ Observed   ╌╌ Hybrid PINN (Physics-Informed Neural Network)
Water Temperature (°C)
PH Parameter
Dissolved Oxygen (mg/L)
Turbidity (NTU)
Salinity (ppt)
TDS (mg/L)
ORP (mV)
0.118
0.165
0.962
1.82
Training Loss vs Validation Loss MSE per epoch
MAE Convergence by Epoch
R² Score theo Epoch
Predicted vs Actual – Scatter (Temperature °C)
Physics Residual Loss (PDE constraint) ∂C/∂t – D·∇²C + k·C
Loss Breakdown: L_data / L_physics / L_bc
Model Architecture – Hybrid PINN (Physics-Informed Neural Network)
Architecture Hybrid PINN
  • Trunk net: 4 × Dense(64), tanh activation
  • Physics constraint: oxygen diffusion PDE
  • Loss = λ₁·L_data + λ₂·L_physics + λ₃·L_bc
  • λ₁=1.0 | λ₂=0.1 | λ₃=0.05
  • Optimizer: Adam warm-up (200) + L-BFGS (300)
  • Total epochs: 500
Physical Constraints
  • PDE: ∂C/∂t = D·∇²C − k·C + S(t)
  • Temperature-dependent D (Arrhenius)
  • Boundary: surface oxygen concentration (saturation)
  • Thermal balance: Newton's cooling law
  • pH buffer: carbonate system