PHA-ANN Predictor — Bioprocess Simulation Tool

Neural-network surrogate for polyhydroxyalkanoate fermentation. Predicts time-resolved biomass (C_x) and PHA (C_p) concentration from initial substrate loading. Trained on 5,139 samples (5,000 synthetic trajectories + 139 literature points across 12 published datasets).

ModelMLP · dual-head

R² C_x0.9969

R² C_p0.9922

READY

1User Input

2Prediction

3Time Course

4Architecture

5Validation

Biomass C_x at selected t

—g/L

PHA C_p at selected t

—g/L

Peak C_p

—g/L

Y_p/s (PHA / sugar consumed)

—g/g

Process parameters [01]

Initial substrate

Start C_s (carbon source)g/L 10.00

Start C_n (nitrogen source)g/L 2.50

Evaluation

Time pointhr

Preset regimes

All species

Biomass

PHA

Substrates

ANN forward pass · 201 points · domain 0–50 hr —

Prediction summary [02]

Biomass C_x —g/L

PHA C_p —g/L

Residual C_s —g/L

Residual C_n —g/L

Derived metrics

PHA / biomass—

Peak C_p—

Peak C_x—

Time to C_s depletion—

Network architecture [04]

Dual independently-trained multilayer perceptrons sharing 5 inputs (Time, Start C_s, Start C_n, C_s, C_n). MinMaxScaler [0, 1] applied to both inputs and outputs. Early stopping on 12% held-out validation fold. Random seed 42.

This build runs the trained MLP weights directly in the browser via a pure-JS forward pass. Biomass (C_x) and PHA (C_p) come from the two primary networks documented in the Model Summary; substrate trajectories C_s(t) and C_n(t) come from two auxiliary MLPs trained alongside on (Time, Start C_s, Start C_n) → current substrate. All four networks are embedded inline.

Validation & test-set performance [05]

Metric	C_x (biomass)	C_p (PHA)
R² — overall test	0.9969	0.9922
R² — synthetic only	0.9997	0.9994
R² — literature only	0.9223	0.9512
RMSE	0.2506 g/L	0.1480 g/L
MAE	0.0723 g/L	0.0357 g/L
MAPE	4.27 %	11.94 %
NRMSE	1.64 %	1.06 %
Test samples	1,028 (1,000 synth + 28 lit)
Training samples	4,111 (4,000 synth + 111 lit)