📊 Results
🎯 Model Accuracy
🧠 ANN Architecture
🌐 3D Surface
📈 Sensitivity
📐 MATLAB/RSM
⚗️ Isotherms
🧪
Set parameters & click Predict
5 models will run in parallel — RSM, Random Forest, XGBoost, GPR, ANN (deep learning)
Best model is highlighted automatically based on validation R²
Best model is highlighted automatically based on validation R²
🏆 Prediction Summary
BEST MODEL
—
EXTRACTION EFF.
—
%
DIST. COEFF. KD
—
LOADING Z
—
MODEL COMPARISON — Extraction Efficiency E%
📊 All Targets — Model Comparison
📖 How to read: Each group of bars = one output variable (KD, E%, Z, SF). Each colored bar = one ML model. Taller bar = higher predicted value. Compare bar heights to see which model gives the highest or most consistent prediction for your input.
🎯 Model Accuracy (R² & RMSE) on Test Set
Trained on 240 samples derived from paper RSM equations (Yıldız et al., 2023).
ANN is the only deep-learning model; RF, XGBoost, GPR, RSM are classical ML.
High R² = good fit. Low RMSE = low prediction error.
📖 How to read accuracy charts: R² (top chart): closer to 1.0 = better fit — the model explains more variance. RMSE (bottom chart): lower = smaller average prediction error. GPR typically gives highest R² for E% because it captures non-linear patterns. ANN needs more data to match GPR performance.
Loading metrics…
🧠 Neural Network Architecture
Deep feedforward ANN with ELU activations, BatchNormalization,
Dropout regularisation, and L2 weight decay.
Trained with Adam optimiser + EarlyStopping (patience=60) on
240 paper-derived samples. ANN is the only DL model
(all others are classical ML).
⚠️ Dataset & ANN Notes
✅ Dataset source: 240 rows generated from paper RSM equations (Yıldız et al. 2023) — formic, acetic, propionic acid × CCD design + dense grid
✅ Balanced: Equal samples per acid type (80 each); continuous coverage from low to high efficiency regions
✅ ANN vs dataset size: Neural networks normally need large datasets. With only 240 rows, ANN may underperform vs RF/XGBoost. R² is shown on the Accuracy tab after training.
✅ Training callbacks: EarlyStopping + ReduceLROnPlateau prevent overfitting
✅ Recommendation: If ANN R² < 0.92, prefer GPR or XGBoost for predictions
🌐 3D Response Surface (Paper RSM Equations)
Direct implementation of the quadratic RSM equations from Yıldız et al. (2023).
Equivalent to the 3D plots in the paper (Figs. 3–5).
📖 How to read the 3D surface: The X and Y axes are the two input variables you selected. The Z axis (height) = Extraction Efficiency E%. Higher peaks = better extraction. Rotate by clicking & dragging. Bright green/teal = high efficiency (>90%), dark blue = low efficiency. The peak region shows the optimum operating conditions.
🌡️ Heatmap — ML Model Prediction Grid
📖 How to read the heatmap: Each cell = a predicted output value at that X/Y combination. Brighter yellow = higher value, darker purple = lower value. Look for the brightest region — that is the optimal operating zone. Use this to quickly find which combination of two variables gives the best result.
📈 Sensitivity Analysis — All Models
📖 How to read: The X axis sweeps one input variable across its full range while all others stay fixed. The Y axis shows the predicted output. Each line = one ML model. Steep rising lines = that variable strongly increases the output. Flat lines = variable has little effect. When all model lines agree → high confidence.
📐 MATLAB / RSM Modelling
Direct implementation of the Response Surface Methodology quadratic models
from Yıldız et al. (2023), originally computed using Design-Expert® software
(equivalent to MATLAB rsm/regress). R² ≥ 0.9985 for all three acids.
Propionic Acid
Acetic Acid
Formic Acid
RSM Quadratic Equation (Coded Variables)
Select an acid and click Compute…
ANOVA Table (from Paper)
🗺️ RSM Contour — E% Response Surface
📖 How to read the contour: Lines connect points of equal E%. Tightly packed lines = steep change (high sensitivity). Widely spaced = flat region. The top-right corner (high TOA + optimum HDES) typically shows the highest efficiency (dark yellow contour lines).
⚗️ Adsorption Isotherms
📖 How to read: The cyan dots are GPR model predictions at different acid concentrations (Ce). The green line is the Langmuir isotherm fit (assumes finite adsorption sites). The dashed purple line is the Freundlich fit (assumes heterogeneous surface). A good fit means the isotherm model describes the extraction equilibrium well.