Machine learning-assisted optimization of green-synthesized CuO nanoparticle Yield and particle size

Description

The Matlab code presented here is based on green synthesized CuO nanoparicle with varying CUSO4 salt concentration, pH, and plant extract. The BBD design data was augmented with Gaussian noise, ANN model developed, and multi-objective optimization of process responses. SHAP, PDP analysis of the model is reported.

Steps to reproduce

1. The end-to-end workflow begins with statistical benchmarking via "a2_Fit_lm_model.m", which fits standard linear, interaction, and quadratic response surface models to experimental Box-Behnken Design (BBD) data to evaluate baseline metrics such as ordinary R2, AIC, and cross-validated R2. 2. Machine learning modeling is then initiated in b1_Bayesian_fitrnet_normalized.m, where the real dataset is augmented with controlled noise-injected synthetic data across varying sample folds and optimized for artificial neural network (ANN) architectures using Bayesian hyperparameter optimization. 3. Following training, b3_modelscatter.m screens the optimization results to select the top-performing networks based on cross-validated fitness, generating validation scatter plots and standardized residual diagnostics. 4. Data fidelity between the original experimental design and the augmented subsets is subsequently verified via b5_boxplot_dataquality.m by comparing raw and standardized feature distributions. 5. The selected neural networks are then transferred to b5_min_max_GA.m to execute single-objective optimization through a Genetic Algorithm (GA), identifying the extreme upper and lower boundaries for each individual response variable. These boundary constraints serve as inputs for b8_Multi_Desire.m, which performs GA-driven multi-objective optimization across predefined weighting vectors by maximizing a composite desirability function. 6. Finally, explainable artificial intelligence (XAI) routines are deployed to unpack model behavior: c1_LIME_Shaply.m calculates local and global predictor contributions using LIME and Shapley algorithms, c3_Shaply_finalplot.m visualizes global feature rankings and swarm attribution plots, and c4_PDP_plots.m generates partial dependence plots to map the marginal effects of each input factor across all modeled responses.

Institutions

• Karunya University

  Tamil Nadu, Coimbatore

Categories

Licence