Modeling and active constrained optimization of C5/C6 isomerization via Artificial Neural Networks

This repository contains the code, datasets, and results published in our paper that was published in the Chemical Engineering Research and Design journal.

Note on Data: While we wait for UniSim approval, we have encrypted the dataset files r490.csv and datasets/*.

Scripts

• create_dataset.py: compiles raw UniSim result Excel input sheets into one comma delimited file, such as r490.csv.
• split_train_test_repeated.py: repeatedly splits the dataset into 10 random splits that are sampled via Latin Hypercube Sampling. Generates a split file that indicates which samples belong to the training or testing sets, such as r490_splits.npy.
• evaluate_model.py: trains and evaluates a given model configuration on a given dataset, on all given splits of the dataset.
• nn.py: artificial neural network model (ANN)
• null_model.py: a model that just predicts average proportions on the training set.
• rbfm.py: radial basis function (RBF) model
• rm.py: basis function linear regression model that can be configured as a linear regression model or, as shown in the paper, a quadratic nonlinear regression model.
• opt_mp.py: runs active experimentation simulations on a given dataset, using the given experimental configuration (e.g., noiseless, noisy, restricted transitions) and computational model/acquisition function configuration.
• pipeline_evaluation.py: evaluates all models, as reported in the paper.
• pipeline_run_active_exp.py: runs all active experiment simulations, as reported in the paper.
• fit_models_to_traces.py: for each active experiment trace in the given trace file (one of the files in traces/r490), this fits an ANN model to all the input/output pairs collected during that active experiment and then uses the trained model to make predictions over all 10,000 combinations of reactor temperatures and pressures.

Directories

• cfgs/models/*.json: configurations of models evaluated in Figure 5.
• cfgs/*.json: configurations of acquisition functions evaluated in Figure 10.
• datasets/r490: raw UniSim excel sheets containing the outputs of 10,000 simulations covering different combinations of reactor temperatures and pressures.
• exp_cfgs: active experiment configurations in Figure 10.
• r490_results: results of model evaluation in Figure 5.
• traces/r490: active experiment results (traces) for all combinations of experiments and acqusition functions.

Analysis Notebooks

• analysis_r1_temp.ipynb: plots Figures 2 and 3.

• analysis_product_correlations.ipynb: plots Figure 4.

• analysis_ann_results.ipynb: plots Figure 5.

• analysis_temp_vs_kpi_heatmaps.ipynb: plots Figure 6.

• analysis_constraint_ch_opt.ipynb: plots Figure 7.

• analysis_nn_uncertainty.ipynb: plots Figure 9.

• analysis_active_exp_traces.ipynb: plots Figure 10.

• analysis_input_kpi_correlation.ipynb: generates Table 1.

• analysis_global_optima.ipynb: generates Table 2.

• analysis_compare_surrogate_model_to_actual.ipynb: generates Table 3.

Requirements

The code has been tested on:

• Python 3.8.10
• tensorflow 2.8.0
• scikit-optimize 0.9.0
• pandas 1.3.4
• numpy 1.22.3
• sklearn 1.0.1
• xmltodict 0.12.0
• PIL 8.4.0
• matplotlib 3.4.3
• pingouin 0.5.1
• scipy 1.7.1
• seaborn 0.11.2