README.md

metadata

license: cc-by-4.0
task_categories:
  - tabular-regression
language:
  - en
tags:
  - chemistry
  - chemical-engineering
  - biodiesel
  - transesterification
  - process-optimization
  - reinforcement-learning
  - surrogate-modeling
  - tabular
pretty_name: Two-Stage Transesterification Operating-Parameter Dataset
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files: Dataset.csv

Two-Stage Transesterification Operating-Parameter Dataset (21,001 records)

A tabular dataset mapping six manipulated operating parameters of a two-stage alkali-catalysed transesterification process to the resulting biodiesel conversion rate (%). It was built as a sensitivity-analysis sweep over the process operating envelope and is intended as a benchmark for surrogate modelling, static optimisation (RSM / Bayesian optimisation), and reinforcement learning for continuous process control.

This is the dataset accompanying the paper "A Hybrid LSTM-DDPG Framework for Operating-Parameter Optimisation of a Two-Stage Transesterification Process."

Dataset structure

21,001 rows × 7 columns. No missing values, no duplicate rows.

Column	Description	Unit	Min	Max	Levels
MOH/Oil ratio-1	Methanol-to-oil molar ratio, stage 1	mol/mol	7.0	9.0	4
Temp-1 (C)	Reaction temperature, stage 1	°C	80	100	5
R. Time-1 (hr)	Reaction time, stage 1	h	3.5	8.0	7
MOH/Oil ratio-2	Methanol-to-oil molar ratio, stage 2	mol/mol	3.035	4.0	5 (+1)
Temp-2 (C)	Reaction temperature, stage 2	°C	40	80	6 (+1)
R. Time-2 (hr)	Reaction time, stage 2	h	0.5	1.5	5 (+1)
CONV, y (%)	Target — biodiesel conversion rate	%	81.95	99.58	continuous

Design of experiments

The first 21,000 rows form a complete full-factorial grid: 4 × 5 × 7 = 140 stage-1 settings crossed with 5 × 6 × 5 = 150 stage-2 settings. Every stage-1 combination appears with all 150 stage-2 combinations, so the design is fully balanced and orthogonal.

Row 21,001 is a single off-grid point (MeOH/oil-2 = 3.835, T2 = 75 °C, t2 = 1.055 h, conversion = 99.26%) appended outside the factorial lattice. Users who require a strictly rectangular design should drop it; users benchmarking against the source paper should keep it, since all reported statistics use the full 21,001 records.

Target distribution

Mean conversion is 95.73%, with the third quartile at 98.35%. The response surface is strongly monotone in the stage-1 alcohol-to-oil ratio, which dominates the variance; stage-2 variables have comparatively narrow optimal windows. The maximum conversion of 99.583% occurs at the upper corner of the envelope (9.0 mol/mol, 100 °C, 8.0 h / 4.0 mol/mol, 80 °C, 1.5 h) — this value acts as a hard ceiling for any surrogate-based optimiser trained on the data.

Usage

import pandas as pd

df = pd.read_csv("Dataset.csv")
X = df.iloc[:, :6].values   # six manipulated variables
y = df.iloc[:, 6].values    # conversion rate (%)

To reproduce the split used in the paper: remove outliers with a 3 × IQR filter, then split 70/15/15 with random_state=42, fitting StandardScaler on the training partition only.

Intended uses

• Regression / soft-sensor benchmarking on a smooth, low-dimensional, noise-free response surface.
• Static optimisation baselines (response surface methodology, Bayesian optimisation, genetic algorithms).
• A nearest-neighbour or learned surrogate environment for reinforcement learning in continuous action spaces, avoiding the cost of physical reactor interaction.
• Ablation studies comparing one-shot optimisation against sequential decision-making under simulated constraint changes and feedstock drift.

Limitations and caveats

• Simulation-derived, not experimental. Values come from a sensitivity analysis of the process, not from wet-lab measurement. Absolute conversion figures should not be treated as validated plant data.
• Noise-free and deterministic. Each parameter combination maps to exactly one conversion value. Models trained here will not encounter the measurement noise, catalyst deactivation, or feedstock variability of a real reactor.
• Single feedstock. The grid varies operating conditions only; feedstock composition, free-fatty-acid content, and catalyst loading are held fixed.
• Extrapolation is unsupported. Nearest-neighbour or interpolating surrogates are trustworthy only inside the convex hull of the grid. Optimisers cannot discover conversions above 99.583% and any such result is an artefact.
• Coarse grid spacing. With only 4–7 levels per variable, gradients near the optimum are resolved coarsely.

Citation

@inproceedings{butarbutar2026hybrid,
  title     = {A Hybrid {LSTM-DDPG} Framework for Operating-Parameter
               Optimisation of a Two-Stage Transesterification Process},
  author    = {Butarbutar, Chardinal Martin and
               Widada, Bartolomeus Priya Perkasa Utama and
               Putra, Cendra Devayana and
               Putri, Berliana Devianti and
               Taipabu, Ikhsan},
  booktitle = {TBA},
  year      = {2026}
}

Contact

Cendra Devayana Putra — putracendra@unesa.ac.id (Surabaya State University)