Machine Learning-Driven Pyrolysis Optimization: Capturing Hidden Feedstock Effects on Biochar and Bio-Oil Yields

Biomass pyrolysis in fixed-bed reactors converts diverse
wastes
into biochar and bio-oil, yet predictive models rarely integrate feedstock
heterogeneity into process optimization. Here, we compile 1691 experimental
runs spanning 69 feedstock classes and 15 physicochemical and operating
variables to train a Gaussian process regression (GPR) meta-learner
that stacks 4 primary machine learning models. The integrated framework
explicitly captures feedstock-driven variability and achieves cross-validated R2 values of 0.86 (biochar) and 0.87 (bio-oil)
while providing uncertainty estimates. SHAP analyses identify feedstock
type, temperature, and ash as dominant drivers, and partial-dependence
trends align with thermochemical expectations. A constrained genetic
algorithm then delivers feedstock-specific operating windows, yielding
feasible, real-world optimizations: cotton stalk maximizes biochar
yield (predicted 91.48%) and pure cellulose maximizes bio-oil yield
(65.60%). This study introduces a novel integration of feedstock heterogeneity
with ML-driven optimization and establishes a scalable, data-driven
workflow for feedstock selection, reactor design, control, and industrial-scale
deployment of the biomass pyrolysis process.