Note: Modeling methods are still
experimental in nature.
Results should be interpreted with caution.
Analyze costs and life-cycle impacts.
Select an output molecule to get started:
Process Flow Diagram
Electricity Generated & Used On-site
Surplus Electricity Generated
Choose modeling options and input parameters.
Then click Run Model to visualize and
Features Under Development
Optimizing energy consumption for aerobic bioconversion, wastewater
treatment, onsite energy generation, and utilities stages.
Identifying promising avenues for reducing natural gas inputs to the
boiler, which is relatively less influential to the minimum selling
could largely reduce GHG emissions.
Integrating the recovery and separation decision tree, which allows
user to choose an economic recovery option for their products among
the several possible recovery and separation options.
Adding more functionality to biomass feedstock supply chain,
including soil organic carbon sequestration, field-to-biorefinery
supply distance, biomass yield, and feedstock forms at the biorefinery
gate such as silage, bales, and pellets.
Adding more biofuels
and bioproducts, including but not limited to methyl ketones, and
indigoidine (blue dye).
Updating carbon footprint of process chemicals assuming that those
will produce utilizing the Zero-Carbon Grid Electricity.
Improving accuracy of greenhouse gas footprint calculations for
simple sugar feedstocks.
Updating the water footprint of simple sugar feedstocks.
Incorporating water withdrawal footprint model.
Minimum Selling Price (MSP) Estimator
This tool provides two modeling options for
estimating the MSP of target molecules:
pyTEA and spLearn. Here, we describe both models in
greater detail and discuss the prediction error
associated with each.
pyTEA is a process-based simulation model designed to
of major supply and production stages for a given molecule.
While lacking the nuance and precision of
industry-grade software such as SuperPro Designer
or Aspen Plus, it aims to roughly emulate the methods of
these programs and generate comparable outputs
in a fast and lightweight manner.
In constrast to the mechanisic approach of pyTEA,
spLearn is a
approach that leverages
machine learning (ML). Specifically, we make use of
The Tree-Based Pipeline Optimization Tool
(TPOT), an Automated ML framework that searches and tests a broad
range of ML algorithms to ultimately identify a model
architecture and set of hyperparameters that achieves the
prediction error for a given dataset.
Oftentimes, multiple ML algorithms are combined using a technique
to obtain the best prediction accuracy.
To build and train spLearn, we ran TPOT over a
dataset of 8,000 simulation trials
of a SuperPro Designer model. For each trial, values of the
input parameters were drawn randomly from their respective
probability distributions. We treat the MSP values generated
SuperPro in these trials as the 'ground truth' as they are
derived from the most precise mechanistic method available.
Ultimately, it is the relationship between these SuperPro MSP
values ther corresponding input parameters
them that we are ultimately seeking to learn through
To evaluate and compare the relative performance of pyTEA and
spLearn we use each model to make predictions on a test
set of 2,000 trials (withheld from model training in
the case of spLearn). The scatterplots below show the
correlations between the SuperPro MSP values for each trial
and the corresponding pyTEA and spLearn predictions
for the same set of input parameter values. Both pyTEA and
spLearn yield stage-specific MSP predictions.
In the case of spLearn, this requires training and tuning
separate ML algorithms for each stage.
A greater clustering of points around the y = x line
(dashed red) in these plots indicates closer agreement between
predicted and actual values, and thus higher prediction accuracy.
Quantitatively, we rely on
Mean Absolute Error
(MAE) as our primary metric of model prediction error.
Conceptually, MAE simply measures the average difference between
predicted and ground truth values and therefore offers easily
interpretable error margins in the same units as the target
variable. As shown in the parity plots,
spLearn achieves a lower MAE for all stage-specific MSP
predictions. In the aggregate, the MAE of spLearn for
predicting total MSP is 0.43 $/gal while that of
pyTEA MAE of our pyTEA model stands at 1.18 $/gal.
Due to the lower error rate of
spLearn, we advise the use of this model for generating
MSP estimates. However, we note that both models perform with
very high accuracy and provide both modeling options in this
tool to grant users the ability to choose between models at their