The emission of CO₂ into the atmosphere from the burning of fossil fuels has contributed to climate change. Atmospheric CO₂ concentration reached 419 parts per million in 2021 (Rees and Columbia, 2021). Carbon dioxide removal from the atmosphere has been proposed to keep global temperature rise within 1.5°C (IPCC, 2021). Production of bioenergy with carbon capture and storage (BECCS) is one approach to carbon removal. This can be accomplished by growing lignocellulosic crops and converting them to useful bioproducts while capturing the CO₂ emitted (Hickman, 2016; Rees and Columbia, 2021). Production and land sequestration of biochar via pyrolysis of biomass is another promising approach to carbon removal (Laird et al., 2009). The sustainability of biomass pyrolysis systems depends on feedstock availability and economic and life cycle performance specific to enterprise location and technology employed (Li et al., 2019).

According to the Billion-Ton Study, lignocellulosic resources, comprising mainly agricultural and forest residues, are presently about 365 million dry tons and are expected to rise to around 826 million dry tons by 2040 (Langholtz et al., 2016). Lignocellulosic biomass like corn stover (Ganguly et al., 2020), red oak (Elliott et al., 2015), soy hulls (Oliveira et al., 2015), switchgrass (Mullen and Boateng, 2008), etc. have been widely studied for thermochemical processing. Techno-economic assessment (TEA) of biomass-fed fast pyrolysis systems has demonstrated their economic and environmental viability. Wright et al. (2010) reported that naphtha and diesel range fuel could be produced in a 2,000 metric ton (MT)/day corn stover fed biorefinery at a minimum fuel selling price (MFSP) of $2.11/gallons. Hu et al. (2016) investigated the economic feasibility of a red oak-fed biorefinery producing biofuels and biochemicals. The study estimated a minimum selling price of $3.09/gallon and $433.7/MT for biofuels and biochemicals, respectively. It was observed that the co-production of biochemicals and biofuels was more profitable than the production of biofuels alone. In a study by Meyer et al. (2020) on the economics and life cycle performance of plants pyrolyzing woody and herbaceous biomass (six pure feedstocks—tulip, pine, corn stover, switchgrass, etc.), calculated MFSP of between $3.7/gge to $5.1/gge. They also observed that woody biomass produced more biofuel than herbaceous feedstock. LCA results of this study observed GHG reductions of around 72% compared to petroleum-based systems with the GHG emissions depending mostly on the feedstock type and other process parameters.

Scientists have investigated the impacts of feedstock properties on the bio-oil yield (Li et al., 2016), which affects the economic feasibility Ou et al. (2018). Li et al. (2017) reported that biomass ash content varying between 0.3 and 7.7 wt.% decreased the biofuel yield from 87.3 to 40.7 gallons/ton and increased the corresponding MFSP from $2.3/gallons to $4.8/gallons. Ou et al. (2018) developed a fast pyrolysis process model sensitive to feedstock characteristics obtained from experimentations. It was found that increasing the ash content from 1 to 7 % increased the MFSP from $0.97/liter to $1.06/liter due to lower biofuel yields. Although these findings show the importance of feedstock properties on biofuel production, few studies have explored the economics and environmental performance of pyrolysis applied to carbon removal.

The versatility of pyrolysis systems allows for the production of various bioproducts including bioasphalt (Peralta et al., 2012), sugars (Rover et al., 2014), and levoglucosan (Rover et al., 2019). Levoglucosan-rich pyrolytic sugars have gained attention due to their wide range of applications, like the fermentation of pyrolytic sugars to produce ethanol (Ganguly et al., 2020), biodegradable plastics, and pharmaceuticals (Junior et al., 2020). Rover et al. (2019) observed levoglucosan crystal production, along with phenolic compounds and biochar via autothermal pyrolysis (ATP) of red oak at a minimum selling price of $1,333/ton, which is much lesser than the market price. Autothermal pyrolysis has advantages of process simplification and intensification compared to conventional fast pyrolysis (Polin et al., 2019). LCA of an integrated fast pyrolysis-anaerobic digestion-fermentation plant producing ethanol, phenolic compounds, and power observed greenhouse gas emission of −16.6 g CO_2e/MJ of ethanol (Ganguly et al., 2020). Other studies evaluated the simultaneous production of multiple products in integrated pyrolysis biorefineries (Dang et al., 2016; Hu et al., 2016).

Researchers have mostly employed regression and interpolation-based models on laboratory experiments to further predict pyrolysis output (Gronli and Melaaen, 2000; Neves et al., 2011; Li et al., 2017). Hansen and Mirkouei (2018) noted that typical mathematical models are not able to capture the complexities of biomass pyrolysis and bio-oil upgrading and suggested the adoption of machine intelligence techniques to overcome these shortcomings. They compared three types of models: a cyber physical system, an artificial neural network, and a “k-Nearest Neighbor” model of pyrolysis-based systems. They found the “k-Nearest Neighbor” model to be the most promising. The “k-Nearest Neighbor” successfully classified 214 biomass datasets obtained from various published resources with an accuracy of around 71% (Olatunji et al., 2020). To the best of our knowledge, machine learning has not been previously used to generate biochemical properties data sets for biomass feedstocks. This study introduces a novel approach to generating the biochemical composition of lignocellulosic biomass samples and demonstrates its application to economic and environmental technology assessment with the case study of a carbon removal pyrolysis system. The objective of this paper is to elucidate relationships between feedstock properties and the economic and environmental performance of an ATP biorefinery that produces sugars, phenolic oil, and biochar.

This study describes the application of Machine Learning (ML) models to process modeling, TEA, and LCA assessment of biomass ATP to produce sugars, phenolic oil, and biochar. Two ML models are compared for the generation of datasets containing 3,000 lignocellulosic feedstock samples with unique cellulose, hemicellulose, and lignin composition. An Aspen Plus™ chemical process model predicts the mass and energy balance of a 250 tons per day ATP system. Process simulations are conducted using Aspen Simulation Workbook to gather simulation results based on the 3,000 feedstock samples. A range of minimum sugar selling prices (MSSP), lifecycle GHG emissions, and carbon removal rates are estimated from the simulation results. The results are interpreted using regression and sensitivity analyses to understand the influence of various parameters on the MSSP and environmental footprint. The MSSP values are compared to historical prices to evaluate the economic feasibility of the ATP to sugars platform.

The GAN ML model generated 3,000 samples of feedstock biochemical data. Figures 2, 3 show the similarity of the GAN ML samples to the original input data set through Van Krevelen-type diagrams. These diagrams describe the relationship between each sample's hemicellulose/cellulose (HC) and lignin/cellulose (LC) mass ratios. These diagrams facilitate visual inspection of the data to support the statistical metrics.

Figure 2 includes the 400 input data's biomass type classification: untreated wood, grass/plant, straw, husk/shell/pit, organic residue, treated wood, and INL samples. As shown, there are a few clusters, including most grass plants having LC ratios of <0.75 and most untreated wood having HC ratios of <1. There are several outliers in the top right corner with high HC and LC ratios due to the samples' low cellulose content.

Figure 3 shows a Van Krevelen-type diagram of the 3,000 GAN-generated data samples, compared to Figure 2. The HC and LC values of the GAN data are closer to the input data than the KDE data. GAN generated fewer outliers and more samples around the clusters in the original data. Henceforth, the result section focuses on the GAN results. KDE results are included in Supplementary Material.

We predicted biomass pyrolysis yields for the 3,000 ML samples using the Kriging-ML model of the ATP system. Figure 4 shows the ATP product yield relationships to cellulose content. We observe that sugar, phenolic oil, and biochar yields have an approximately linear relationship with cellulose content. Sugar and phenolic oil yields increased with cellulose content, but biochar yield decreased. This finding agrees with experimental observations where cellulose pyrolysis produces more pyrolysis oil than lignin, which favors biochar production. Cellulose decomposes at a temperature between 315 and 400°C, which is lower than pyrolytic temperature (typically 500°C), and quickly converts to pyrolytic vapors (Yang et al., 2007; Lee et al., 2013).

As shown in Figure 5, sugar and phenolic oil yields exhibit an exponential decay with increasing hemicellulose/cellulose (top) and lignin/cellulose ratios (bottom) while biochar yields increase in a logarithmic manner. The machine learning yield-based trends were similar to those reported in other studies (Zhou et al., 2018; Tomczyk et al., 2020). Unlike the hemicellulose/cellulose ratio, the biochar yield had an increasing pattern at a high lignin/cellulose ratio. Higher lignin content in a feedstock induces char production (Demirbas, 2004; Shariff et al., 2016; Tomczyk et al., 2020). These results do not account for ash content, reaction temperature changes, differences in particle mechanical properties, and others that might affect ATP yield.

Figure 6 shows the MSSP obtained from evaluating the 3,000 plus sample data gathered from the GAN feedstock database. The ATP system incurs a total fixed capital investment of $16 MM. We compared the generated values of MSSP with the historical price of sugar obtained from the U.S. Department of Agriculture. The blue dashed line is the current price of sugar which is around $369/MT (Hunt, 2021), and the red dashed-dot line represents the average sugar price over the last 5 years. The result showed that the MSSP generated using the GAN model was expected to be lower than the historical sugar price over the last 5 years.

Figure 7 shows the probability density of GHG emissions obtained from the GAN values expressed in kg CO_2e/kg of lignocellulosic biomass. The GHG emissions varied between −0.56 and −0.74 kg CO_2e/kg lignocellulosic biomass. It was observed that all the values obtained show negative emissions. Based on the global market demand for sugars (Statistica, n.d.), phenolic oil (asphalt) (Freedonia Group, World Asphalt (Bitumen), 2020), and biochar (Grand View Rearch, 2019), ATP systems could achieve a potential removal of 0.0006 to 0.8 billion-tons CO₂ per year if these markets are completely saturated by ATP products. This supports the classification of lignocellulosic biomass ATP plant producing sugar, phenolic oil, and biochar as a carbon-negative system.

There are several techno-economic and lifecycle assessment assumptions that could vary significantly depending on the location of the facility. For example, electricity costs and emissions vary based on the predominant resource available. Switching from the US grid mixture to wind power generation could lower costs from $185 to $183/MT of sugar and emissions from −0.65 to −0.67 kg CO_2e/kg corn stover. A detailed analysis of regional factors is beyond the scope of this study.

This study investigated the impact of feedstock variability on a 250 ton/day lignocellulosic biomass-driven autothermal pyrolysis biorefinery producing sugar, phenolic oil, and biochar. This study utilized GAN and KDE models for generating biochemical feedstock data samples based on real samples obtained from Phyllis 2 and INL. We compared the GAN and KDE biochemical composition to the real samples. GAN-generated samples showed a closer agreement with the input data. Pyrolysis yield predictions based on the GAN samples showed a linear relationship between the biomass cellulose content and the sugar and phenolic oil yields. Biochar yield showed a logarithmic relationship with an increase in the lignin content of biomass feedstock. The MSSP and GHG emissions from the GAN generated model varied between $66/MT and $280/MT and −0.56 kg CO_2e/kg and −0.74 kg CO_2e/kg lignocellulosic biomass used for the ATP system, respectively. These findings suggest that lignocellulosic biomass autothermal pyrolysis produces pyrolytic sugars, phenolic oil, and biochar is an economic and sustainable carbon-negative system.

However, this study does not explicitly consider the underlying uncertainty in the experimental data. We also assume that ash content does not play a major factor in the output yields, and the pyrolytic reactor operates at a uniform temperature. Additionally, we also assume that all particles have relatively similar characteristics and do not consider the impact of particle variability on the output yield.

Future studies will investigate the impact of facility location on costs and emissions due to regional factors like local feedstock, fuel, and electricity resources. We are also interested in comparing the ATP carbon-negative platform to alternative carbon capture and sequestration systems. The machine learning techniques from this study could be applied to other systems to evaluate their potential for improving sustainability assessments. ML could also be used to directly predict the sugar, phenolic oil, and biochar output from the 35 pyrolysis intermediate products instead of using Aspen Plus. Machine learning models could also be applied to look into how other experimental factors like reactor temperature changes mechanical particle properties affect the product yield and, in turn affect the economical and environmental aspects of the considered system.

The original contributions presented in the study are included in the article/Supplementary Materials, further inquiries can be directed to the corresponding author/s.

This study's supervision and concept building were provided by MW and RB. Methodology and analysis of this study were done by AG and MW. Writing and editing were performed by AG, MW, and RB. All authors contributed to the article and approved the submitted version.

This research was funded in part by the American Institute of Chemical Engineers as part of the Department of Energy sponsored RAPID Institute [subcontract #138846].

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.