From Electrons to Biorefineries: How Multiscale Modeling and Machine
Lignocellulosic biomass is the tough, fibrous material that makes up plant cell walls. It is widely seen as one of the most important renewable sources of carbon for replacing fossil fuels and fossil-based raw materials used to make fuels, chemicals, and materials. A detailed review argues that the main challenge is no longer finding possible ways to convert this plant material into useful products. Instead, the challenge is connecting knowledge across different scales, from the quantum mechanics of a single chemical bond to the costs and performance of an entire biorefinery. The review brings together computer-based methods and data-driven approaches for turning biomass into useful products. It shows that progress depends less on any one method and more on whether useful information can be passed between molecular models, reactors, process simulations, and studies of environmental sustainability.
The first challenge is the natural complexity of biomass. It is not simply a mixture of cellulose, hemicellulose, and lignin. Instead, these materials form a layered structure in which the polymers are held together by hydrogen bonds, interactions between water-repelling surfaces, and covalent chemical bonds. Cellulose forms stiff, tiny fibers called microfibrils. Dense networks of hydrogen bonds hold these fibers together and make it difficult for solvents to enter them. Lignin forms a chemically varied structure containing aromatic compounds that protects the carbohydrates. Recent structural studies discussed in the review show that lignin can organize itself into tiny water-repelling regions called nanodomains. They also show that the connections between lignin and carbohydrates change as plant cell walls mature. These findings challenge the older view that plant cell walls are simply fixed mixtures of polymers.
Numerical examples show why simple measurements and descriptions are not enough. In one study of the chemical bonds linking lignin and carbohydrates, researchers used ab initio molecular dynamics, metadynamics, and density functional theory to study the energy needed for cellulose decomposition. The calculated energy barriers increased from about 61–68 kilocalories per mole for cellobiose, a small molecule made of two linked sugar units, to about 108–117 kilocalories per mole in models where the polymers were cross-linked. This is roughly a twofold increase. The energy changes during the reactions also increased by three to four times. In other words, the chemical bonds linking these polymers do more than make the material harder to reach. They fundamentally change the energy needed to break it down. The review also explains why measurements of crystallinity, or how regularly molecules are arranged, can be misleading. In a study of enzyme-driven breakdown, researchers examined Avicel, bleached softwood, and bacterial cellulose. The measured crystallinity index of bleached softwood rose from about 67 percent to 80 percent during the first 12 hours, even though about 50 percent of its carbohydrates had already been converted. This happened because the process removed less-ordered material, changing the measurement without meaning that the underlying chemistry had changed in the same way.
To manage this complexity, the review describes a range of computer-based methods, each with a specific purpose and limitations. Density functional theory can examine how chemical bonds are activated, how strongly molecules stick to surfaces, and how reactions move through intermediate states. It is useful for reactions such as breaking lignin apart with hydrogen and removing oxygen from molecules using hydrogen. However, it is usually limited to small model molecules and simplified surfaces. Classical molecular dynamics can show how solvents interact with biomass, how hydrogen bonds are disrupted, and how lignin gathers into groups in different solvents. But it cannot directly describe the breaking of chemical bonds. Reactive force-field simulations, such as ReaxFF, help fill this gap. They can track chemical reactions involving free radicals and the breaking apart of molecules in systems containing thousands of atoms during pyrolysis, a process that heats material without enough oxygen for it to burn completely. However, their predictions depend on how their parameters are set. They also require heating rates much faster than those used in real processes to fit the reactions into computer simulations lasting only nanoseconds.
Between molecular research and engineering are methods for screening materials and predicting reaction rates. COSMO-RS calculations can compare thousands of possible solvents for extracting lignin before any laboratory experiments are carried out. In one example highlighted by the review, researchers screened 10,638 combinations of eutectic solvents, a type of solvent mixture, and narrowed them down to 22 candidates for experimental testing. The best-performing system dissolved 46.5 percent of its weight in lignin and achieved 111 percent relative laccase activity, a measure of the activity of an enzyme that helps break down lignin-related compounds. Kinetic models and models that describe reactions in moderate detail then turn molecular energy calculations into estimates of reaction rates that can be used in reactor design. One model for biomass pyrolysis included 32 reactions and 58 chemical species. It captured meaningful differences between feedstocks while remaining simple enough to connect with models of individual particles and entire reactors. The review warns that even a small error in the energy barrier of a rate-controlling reaction can lead to predicted reaction rates that differ by many orders of magnitude. For this reason, uncertainty must be included alongside the numerical predictions.
Machine learning acts as a link between the different levels of this modeling system, rather than replacing the other methods. Models designed to explain their predictions and trained on 467 pretreatment conditions predicted how much lignin was removed during treatment with deep eutectic solvents. They achieved an R² value of 0.8034 on a separate validation dataset. R² measures how well a model's predictions match the observed results. Other models predicted the amounts of bio-oil, char, and gas produced by pyrolysis using 273 experiments reported in scientific papers. Their R² values on the training data were above 0.90. The review also clearly explains the risks. Randomly dividing data into training and test sets can place almost identical conditions from the same study in both groups, making a model appear more accurate than it really is. In addition, most machine-learning research on biomass still relies on checking models against existing data rather than running new experiments to test candidates selected by those models.
The strongest examples are studies that connect different scales directly and test the results against real data. In one fast-pyrolysis study, a research team combined detailed reaction calculations with a model of how heat and materials move inside a particle. The team used a machine-learning model, called a surrogate model, to reproduce the particle model's results more quickly. They then connected it to a computer simulation of fluid flow in a fluidized-bed reactor, where solid particles are suspended by a moving gas. The combined model predicted bio-oil yields with an average error of about 6.4 percent across four feedstocks. At the level of entire processes, a study that looked ahead at the environmental effects of a scaled-up deep eutectic solvent pretreatment process found that its environmental impacts were at least 67 percent lower than those of the laboratory-scale process when solvent recovery was included. This shows how assumptions about scaling up a process can change conclusions about sustainability drawn from laboratory results.
The review also identifies data systems as a key part of reliable research, not just an administrative task. It presents the FAIR principles—findability, accessibility, interoperability, and reusability—as essential for building trustworthy models. Predictions about biomass depend on information such as where the feedstock came from, how it was pretreated, its ash and moisture content, the shape and size of the reactor, and the method used to analyze the results. Much of this information is often missing from published datasets. Examples of useful tools already in use include automated reactors for testing catalysts that record information in machine-readable formats through standard procedures and application programming interfaces (APIs), permanent sample identifiers with QR codes that track each sample's history, shared vocabularies based on formal definitions, and knowledge graphs that connect information through its meaning. However, the authors warn that a complete FAIR database covering the entire process of turning lignocellulosic biomass into products in biorefineries does not yet exist. Current systems are useful examples that can help make such a database possible.
The review takes a careful view of how mature these technologies are. Digital twins, which are computer models of real processes, are becoming practical at the process level. Researchers have also proposed a simpler digital twin that can work with realistic amounts of sensor data instead of requiring sensors to measure everything. However, fully adaptive digital twins that connect molecular models to industrial processes remain mostly a future goal. Such systems would need to combine biomass structure, quantum and molecular simulations, catalyst chemistry, fluid movement inside reactors, economic analysis, and real plant data. The review proposes levels of progress, starting with methods that pass information between separate models, moving through connections with parameters that have been tested and validated, and ending with predictions that work reliably across different feedstocks and scales. The authors conclude that validated connections between neighboring scales are becoming increasingly achievable, but reliable predictions that transfer all the way from individual atoms to industrial production are still uncommon.
The authors argue that the main goal for the next generation of biomass research should be to move beyond explaining past results and toward designing future experiments and processes that are tested in advance and account for uncertainty. They recommend a modular approach: use chemistry-based models where chemical reactions matter most, data-driven methods where speed and complexity are important, models of heat and material movement where scale affects performance, and whole-system analysis where costs and environmental impacts determine whether a process is practical. Researchers should test solvents and catalysts over several reuse cycles, keep unsuccessful predictions as evidence of where models stop working, and use sensitivity studies of reactors, costs, and life-cycle impacts to improve molecular design. This would help ensure that computing time and effort are spent on questions that can change real engineering decisions. When reliable knowledge can be passed between these different parts of the system, computer simulations and data-driven methods can become genuinely predictive tools for designing biorefineries that support a circular chemical economy with lower emissions.
Subject of Research: Computer-based and data-driven modeling methods that connect different scales to turn lignocellulosic biomass into useful products and design biorefineries
Article Title: Computational and Data‐Driven Strategies for Lignocellulosic Biomass Valorization: From Multiscale Modeling to Digitalized Biorefinery Design
Article References: Umar, A. B., Zheng, J., Lin, X., Liao, H., Xu, H., Huang, Y., Garba, Z. N., Uzairu, A., Zhao, W., & Yuan, Z. (2026). Computational and Data‐Driven Strategies for Lignocellulosic Biomass Valorization: From Multiscale Modeling to Digitalized Biorefinery Design. Advanced Science, 13(59), Article e78026. https://doi.org/10.1002/advs.78026
Image Credits: AI Generated
DOI: 10.1002/advs.78026
Keywords: lignocellulosic biomass, biorefinery, multiscale modeling, density functional theory, molecular dynamics, ReaxFF, machine learning, COSMO-RS, pyrolysis, FAIR data, digital twin, life-cycle assessment