The research team, led by Sunyong Park and colleagues including Sunhwa Ryu, assembled a literature-derived dataset of 552 thermally treated biomass samples, spanning torrefaction and pyrolysis experiments drawn from dozens of published studies. Each record contained ultimate analysis, proximate analysis, higher heating value, and the process conditions—treatment temperature and residence time—that produced the fuel. Rather than feeding these numbers raw into an algorithm, the researchers invested heavily in preprocessing. Because both ultimate and proximate analyses are compositional data whose variables must sum to 100 percent, they applied closure normalization separately to each block and computed centered log-ratio transforms, a technique from compositional statistics that suppresses the spurious correlations that fixed-sum data otherwise generate.

Feature engineering went further. The team built process-aware descriptors—squared temperatures, logarithmic residence times, temperature–time interaction terms, and a combined severity index—to capture the nonlinear way thermal treatment intensity drives devolatilization, oxygen removal, and carbon enrichment. Composition-aware descriptors such as H/C and O/C ratios encoded the degree of carbonization, while proximate-derived quantities like the volatile-matter-to-fixed-carbon ratio captured the balance between volatile release and char formation. Critically, the authors also built in safeguards against information leakage and circular reasoning: auxiliary features were constructed only from training data, and a Dulong-equation-derived heating value surrogate was deliberately excluded from the final heating value models to ensure the algorithm was not simply copying a textbook formula hidden inside its own inputs.

After outlier removal using the interquartile range rule—which cut 234 records, or 42.4 percent of the dataset, leaving 318 cleaned samples—the team trained and compared linear, bagging, boosting, and tree-ensemble models, including Ridge regression, partial least squares, Random Forest, Extra Trees, XGBoost, and CatBoost. Each target variable got its own single-output model, selected by validation performance and tuned with grid search. To avoid the common trap of reporting a flattering single data split, the researchers re-evaluated every selected model across five repeated random train-validation-test partitions and reported the mean and standard deviation of test-set performance.

The results reveal a striking asymmetry between the two prediction directions. Predicting proximate properties and heating value from elemental composition—the ultimate-to-proximate direction—worked well. CatBoost models achieved repeated test-set R-squared values of 0.756 for volatile matter and 0.751 for higher heating value, with XGBoost reaching 0.702 for fixed carbon and Extra Trees 0.607 for ash. In plain terms, knowing what elements a treated biomass contains, along with how hot and how long it was treated, is enough to estimate its volatile content and energy density with useful accuracy. The heating value model predicted values with a mean absolute error of roughly 0.75 megajoules per kilogram—impressive for a screening tool that never lights a flame.

Going the other way proved harder. When the models tried to infer elemental composition from proximate data, performance dropped and became far more sensitive to how the data were partitioned. Carbon, hydrogen, oxygen, and heating value all landed in a moderate range, with repeated R-squared values between roughly 0.41 and 0.57, and large standard deviations betraying instability across splits. The reason is mathematical as much as chemical: proximate analysis compresses the fuel into three broad fractions, and many different elemental combinations can produce nearly identical volatile, fixed-carbon, and ash profiles. The inverse problem is simply underdetermined. Sulphur fared worst of all, with a negative mean R-squared of −0.085, meaning the model performed worse than naively guessing the dataset average. The authors are refreshingly blunt about this: sulphur, present at very low and heavily skewed concentrations, is flagged as unreliable and excluded from quantitative decision-making, and nitrogen is marked as a low-confidence screening estimate only.

The team also stress-tested their own pipeline. A filtering sensitivity analysis compared no outlier removal, gentle filtering, and the strict final setting, and found the effects were target-dependent rather than uniformly beneficial. Volatile matter and heating value barely changed across scenarios, but ash prediction improved substantially with stricter filtering—its R-squared rising from 0.44 to 0.64—because extreme ash values from heterogeneous feedstocks were dragging the model off course. Meanwhile, several inverse-direction targets actually got worse under strict filtering, suggesting that aggressive pruning had removed compositional variability the models needed. This nuanced outcome matters: it demonstrates the reported performance was not an artifact of quietly discarding nearly half the data, while honestly acknowledging that the final models cover a narrower slice of the biomass universe than the raw literature does.

The benchmark against human-devised formulas may be the study’s most compelling result. On the same held-out test set, the Park et al. empirical equations for predicting proximate properties from elements managed R-squared values of 0.72 for volatile matter and 0.40 for ash, while the machine-learning models reached 0.89 and 0.64 respectively. The gap widened dramatically for heating value: the century-old Dulong equation produced a negative R-squared of −1.01 on this thermally treated dataset, whereas the Dulong-free machine-learning model scored 0.87. In the inverse direction, the Shen et al. correlations from 2010 essentially failed on modern torrefied biomass, with carbon prediction collapsing to an R-squared of −0.54, while the machine-learning framework achieved 0.62. The lesson is that fixed-form equations derived decades ago on different feedstocks simply do not transfer to the complex, treatment-dependent fuels now entering the market, whereas flexible nonlinear models trained on relevant data do.

Interpreting the models with SHAP analysis added physical plausibility to the statistics. For heating value prediction, carbon- and oxygen-related descriptors—including closure-aware carbon terms and O/C variables—dominated the feature importance, exactly matching the established chemistry that carbon enrichment and deoxygenation raise energy density during torrefaction. Severity-weighted terms and temperature–residence-time interactions also contributed, indicating the models had genuinely learned treatment-intensity effects rather than memorizing feedstock labels. For ash, the influential features were hydrogen, nitrogen, and sulphur descriptors, which the authors interpret not as causal drivers of ash but as statistical fingerprints of feedstock type and inorganic-rich sample groups—a candid reminder that SHAP values describe model behavior, not experimental mechanism.

To make the framework usable beyond the modeling team, the researchers wrapped it in a lightweight Streamlit graphical interface that accepts either ultimate or proximate inputs plus process conditions and returns cross-predictions with heating value estimates, input validation, compositional closure checks, and warnings when users venture outside the training domain. Reliability flags are baked in: sulphur outputs are explicitly marked unusable for quantitative decisions. The authors are careful to position the tool as a screening aid, not a laboratory replacement—regulatory evaluation of low-concentration elements still demands real analysis, and the literature-derived dataset carries inherent heterogeneity in feedstock, reactor design, and analytical protocol. But for traders, plant operators, and researchers needing a fast preliminary read on whether a biomass batch is worth pursuing, the message is clear: a well-engineered machine-learning model, honest about its own uncertainty, can now deliver in seconds what once required a destructive, instrument-intensive afternoon—and it can do it more accurately than the empirical formulas the field has relied on for decades.

Subject of Research: Machine-learning cross-prediction of ultimate and proximate properties of thermally treated biomass for cleaner fuel assessment

Article Title: Machine-Learning-Based Cross-Prediction of Ultimate and Proximate Properties for Cleaner Biomass Fuel Assessment

Article References: Park, S., Lee, J., Yang, J., Kim, S., Hwang, K., & Ryu, S. (2026). Machine-Learning-Based Cross-Prediction of Ultimate and Proximate Properties for Cleaner Biomass Fuel Assessment. Cleaner Engineering and Technology, Article 101333. https://doi.org/10.1016/j.clet.2026.101333

Image Credits: AI Generated

DOI: 10.1016/j.clet.2026.101333

Keywords: biomass, machine learning, torrefaction, proximate analysis, ultimate analysis, higher heating value, fuel quality, compositional data, SHAP, XGBoost, CatBoost, cleaner energy