INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1192
www.rsisinternational.org
Methane Yield Prediction for Anaerobic Digestion using Machine
Learning Models
Sharan Aditya
Chemical Engineering, Valdel Engineers, Bengaluru, Karnataka, India
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150100097
Received: 06 February 2026; Accepted: 11 February 2026; Published: 17 February 2026
ABSTRACT
Anaerobic digestion (AD) systems exhibit complex, nonlinear interactions between operational parameters,
making methane yield prediction challenging using traditional analytical methods. This study evaluates multiple
machine learning (ML) techniques to model methane production from an industrial reactor dataset containing
operational and environmental variables. A structured workflow incorporating correlation analysis,
dimensionality reduction, clustering, and supervised learning was implemented to identify key predictors and
assess model performance. Reactor temperature emerged as the dominant factor influencing methane yield, while
most other variables showed weak direct correlations. Principal component analysis and K-means clustering
revealed distinct operational regimes associated with different performance levels. Baseline linear regression
models achieved moderate predictive accuracy (R² ≈ 0.5), while nonlinear models—including ensemble methods
and artificial neural networks—provided only marginal improvements, suggesting dataset limitations rather than
algorithmic constraints. The results highlight the importance of richer biochemical and temporal data for
improving predictive modelling and supporting data-driven optimization of industrial AD processes.
Keywords: Machine Learning, Anaerobic Digestion, Methane Yield, ANN, K-means cluster, Regression
INTRODUCTION
The UN Paris Agreement is an accord adopted in 2015 and ratified by 195 countries worldwide. Its goal is to
decarbonise the planet, mitigate climate change, and reduce emissions intensity by 45% before 2030. To this
aim, countries have been incentivising the adoption of programs that contribute to the effort of going green in
energy generation and utilisation, curbing the use of carbon fuels, and cutting carbon emissions on a global scale.
Energy businesses are offered major benefits to shift towards using clean fuel and farms for clean energy
generation. Biofuels are a major component of available clean fuels in the market, accounting for 25% of the
global energy supply in 2024, with ever-increasing demand [1]. Biofuel offers advantages over solar, wind, and
other renewable energy resources: easy physical storage, a fixed, reliable rate of production and utilisation; an
option for waste treatment and recycling, including limiting methane release as a greenhouse gas; and is low-
cost to build.
Large-scale industrial processes to optimise the production of biogas from waste treatment plants have been
researched extensively. A popular method for extracting methane is Anaerobic Digestion (AD). A great way to
naturally process waste matter, it does not require high temperatures or large forms of energy to produce methane,
is easily stored, and is a highly economical and high-yielding industrial process for methane extraction [2].
Extensive research in the field has accumulated a vast amount of data, and the scope of using machine learning
models to optimise the operating conditions of this industrial process is in its early stages. Preliminary models
using regression techniques, ANNs, and SVM have been used to evaluate and optimise operating conditions.
However, many concerns arise in ascertaining the efficacy of these models owing to overfitting, inconsistent
feature engineering, and insufficient modelling [3]. Along with the difficulty of obtaining data, the practical
application of these datasets is often overlooked because OEMs who prepare the reactor tanks that host the AD
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1193
www.rsisinternational.org
process often optimise their tanks based on theoretical information and past data. The tanks cannot account for
live tracking and optimisation of their performance. Since standardisation of feed composition is difficult, the
yield percentage of methane cannot be standardised without actively studying live data from the reactor, and
tank operators often only run the tank according to the manufacturer’s manuals, leading to inconsistent results.
This paper aims to focus solely on the model training aspect of the various studies conducted, including feature
engineering, selection of relevant input variables, handling of multicollinearity, normalization and scaling
strategies, cross-validation techniques, and hyperparameter optimisation. The discussion will emphasise how
these methodological choices influence model generalizability and predictive performance, and will propose
best practices for developing reliable and consistent data-driven tools to optimise the yield percentage of methane
from AD.
Anaerobic Digestion: Principles, Benefits, and Process Overview
Anaerobic digestion (AD) is a well-established biotechnological process in which microorganisms convert
organic matter into biogas in oxygen-deficient conditions. The biogas produced consists primarily of methane
(CH₄) and carbon dioxide (CO₂), and can be used for heating, converted into electricity via combined heat and
power (CHP) systems, or upgraded to biomethane for grid injection and transport fuel applications.
The AD process occurs through four sequential biochemical stages: hydrolysis, acidogenesis, acetogenesis, and
methanogenesis. During hydrolysis, extracellular enzymes degrade complex macromolecules such as proteins,
lipids, and polysaccharides into soluble monomers, including amino acids, fatty acids, and sugars. Acidogenesis
converts these monomers into volatile fatty acids (VFAs) and CO₂, which are subsequently transformed into
acetate and hydrogen during acetogenesis. In the final stage, methanogenesis, acetate and CO₂ (in the presence
of hydrogen) are converted into CH₄. Among these stages, hydrolysis is frequently identified as the rate-limiting
step in biogas production from recalcitrant biomass, thereby constraining the overall efficiency of the AD process
[2].
The biochemical composition of food waste (FW) plays a critical role in determining methane yields and
production kinetics in anaerobic digestion. FW can be broadly categorized into high protein and lipid content
food waste (HPLFW) and high carbohydrate content food waste (HCFW). HPLFW - such as meat, dairy, and
mixed food wastes—typically achieve significantly higher methane yields than carbohydrate-dominated feeds.
Meat-dominant FW exhibits yield reaching approximately 337 mL CH₄/gCOD and conversion efficiencies above
96%, indicating high biodegradability. Dairy-dominant FW yields around 307 mL CH₄/gCOD with conversion
efficiencies close to 88%, while mixed FW achieves approximately 297 mL CH₄/gCOD with conversion rates
near 86%. These high yields are associated with balanced fermentation and methanogenesis rates. Lipids possess
a high energy content and a theoretical methane potential exceeding 1,000 mL CH₄/gVS. However, excessive
concentrations of long-chain fatty acids (LCFAs) can inhibit methanogenesis. In contrast, HCFW—such as fruit,
vegetable, and grain wastes—generally produce lower methane yields. Vegetable-dominant FW yields
approximately 238 mL CH₄/gCOD with conversion efficiencies around 68%, while grain-dominant FW yields
only about 171 mL CH₄/gCOD with conversion efficiencies below 50%. The reduced performance is attributed
to the inherently lower methane potential of carbohydrates and the slower degradation rates of fibrous
carbohydrates [4]
Two primary factors lead to higher efficiencies in methane yield: FW pretreatment strategies and a high
concentration of Anaerobic bacterial sludge maintained in the processing tanks. At least 10kg/m
3
/day of
concentrated sludge is required to meet the upkeep requirements for methane generation. Pretreatment strategies
generally involve increasing the lipid content in FW to increase methane output generation at the hydrolysis
stage. These methods include adding protein-rich substrates that possess a high theoretical methane potential of
approximately 0.5 Nm
3
/kg, but are susceptible to ammonia inhibition during degradation; Lipid-rich substrates
can achieve higher biogas production with a theoretical methane potential of about 1.0 Nm³/kg, but the high
LCFA concentration can inhibit microbial activity; Lignocellulosic=rich substrates are often abundantly
available but show variable degradability due to different fibre compositions, which hinder fatty acid generation
and molecular breakdown during acidogenesis. Mixed substrates can exhibit vastly varying results due to their
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1194
www.rsisinternational.org
highly variable composition, depending on diet, season, and cultural factors, with no clear benefit to the
process[4]
Temperature management is a key operational factor in anaerobic digestion, with controlled systems typically
operating in psychrophilic (10–20 °C), mesophilic (20–40 °C), or thermophilic (50–60 °C) ranges. Methane
production rates generally increase with temperature, reaching optimal levels around 35–37 °C for mesophilic
microorganisms and 55 °C or higher for thermophiles. Temperature shifts influence microbial community
composition and metabolic activity, with increases from 15 °C to 35 °C often resulting in enhanced biochemical
performance. While fluctuations can disrupt stability, anaerobic systems can recover following an acclimation
period, allowing for controlled changes in operating regimes without irreversible loss of performance [2].
A promising alternative is the integration of electrodialysis into anaerobic digestion (ADED). This approach
enables in-situ removal of ammonium ions while simultaneously enhancing process stability and methane
production. ADED systems have demonstrated the ability to reduce ammonium concentrations from influent
levels of up to 10,000 mg/L to below 2,000 mg/L. At moderate ammonia levels, methane production can increase
by more than 40% compared to conventional digestion, and at extreme concentrations, ADED can prevent
complete inhibition. The process also offers competitive energy efficiency, with consumption for ammonium
recovery comparable to industrial ammonia synthesis. From an environmental standpoint, electrodialysis
integration supports nutrient recovery, reduces the global warming potential of the process, sometimes to net-
negative levels, and lowers acidification and eutrophication impacts, contributing to more sustainable waste
management practices [5].
The economic feasibility of AD projects depends on multiple revenue and cost factors. For large-scale centralised
plants, gate fees for waste intake remain an important income source, while the sale of electricity, heat, and co-
products such as fertilizer enhances profitability. Compared to composting, anaerobic digestion offers a
competitive advantage as a net energy producer rather than an energy consumer. The viability of projects is
strongly influenced by supportive policies, including feed-in tariffs for renewable energy and restrictions on
landfilling organic waste [6].
All these parameters controlling the output and efficiency of methane production have given rise to many
machine learning models being used to study operating conditions, as their popularity has grown over the last
decade. Automation and process monitoring have become increasingly important for optimising plant
performance. With the increasing complexity of datasets required to measure the various input parameters of
production units for methane yield, traditional methods fall short in improving process efficiency. ML models
have the added advantage of better recognizing and predicting non-linear correlations between inputs, for
example, finding links between operating temperatures and pressures with the FW composition, among other
inputs. In essence, the more detailed the input data is from methane production units, the more efficient ML
models become in predicting their behaviour. ML models are also much more efficient when processing datasets
collated at an industrial scale, with tens of thousands of data points requiring much less computing power than
traditional methods.
METHODOLOGY
The dataset used in this study contains a comprehensive set of process variables relevant to anaerobic digestion
in a continuous stirred-tank reactor (CSTR). Key parameters include feed composition, nutrient solution usage,
reactor and biogas temperature, system pressure variables, and waste vapor characteristics. Methane yield was
designated as the target variable, representing the key performance indicator for process efficiency. The first
stage of analysis involved a complete inventory of available features, followed by statistical correlation analyses
to identify the most influential predictors of methane yield. The Pearson/Spearman correlation coefficient was
employed to measure linear dependencies, and the mutual information was used to capture nonlinear
relationships. Pearson correlations quantify the general covariance to assess the strength and direction of linear
relationships between two continuous random variables, which can explain the target’s dependencies on
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1195
www.rsisinternational.org
individual weights, if present. The Spearman correlation measures the strength and direction of a monotonic
relationship, whether linear or non-linear, giving a ranking to the Pearson correlation.
These methods ensured that both direct and complex dependencies between variables and methane yield are
quantified. Additionally, exploratory data analysis (EDA) was conducted to examine the distribution, variance,
and potential interactions among the input features [7].
Preprocessing of the dataset was performed to improve model performance. Continuous features were
normalised using standard scaling to account for differences in measurement units. Feature engineering strategies
were used to find improved correlations using a theoretical basis, such as creating derived ratios (e.g., methane
yield per unit of substrate input) and interaction terms between the input variables[8], to improve the richness of
the input space.
Machine learning model development was structured hierarchically. Baseline regression models, including linear
regression and support vector regression (SVR), were initially applied to establish reference performance.
[9]Subsequently, more advanced nonlinear models were implemented, including random forests and gradient
boosting algorithms (XGBoost), to capture higher-order interactions and improve prediction accuracy. For
potential scalability, artificial neural networks (ANNs) were also considered for modeling the complex
relationships inherent in biochemical processes.
To ensure robustness, the random forest regressor was trained and evaluated using k-fold cross-validation. Model
performance was assessed using multiple error metrics [7], including the coefficient of determination (R²), root
mean square error (RMSE), and mean absolute error (MAE), to provide a comprehensive view of predictive
accuracy. Hyperparameter optimisation was performed using grid search and Bayesian optimization techniques,
while regularization strategies and feature selection were employed to mitigate overfitting and enhance model
generalizability.
The methodological framework developed in this study provides a systematic approach to replacing traditional
empirical methods with machine learning-based predictive models. By integrating statistical analysis, feature
engineering, and multi-model evaluation, this methodology is designed to maximise the efficiency of anaerobic
digestion and provide actionable insights for process optimization in methane production.
Correlation Analysis
To investigate the relationships between process parameters and methane yield, Pearson’s correlation coefficient
(r) and Spearman’s rank correlation coefficient (ρ) were computed between all numeric features and the target
variable, Methane Standard Volume.
The results consistently highlight Reactor Temperature [°C] as the strongest predictor of methane yield, with a
Pearson r of 0.704 and a Spearman ρ of 0.706, indicating a strong positive correlation that is both linear and
monotonic. This suggests that higher reactor temperatures are strongly associated with increased methane
production. We also see a strong negative correlation with the organic loading rate (OLR), calculated as (standard
volume/reactor temperature), with a Pearson r of -0.577 and a Spearman ρ of -0.562.
Biogas Temperature [°C] also exhibited a positive association with methane yield, though weaker in magnitude
(r = 0.188; ρ = 0.296), implying that, while it has an effect, it is less influential than reactor temperature. Other
process variables, including Biogas Pressure (BP), Nutrient Solution Usage, and Feed, showed very weak
correlations with methane yield, suggesting a limited direct role in influencing methane generation under the
studied conditions. Variables such as BP, Waste Vapor Pressure, and Standard Volume were either weakly
negative or near zero, further confirming their negligible direct effect on methane production.
The heatmap visualization of the Pearson correlation matrix (Figure X) further reinforces these findings, showing
a strong correlation between methane yield and reactor temperature, while other parameters remain largely
uncorrelated. To explore the influence of operating parameters on methane yield, scatter plots were generated
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1196
www.rsisinternational.org
comparing methane yield against the remaining input parameters. The resulting visualizations reveal distinct
patterns that inform the selection of predictive features for subsequent modelling. Figure 03 illustrates the
pairwise relationships between methane standard volume and the other input variables. Across all scatter plots,
methane yield exhibits a pronounced banded structure, indicating the presence of discrete operational regimes
rather than a continuous response, with low, medium, and high yield states consistently observed. Variables such
as BP, nutrient solution usage, standard volume, biogas pressure, and waste vapor pressure show weak monotonic
trends and large within-band dispersion, suggesting limited direct influence on methane yield when considered
independently. In contrast, reactor temperature and, to a lesser extent, biogas temperature show clearer
structuring of the data, with higher temperature regimes more frequently associated with elevated methane
production, consistent with established anaerobic digestion kinetics[10]. The appearance of distinct vertical
alignments in several plots arises from discretised or fixed operating setpoints (e.g., reactor temperature, feed
type, pressure), reflecting controlled experimental or synthetic design conditions. These discrete inputs reinforce
the need for non-linear and tree-based machine learning models capable of capturing regime transitions and
interaction effects rather than relying on purely linear formulations.
Reactor temperature was found to vary discretely across the dataset, with values centred around approximately
28°C, 42°C, and 55°C. Substantial variation in yield was observed within each band, indicating that reactor
temperature alone does not fully determine methane output. This highlights the need to consider additional
process variables in predictive modelling. Alternatively, biogas temperature shows a continuous distribution and
a clearer monotonic relationship with methane yield.
As the biogas temperature increased, methane yield generally rose as well, though the relationship was not
strictly linear [10]. The scatter plot suggests that the yield approaches a plateau at higher biogas temperatures,
implying nonlinear or threshold effects. This variable, therefore, emerges as a critical determinant of yield, likely
reflecting its influence on microbial activity and fermentation kinetics.
Biogas pressure showed only marginal variation, with values concentrated in a narrow band around 1.01–1.03
bar. Other parameters also show a similar lack of correlations with methane yield, suggesting that these variables
are not significant in this process under the studied conditions. Reactor temperature as a categorical operating
regime and biogas as a continuous non-linear driver will primarily influence the model behaviour in a non-linear
study.
To study the distinct operational regimes seen in the scatter plot further, an unsupervised clustering approach
was applied to the input features. Principal Component Analysis (PCA) was first used to reduce the high-
dimensional operating data into two principal components (PC1 and PC2) that capture the majority of the
variance. The transformed dataset was then subjected to K-means clustering with k = 3, based on prior evidence
in the literature[7] suggesting the existence of three typical operating regimes in biogas reactors.
The clustering results (Fig 04) revealed three distinct groups of operating conditions. Cluster membership was
subsequently compared against methane yield to assess whether specific clusters were associated with more
favourable outcomes. A boxplot of methane yield distribution across the clusters (Figure Y) demonstrated clear
differences between groups. Cluster 0 exhibited an average methane yield of 1496 NmL d⁻¹, with a wide
distribution spanning from low to moderately high values. Cluster 1 was characterized by the lowest methane
yield, averaging 1125 NmL d⁻¹, and displayed a comparatively narrow range. Cluster 2 consistently showed the
highest methane yield, with a mean of 1803 NmL d⁻¹ and tighter variance, suggesting more stable operation
under these conditions.
These findings indicate that the operating features of the reactor naturally segregate into regimes that align with
performance differences. In particular, the high-yield cluster (Cluster 2) may represent a set of operating
conditions optimal for methane production. Conversely, Cluster 1 corresponds to a low-yield regime that may
reflect suboptimal or stressed operating states.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1197
www.rsisinternational.org
Importantly, cluster labels can now be incorporated as an additional feature in predictive modelling of methane
yield, or used to stratify the dataset for regime-specific analyses. This provides a pathway to integrate both data-
driven pattern recognition and mechanistic process understanding.
Machine Learning Models
Baseline Linear Models
To establish a baseline, Ordinary Least Squares (OLS) regression and its regularized variants (Ridge and Lasso)
were applied to the dataset. Model performance was assessed using the coefficient of determination (R²), root
mean squared error (RMSE), mean average error (MAE), and mean squared error (MSE).
Ordinary Least Squares (OLS), Ridge, and Lasso regression yield comparable coefficients of determination (R²
≈ 0.51) and high RMSE values (~443 NmL d⁻¹), suggesting that only about half of the variance in methane
production is explained by linear combinations of the other process variables. The near-identical performance of
OLS and its counterparts implies that multicollinearity is not the primary limitation and that regularization does
not improve generalization. Instead, these results point to an inherently weak linear relationship between the
available operational features and methane yield, consistent with the non-linear and threshold-driven nature of
anaerobic digestion. Consequently, linear models serve as a useful baseline but are insufficient for accurate
prediction, reinforcing the need for non-linear and ensemble-based approaches to capture the complex process
dynamics.
The MAE values remain relatively high and comparable between training and test sets (~354–364 NmL d⁻¹),
indicating moderate average prediction error and reinforcing that linear models capture only a limited portion of
the variability in methane production. Similarly, the large and consistent MSE values (~1.85–1.97 × 10⁵) suggest
the presence of substantial residual variance and occasional larger prediction deviations, supporting the earlier
observation that linear regression provides only a baseline representation of the process dynamics.
The close agreement across all three methods demonstrates that the underlying relationships between process
parameters and methane yield are largely non-linear under the studied conditions [11]. Linear models fail to
capture subtle non-linear interactions that are known to influence biogas production dynamics. As such, these
results serve as a strong baseline against which more flexible non-linear models (e.g., tree-based ensembles or
kernel methods) can be evaluated.
Decision Tree Model
The tree-based models demonstrate a moderate improvement over linear regression, indicating that non-linear
relationships are present, but are not sufficiently strong or well-resolved to enable high predictive accuracy. The
Decision Tree regressor shows reasonable training performance (R² ≈ 0.59) but a noticeable drop on the test set
(R² ≈ 0.48), accompanied by an increase in RMSE. This divergence suggests limited generalisation and a
tendency toward overfitting to local patterns in the training data.
Feature importance analysis reveals a dominant dependence on reactor temperature, which accounts for the
majority of variance explained by the model, while all other operational variables contribute marginally. This
concentration of importance indicates that the model relies heavily on a single controlling variable rather than
capturing broader multivariate process interactions.
The Random Forest model improves robustness relative to the single-tree approach, achieving higher training
performance (R² ≈ 0.76) and slightly improved test performance (R² ≈ 0.51). However, the persistence of a gap
between training and test metrics, along with a cross-validation mean R² close to the test value, suggests that the
ensemble primarily stabilizes variance rather than uncovering substantially new predictive structure. Similarly,
the Gradient Boosting model exhibits strong training performance (R² ≈ 0.75) but does not translate this
advantage to the test set, where performance remains comparable to the Random Forest and only marginally
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1198
www.rsisinternational.org
better than linear baselines. The higher training accuracy compared to testing again points to partial overfitting
and limited data in the feature space.
Overall, while ensemble methods outperform linear and single-tree models, their gains are modest, indicating
that the current feature set does not capture the dynamics of the governing methane production in its entirety
[12]. The dominance of temperature-related variables across all models reinforces their central role in anaerobic
digestion performance, but the inability of more sophisticated models to achieve substantial predictive
improvement suggests that key biochemical, microbial, or temporal descriptors are missing. These results
highlight the need for richer feature engineering and possibly dynamic or time-resolved inputs to meaningfully
enhance data-driven optimization of anaerobic digestion systems.
SVR and KNN Performance
The Support Vector Regression (SVR) model exhibits limited predictive capability for methane yield, with
training and testing R² values of approximately 0.48 and 0.43, respectively. The relatively close alignment
between train, test, and cross-validation performance indicates stable generalisation but also suggests that the
model is unable to extract a strong nonlinear structure from the available features. The high RMSE values further
confirm that prediction errors remain substantial, implying that the kernel-based mapping does not sufficiently
capture the underlying process complexity or that the feature space lacks the resolution required for effective
margin-based regression.
In contrast, the k-Nearest Neighbours (KNN) model demonstrates extreme overfitting, achieving perfect training
performance (R² = 1.00) with zero error, while generalisation performance deteriorates markedly on the test set
(R² ≈ 0.45). This behaviour is characteristic of instance-based learning methods in relatively noisy or sparsely
informative datasets, where local neighbourhoods in the training data do not represent the broader input space.
Overall, these results reinforce the conclusion that while non-linear models are necessary, models that rely
heavily on local interpolation or fixed kernel mappings are insufficient, further emphasizing the need for richer
process descriptors or hybrid modelling strategies.
ANN
The artificial neural network implemented as a multi-layer perceptron (MLP) demonstrates moderate predictive
capability, achieving a training R² of 0.58 and a test R² of 0.51. The model was run with 128 hidden layers to
improve performance and avoid overfitting, resulting in a relatively small gap between training and testing
performance, particularly compared to more flexible models such as KNN. However, the magnitude of the
RMSE on both sets indicates that prediction errors remain substantial, reflecting the limited explanatory power
of the available input features.
These results indicate that while the ANN can capture some non-linear relationships in the anaerobic digestion
process, its performance does not markedly exceed that of ensemble tree-based methods [13]. This points
towards the dataset's coverage being rather limited as opposed to the model’s inability to scope the dataset. In
the absence of additional biochemical, microbial, or time-dependent features, even flexible learning architectures
such as neural networks cannot fully resolve the complex dynamics governing methane production.
CONCLUSION
This study demonstrates the potential—and current limitations—of applying machine learning models to predict
methane yield in anaerobic digestion systems (Table 01) using a static, process-level dataset. Across linear, tree-
based, kernel-based, and neural network models, predictive performance consistently plateaued at a test R² of
approximately 0.5, with RMSE values remaining relatively high. While non-linear and ensemble models
modestly outperformed linear baselines, none achieved a substantial improvement in generalisation, indicating
that model choice was not the primary bottleneck. Correlation analysis, clustering, and feature importance results
consistently identified temperature-related variables as the dominant predictors, reinforcing established
anaerobic digestion kinetics. However, the strong reliance on a small subset of features, coupled with the limited
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1199
www.rsisinternational.org
contribution of other operational variables, suggests that the dataset does not adequately capture the full
biochemical and microbial complexity governing methane production.
The dataset provided by OEMs to improve the efficiency of methane yield from their CSTR tanks sees a primary
correlation only with reactor temperature. From an operational standpoint, OEMs should prioritise tight thermal
management within the optimal mesophilic or thermophilic window observed in the dataset, ensuring minimal
fluctuations and avoiding frequent regime switching [14]. Secondary process inputs such as nutrient dosing or
feed variations should be maintained within stable baseline ranges rather than aggressively tuned unless
supported by additional biochemical monitoring.
The clustered operational patterns further suggest that tanks may benefit from clearly defined operating regimes
rather than broad exploratory ranges, with temperature serving as the primary control variable, and other
parameters acting as stabilising factors. However, since the dataset lacks dynamic biological indicators and
feedstock composition variability, these recommendations should be interpreted as operational guidance based
on historical performance trends rather than universal process optimization rules.
The principal limitation of this work lies in the lack of data richness rather than methodological rigor. The dataset
is constrained by discretized operating regimes, the absence of time-resolved dynamics, limited representation
of feedstock composition variability, and no direct descriptors of microbial population, inhibition effects, or
transient process behavior. As a result, even highly expressive models such as ANNs fail to uncover additional
structure beyond what is already encoded in temperature-driven regimes [15]. These findings underscore that
meaningful gains in predictive accuracy will likely require integrating longitudinal data, real-time sensor
measurements, and mechanistically informed features rather than further algorithmic complexity alone. Future
work should focus on hybrid approaches that combine process knowledge with data-driven models, as well as
on the generation and sharing of standardized, high-resolution industrial datasets, which are essential for
translating machine learning from academic feasibility studies into practical tools for anaerobic digestion
optimization.
BIBLIOGRAPHY
1. “Statistical Review of World Energy - American Gas Association.” Accessed: Jan. 17, 2026. [Online].
Available: https://www.aga.org/research-policy/resource-library/statistical-review-of-world-
energy/?utm_source=chatgpt.com
2. T. Z. D. De Mes, A. J. M. Stams, J. H. Reith, and G. Zeeman, “Methane production by anaerobic
digestion of wastewater and solid wastes.”
3. H. Rutland, J. You, H. Liu, L. Bull, and D. Reynolds, “A Systematic Review of Machine-Learning
Solutions in Anaerobic Digestion,” Dec. 01, 2023, Multidisciplinary Digital Publishing Institute
(MDPI). doi: 10.3390/bioengineering10121410.
4. T. M. Anacleto et al., “Methane yield response to pretreatment is dependent on substrate chemical
composition: a meta-analysis on anaerobic digestion systems,” Sci. Rep., vol. 14, no. 1, Dec. 2024, doi:
10.1038/s41598-024-51603-9.
ML Models
Training R
2
Test R
2
Test RMSE
Test MAE
OLS, ridge, Lasso Regressions
0.53
0.51
443
364.36
Decision Tree
0.58
0.47
460.4
369.55
Random Forest
0.75
0.50
446.9
357.56
Gradient Boosting
0.74
0.49
450.82
361.26
SVR
0.48
0.43
480.23
337.31
KNN
1.00
0.45
471.93
375.95
ANN
0.59
0.5
450.24
359.71
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Page 1200
www.rsisinternational.org
5. J. Meng, L. Shi, Y. Hu, Z. Wang, Z. Hu, and X. Zhan, “Integration of anaerobic digestion and
electrodialysis for methane yield promotion and in-situ ammonium recovery,” Bioresour. Technol., vol.
402, Jun. 2024, doi: 10.1016/j.biortech.2024.130770.
6. A. H. Bhatt and L. Tao, “Economic perspectives of biogas production via anaerobic digestion,”
Bioengineering, vol. 7, no. 3, pp. 1–19, Sep. 2020, doi: 10.3390/bioengineering7030074.
7. S. Ö. Cinar, S. Cinar, and K. Kuchta, “Machine Learning Algorithms for Temperature Management in
the Anaerobic Digestion Process,” Fermentation, vol. 8, no. 2, Feb. 2022, doi:
10.3390/fermentation8020065.
8. R. Llanos-Lizcano, L. Senila, and O. C. Modoi, “Evaluation of Biochemical Methane Potential and
Kinetics of Organic Waste Streams for Enhanced Biogas Production,” Agronomy, vol. 14, no. 11, Nov.
2024, doi: 10.3390/agronomy14112546.
9. S. Mokraoui, A. Halilu, M. A. Hashim, and M. K. Hadj-Kali, “Modeling and simulation of biomass
anaerobic digestion for high biogas yield and CO2 mineralization,” Mater. Renew. Sustain. Energy, vol.
12, no. 2, pp. 105–116, Aug. 2023, doi: 10.1007/s40243-023-00233-8.
10. E. Piercy, X. Sun, P. R. Ellis, M. Taylor, and M. Guo, “Temporal Dynamics of Microbial Communities
in Anaerobic Digestion: Influence of Temperature and Feedstock Composition on Reactor Performance
and Stability.”
11. M. Mohammadianroshanfekr, M. Pazoki, M. B. Pejman, R. Ghasemzadeh, and A. Pazoki, “Kinetic
modeling and optimization of biogas production from food waste and cow manure co-digestion,”
Results in Engineering, vol. 24, Dec. 2024, doi: 10.1016/j.rineng.2024.103477.
12. O. J. Sinayobye et al., “Optimizing Biogas Production with Machine Learning: A Comparative Study
of Predictive Models,” 2024. [Online]. Available: https://www.jisem-journal.com/
13. D. A. Samuel, B. I. Eziefula, G. Orkuma, and A. Usman, “Forecasting of Biogas and Biomethane
Outputs from Anaerobic Co-digestion Using Multilayer Perceptron Artificial Neural Networks (MLP-
ANN),” International Journal of Multidisciplinary Approach Research and Science, vol. 3, no. 02, pp.
561–568, May 2025, doi: 10.59653/ijmars.v3i02.1529.
14. T. Z. D. De Mes, A. J. M. Stams, J. H. Reith, and G. Zeeman, “Methane production by anaerobic
digestion of wastewater and solid wastes.”
15. “Artificial Neural networks for predicting methane content in biogas from livestock waste”.