Performance emission and combustion characteristics of TIO₂ and CEO₂ nanoparticle enhanced Mahua biodiesel diesel blends using experimental and machine learning approaches

Venkatesh Janaki; P S Ranjit; B Balakrishna

doi:10.1038/s41598-026-38657-7

. 2026 Feb 12;16:8594. doi: 10.1038/s41598-026-38657-7

Performance emission and combustion characteristics of TIO₂ and CEO₂ nanoparticle enhanced Mahua biodiesel diesel blends using experimental and machine learning approaches

Venkatesh Janaki ^1,^2,^✉, P S Ranjit ², B Balakrishna ³

PMCID: PMC12976138 PMID: 41680251

Abstract

The present research employs an experimental–computational study for investigating the performance, combustion and emission characteristics of a single-cylinder compression ignition engine fuelled with conventional diesel and Mahua biodiesel–diesel blends doped with TiO₂–CeO₂ hybrid nanoparticles. A comprehensive engine test campaign was conducted at a variety of loads, and critical parameters such as brake thermal efficiency (BTE), brake specific fuel consumption (BSFC), and exhaust emissions (Carbon Monoxide (CO), Hydrocarbon (HC), Oxides of nitrogen (NOₓ), smoke, and Carbon di-oxide (CO₂) were carefully analysed over the entire engine operating range. They also examined the combustion of gasoline in the engine. Compared to the normal diesel, Mahua biodiesel blended with nanoparticles revealed a significant enhancement in brake thermal efficiency of 6–8% for the blends. There was also up to a 7% reduction in brake-specific fuel consumption at higher loads. Emission reductions were substantial, with CO, HC, and smoke opacity reduced by 25–30%, 20–25%, and 35–40%, respectively. The cylinders, being hotter, did however lead to a moderate increase in NOₓ emissions (by 8–12%) because they accelerated oxidation processes. In addition to the experimental study, we developed machine learning-based predictive models to predict engine outputs. The XGBoost model performed better than others. R² values obtained through Leave-One-Out Cross-Validation (LOOCV) were very high: 0.993 for BTE, 0.996 for BSFC and > 0.97 for all emission parameters. Both RMSE and MAE were very low for the model. The analysis of residuals and sensitivity demonstrated strong generalisability and a consistent effect of physical traits. The elaborate experimental and machine learning framework demonstrated the potential of Titanium Dioxide Nanoparticles (TiO₂)–Cerium Oxide Nanoparticles (CeO₂) nanoparticle-enhanced Mahua biodiesel towards a viable alternative fuel that may enable a cleaner and more efficient operation of compression ignition engines.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38657-7.

Keywords: Mahua biodiesel, TiO₂ nanoparticles, CeO₂ nanoparticles, Diesel engine performance, Emission characteristics, Combustion analysis, Brake thermal efficiency, Machine learning prediction, XGBoost, SVR

Subject terms: Energy science and technology, Engineering, Environmental sciences

Introduction

Increased demand for energy and environmental damage from burning fossil fuels have promoted the use of renewable sources, with biodiesel being one of them. Biodiesel is degradable, nontoxic, and releases oxygen during combustion. But the practical application was limited in diesel engines due to drawbacks such as higher viscosity, lower calorific value and poorer volatility than diesel. These features result in incomplete combustion, low brake thermal efficiency and poor fuel economy¹. Mahua (Madhuca indica) is a non-edible oil-yielding tree, which is available in plenty in India, and it does not compete with edible sources as the source for biodiesel production². The current paper explores the influence of Mahua biodiesel as a blending component with mineral diesel without demanding any engine hardware modifications for the perfect utilisation. In the use of biodiesel/diesel blends containing metal oxide nanoparticles, it has been shown to improve the combustion characteristics³. Nanoparticles, including titanium dioxide (TiO₂) and cerium oxide (CeO₂), serve as oxidation catalysts that can enhance fuel-air mixing and facilitate full combustion⁴. TiO₂ acts on promoting oxygen release, and CeO₂ plays the role of oxygen buffer by virtue of its good redox cycling properties. These outcomes are associated with higher thermal efficiency and lower emissions of CO, HC, and smoke⁵. The current work also experimentally analyses the effect of nanoparticles of TiO₂ and CeO₂ at various concentrations (25, 50, 75 Parts Per Million (ppm)) on performance, emission and combustion of Mahua biodiesel–diesel blends in a single-cylinder diesel engine⁶. Moreover, machine learning models are used to facilitate predictive performance evaluation and confirm the experimental observations⁷.

Depletion of fossil fuel reserves and environmental issues have increased the research for alternative fuels in Compression Ignition (CI) (engines. Biodiesel, a renewable fuel produced from non-food materials, has had significant attention⁸. Recent studies on biodiesel blends, nanoparticle additives and their impact on diesel engine performance and emissions are discussed in this review article. Non-edible oils of the second generation, such as Mahua (Madhuca indica), Jatropha (Jatropha curcas), Karanja (Pongamia pinnata) and Neem (Azadirachta indica), have been explored widely for use as feedstocks of biodiesel. The oils are transesterified with methanol and an alkaline catalyst to form fatty acid methyl esters, which have low viscosity and good burning properties. Biodiesel of mahua oil exhibited enhanced performance properties, and the B20 blend had a good trade-off between performance and emissions⁹.

The addition of metal oxide nanoparticles in a blend with biodiesel offers an emerging solution to some of the challenges facing the use of biodiesel. 20–200 mg/L TiO₂ nanoparticles in Mahua biodiesel blends showed enhancement of brake thermal efficiency (1.1–1.5%) and decrease in CO (37.42–46.54%) and HC emissions (22.54–28.4%). A larger active surface area of nanoparticles allows more reactive surfaces for oxidation, good thermal conductivity, and shorter ignition delay. 7.66% decrease in SFC and better combustion behaviour in BD blends doped with 50–100 ppm Al₂O₃ NPs¹⁰. The improved chemical reactivity of nanoparticles leads to the acceleration of the chemical reaction and micro-explosion phenomenon, which results in better fuel atomisation and complete combustion. Numerous biodiesels can be blended in different ways to improve physicochemical properties¹¹. Blends of Jatropha-Mahua and Jatropha-Karanja and Mahua-Karanja Biodiesels with diesel had better cold flow properties and combustion performance. The blend D80JB10KB10 (diesel 80% + Jatropha 10% + Karanja 10%) showed higher brake thermal efficiency by about 7.8% at the maximum load with an insignificant increase in emissions levels¹².

Studies with variation in compression ratio on dual biodiesel blend (Mahua-Jatropha) indicated the possibility to operate B20 at CR 16.5:1, having 1.58% higher brake power and mechanical efficiency of 12.42% more than diesel. Nevertheless, the in-cylinder pressure peak was slightly drop (0.15–0.36 bar) since of the biodiesel’s greater viscosity and lower calorific value. Biodiesel combustion properties exist greatly affected by fuel-injection variables. The heat release rate of B20T50 at 180 bar injection pressure witnessed 5.3% higher performance associated to another pressurisation (180–240 bar). Use of 7-hole fuel injectors (0.85 mm diameter) enhanced spray atomisation and air-fuel mixing, enabling higher brake thermal efficiency from 8.83% to 16.4%. TRCC designs provide improved swirl and turbulence, leading to better fuel-air mixture and complete burning. Brake thermal efficiency of TRCC configurations was higher by 2.9–3.3% than the hemispherical combustion chamber because the squish effect and bulk airflow character were improved¹³.

Studies on Mahua biodiesel combined with biogas in dual mode of operation indicated an increase of 32.39% brake thermal efficiency (BTE) and a significant reduction in NOx (46.32% at 12 lpm) for the B20 blend at an optimal flow rate of 8 lpm as compared to only diesel operation. But, at high biogas flow rates, it occurred at the cost of increased HC and CO emissions by an incomplete combustion. The CO (34.05%), HC (21.62%), and smoke emissions are decreased substantially, but NOx increases by 4–21.29% with BBD blends due to the high O₂ level and the increase in combustion temperature. The trade-off of flames can be partially bypassed by the addition of nanoparticles which have catalytic properties to reduce combustion temperatures without decreasing the complete oxidation¹².

The addition of TiO₂ Nanoparticle resulted in a significant improvement in engine performance and decreases in pollutants, regardless of the type of feedstock sources used for preparing biodiesel mixtures, including waste cooking oil, Eichhornia crassipes, tamanu, and acetone-based mixtures¹⁴. For the study of WCO methyl ester blends (B20) doped with TiO₂, Al₂O₃, and CNTs, B20T100 had an improvement in brake thermal efficiency by approximately 4% and reduced specific fuel consumption by about 4.5% when compared to the plain biodiesel blend. The improvement of thermal efficiency is ascribable to the catalytic effect of TiO₂, which augments the combustion kinetics by increasing atomisation and surface-to-volume ratio effects¹⁵. This led to an increment of 5% in peak cylinder pressure and 4% in heat release rate. Performed tests also showed an 11% decrease in CO, 8% in HC and almost 10% in smoke opacity, representing better combustion. However, the NOx level was slightly increased (5%) due to high combustion temperature as oxidants gradually increased¹⁶.

Correspondingly, Akshay Jain et al. (2023) investigated the effects of adding TiO₂ nanoparticles (50–150 ppm) to Eichhornia crassipes biodiesel on overall brake thermal efficiency and observed an increase of up to > 1% with respect to that of conventional diesel at an injection pressure of 220 bar. Lower ignition delay, higher heat release rate and larger peak cylinder pressure were generated by the increasing number of nanoparticles, and at 150 ppm nanoparticle concentration, it increased to 6.62%. The enhanced mixing quality and the HC ratio reduction of around 39% CO were respectively reduced to levels as low as 49% at elevated fuel injection pressures with stable combustion¹⁷. However, NOx emissions increased slightly (up to 8%) at high nanoparticle loading due to higher in-cylinder temperatures. The researchers noted that the increased injection pressures with the addition of TiO₂ also resulted in higher oxidation efficiency and slightly higher NOx emissions¹⁸.

On the other hand, CeO₂ nano-additives are most effective in emission control by virtue of their excellent redox properties, as they can show a reversible redox cycle between Ce⁴⁺ and Ce³⁺ states to enhance continuous oxidation of unburned carbon soot during combustion. Conferring that 20 ppm CeO₂ was blended with diesel–biodiesel mixtures in the experiments, CO and HC were reduced by 32% and 45%, and NOx increased slightly by 6%. The catalytic reaction 4CeO₂ + C_soot → 2Ce₂O₃ + CO₂ enhanced active oxidation of the soot particles, which decreased the formation of particulate matter. In another investigation, biodiesel with 80 ppm CeO₂ used in B-20 waste cooking oil biodiesel exhibited an increase in brake thermal efficiency of 10.9% and a reduction in brake-specific fuel consumption of 11.1% at the advanced injection timing conditions. More importantly, HC and smoke emissions were 35% and 14% lower compared to neat B20 operation. Furthermore, with retarded injection timing, NOx decreased by approximately 16%, which demonstrates CeO₂ to be able to control the combustion process and emissions improvement via the timing strategy at the same time. By contrast, the combined study of TiO₂ and CeO₂ demonstrated synergistic effects on engine performance¹⁹.

High thermal conductive and catalytic surface energy of TiO₂ enhances mostly ignition and oxidation, resulting in low CO & HC but with a slight increase in NOx. In contrast, CeO₂ acts as an active carrier of oxygen, reducing UHC and smoke by the soot oxidation reactions and stabilising combustion while possibly suppressing NOx via injected timing parameter tuning. All in all, both nanoparticles serve to enhance exergy efficiency, decrease specific fuel consumption and improve the sustainability index of biodiesel-fuelled engines²⁰. An oxidation catalysis of TiO₂ facilitates heat release and thermal conversion processes, while the oxygen storage performance of CeO₂ strongly suppresses hazardous pollutant emissions and particulate matter generation²¹. In general, the best results are obtained at concentrations within 30–90 ppm of TiO₂ and 20–80 ppm of CeO₂. Thus, mixing them could potentially lead to future biodiesel engines that are high-efficient, low-emission, and adaptable to different working conditions¹⁹.

ML algorithms have produced excellent predictive value in biodiesel engine-related research in several domains. Among the B0, B10, and B20 blends of biodiesel-diesel in the comparisons SVR had the lowest RMSE, MAPE, and MAE values and the highest correlation coefficients R. SVR successfully accommodated the complex non-linear dynamic interactions governing combustion processes and exhibited superior performance²². In biodiesel studies, ANN is the most common ML method used as it can be implemented to approximate continuous functions, and also used in modelling complex relationships. ANN to predict BTE and NOx so that the correlation coefficient R² between estimation results provided by ANN algorithms and measurements of those parameters were significant high (R² = 0.93–0.99)²⁴. models had better performance on prediction than other ML models for the study case (Pumpkinmaize biodiesel blend), where R² value was approaching unity for BTE as 0.984 & NOx as 0.959²⁵. XGBoost Extreme gradient Boosting performs very well in predicting the emission parameters, particularly CO₂ and hydrocarbon HC emissions²⁵. The model’s R² results of the gradient boosting decision tree were 0.956 for CO₂ emission prediction and 0.973 for HC emission prediction, superior to ANN and SVM in both uses²⁶. Random Forest RF and Gradient Boosting Regressor GBR show a promising perspective in multi-engine parameter prediction. The results of the predicted engine performance (HC, CO, CO₂, NOx, BTE (brake thermal efficiency), BSEC (brake specific emissions rate) and smoke emissions) using the LASSO and GBR methods showed a high-order agreement with engine experiment measurements. CatBoost and Ridge Regression, included in the engine performance prediction ML toolbox, are not well referenced by the literature reviewed. These are great models when there fails to be a significant categorical variable and for limiting overfitting in regression problems²⁷. Application needs determine algorithm selection. The comparison of ANN, XGBoost and SVM for several biodiesel blends identified that XGBoost was better in the prediction of emission characteristics with respect to ANN concerning the performance parameters²⁸. The adopted ML validation and interpretability strategy follows recent best practices for small-sample scientific machine learning²⁹. This implies that hybrid methods combining the strengths from different algorithms for various prediction objectives might be successful. SVR, RF, GBR, XGBoost, CatBoost, Ridge Regression are working fine as ML models for predicting biodiesel engine performance and emissions, as reported in the different literature reviews.

Although significant experiments have been conducted on biodiesel-fuelled compression ignition engines, there are still many knowledge gaps. Currently available research on biodiesel-nanoparticle blends focuses on nanoparticle additives in isolation. Limited information is available on dual combustion associated with hybrid MG oxide nanoparticles, especially with respect to Mahua biodiesel. Second, past studies have focused strongly on performance and emissions but lacked the development of accurate predictive models that show the engine behaviour across different conditions. Third, machine learning is being used increasingly to analyse internal combustion engines, but often only with simple splits into train and test datasets, and almost never with the cross-validation methods appropriate for small experimental datasets. In addition, across the modelling spectrum little attention has been paid to the fidelity of models, residual behaviour and the requirement for a sensitivity analysis linking data-driven predictions to the fundamental combustion physics.

To overcome these constraints, in this work, a comprehensive experimental-computational framework is presented by synthesising TiO₂–CeO₂ dual-promoted nanoparticle-enhanced Mahua biodiesel along with an advanced machine learning validation. The innovative aspects involve.

(i)
Experimental elucidation of the synergistic catalytic and oxygen-buffering mechanism of TiO₂ and CeO₂ nanoparticles on combustion, performance, and emissions;
(ii)
Generation and validation of different machine learning approaches for CI engine multi-output prediction, with Leave-One-Out Cross-Validation (LOOCV) to ensure robustness in the data-limited regime; and).
(iii)
Path-dependent improvement of physical interpretability via residual and sensitivity analyses. Such an integrated approach enables accurate engine predictions and helps decipher the contribution of nanoparticles in improving biodiesel combustion.

Methodology

Fuel preparation

Mahua oil (Madhuca indica) was transesterified with methanol, using potassium hydroxide (KOH) as a catalyst, to produce biodiesel. The reaction was conducted at 60 °C, with a molar ratio of methanol to oil (6:1) and under constant stirring for 90 min. The solution was left to stand for 24 h in order for the glycerol to settle. The upper layer was separated and washed with warm distilled water to remove impurities, followed by drying to get Mahua Methyl Ester (MME). A B20 (80% diesel + 20% MME) blend was selected as the reference for its well-documented relationship between performance and emissions.

Two types of metal oxide nanoparticles, i.e., titanium dioxide (TiO₂) and cerium oxide (CeO₂), were used as fuel additives. The nanoparticles were diluted in the B20 oil blend at concentrations of 25, 50, and 75 ppm. The desired mass of each nanoparticle was pre-mixed with a small amount of biodiesel and ultrasonicated for 45 min to retard any agglomeration. The solution was subsequently poured into the B20 blend, and magnetic stirring continued for 30 min to obtain a good dispersion. The ultimate fuel samples were visually examined and kept in dark capped amber vials to prevent oxidation and precipitation. The fuel properties shown the Table 1 and fuel preparation is shown in Fig. 1a. The fuel properties were measured by standard ASTM test procedure. Density was measured by hydrometer in accordance with ASTM D1298. All kinematic viscosities at 40 °C were measured in a Redwood viscometer according to the ASTM D445. Flashpoint was measured in a Pensky–Martens closed cup device per ASTM D93. The bomb calorimeter was used to determine the calorific value of the fuels, following the ASTM D240. Cetane number data were obtained directly from supplier specifications and correlated methods according to ASTM D613. All of the measurements were performed under laboratory-controlled conditions to provide for accuracy and repeatability.

Table 1.

Fuel properties.

Fuel nomenclature	Fuel sample	Density (kg/m³)	Kinematic viscosity (mm²/s @40°C)	Flash point (°C)	Calorific value (MJ/kg)	Cetane number	Relevant standard limits
D100	100% Diesel	825.8	3.06	61	44.23	47	ASTM D975
	100% Mahua	872.3	3.45	155	39.38	53	ASTM D6751
B20	B80% Diesel + 20% Mahua	850	3.23	122	41.89	50	ASTM D7467
B20T25	B20 + TiO₂ (25 ppm)	851	3.25	123	41.95	51	ASTM D7467
B20T50	B20 + TiO₂(50 ppm)	852	3.27	124	42	52	ASTM D7467
B20T75	B20 + TiO₂(75 ppm)	853	3.3	125	42.05	52	ASTM D7467
B20C25	B20 + CeO₂(25 ppm)	851.2	3.26	124	41.96	51	ASTM D7467
B20C50	B20 + CeO₂(50 ppm)	852.3	3.28	125	42.01	52	ASTM D7467
B20C75	B20 + CeO₂(75 ppm)	853.5	3.31	126	42.06	53	ASTM D7467

Open in a new tab

Fig. 1 — (a) fuel preparation and (b) experimental setup.

The recorded fuel properties of diesel, Mahua biodiesel and B20 blends imparted with nanoparticles were compared with the respective limits outlined in prescribed international standards for their applicability in the operation of compression ignition engines. The diesel fuel properties are according to the ASTM D975 specifications, while the Mahua biodiesel properties are below the specification limits according to ASTM D6751, EN 14,214 and IS 15,607. The density and kinematic viscosity of all B20 and nanoparticle-doped blends are within the effective limits set by ASTM D7467, maintaining appropriate fuel injection and atomisation characteristics. The incorporation of TiO₂ and CeO₂ nanoparticles only elicits negligible variations in density and viscosity, which are still within the appropriate limits. Biodiesel and blended fuels have flashpoint values far above the minimum allowable values, indicating safe handling and storage of biodiesel as a neat fuel or within a blended fuel. The lower values for calorific value for pure blends of biodiesel than for diesel are due to the fact that biodiesel is an oxygenated fuel; however, these values are still in conformance with typical values for biodiesel. In general, as all the fuels met the requirements of the international fuel quality standards, the results imply the applicability of all the tested fuels for CI engine combustion application.

Thus, the changes in fuel properties due to the addition of TiO₂ and CeO₂ nanoparticles at low concentrations (25–75 ppm) are limited and fall within a narrow range. The changes we observed are similar to the measurement uncertainty associated with the conventional fuel property testing procedure and should not be interpreted as significant changes in the properties of bulk fuel. At these very low dosing levels, it is likely that the main effect of the nanoparticle on combustion behaviour is catalytic activity, rather than a large change in the intrinsic properties of the fuel. Therefore, the ranges of property variation are theoretically bounded and are experimentally acceptable such that they would not cause fuel standard violations.

Further, the present study used Titanium dioxide (TiO₂) and Cerium oxide (CeO₂) nanoparticle commercially available solution obtained from certified supplier used as received. The nanoparticles were purchased with technical details ascertained from the manufacturer, such as particle size and purity, which are comparable to values found in the literature regarding fuel additives. As herein, the main highlight of this work also does not dwell on the in-house synthesis and microscopic characterisation (SEM/TEM) of the nanomaterials like other works which are based more on the material development, and hence that is not performed either. The model assumes a power-law size distribution of stable clusters, and it does not capture any potential effects of normalised nanoparticle morphology and size distribution on fuel behaviour and is thus qualified as a limitation and a further research direction.

Engine test setup

All the experimental investigations were conducted on a single-cylinder, four-stroke, direct-injection, water-cooled diesel engine (Kirloskar TV-1). The main features of the engine are listed in Table 2 and schematic experimental setup shown in Fig. 1b. Load was applied onto the engine using an eddy current dynamometer. The combustion variables were recorded by a Kistler piezoelectric pressure transducer, while the crank angle information was collected using an AVL 364 C crank angle encoder with a resolution of 0.1° of Crank Angle (CA). Exhaust emissions (level) Emissions were measured using an AVL 444 N gas bench to test CO, CO₂, HC, and NO_x and an AVL 437 C opacity meter to measure smoke opacity. The instantaneous pressure, load, and exhaust data were collected via a LabVIEW-based Data-Acquisition (DAQ) system.

Table 2.

Engine specification.

Parameter	Specification
Engine model	Kirloskar TV-1
Type	Single-cylinder, 4-stroke, water-cooled, DI diesel engine
Rated power	5.2 kW @ 1500 rpm
Compression ratio	17.5:1
Bore × Stroke	87.5 mm × 110 mm
Injection pressure	210 bars
Injection timing	23° before TDC
Cooling system	Water cooled
Dynamometer	Eddy current, 7.5 kW capacity

Open in a new tab

Experimental procedure

The engine was then initially run on neat diesel to take the baseline readings. Measurements were taken after stabilization at each operating point. All the fuel mixtures (B20, B20 + TiO₂, and B20 + CeO₂ 25, 50 & 75 ppm) were operated in five load conditions: 0%, 25%, 50%, 75%, and 100%. The engine was run 10 min for each test point, and the average of the triplicate readings was taken to eliminate random errors. Computed performance parameters were Brake Thermal Efficiency (BTE) and Brake Specific Fuel Consumption (BSFC).

Machine learning-based prediction

Machine learning models were created for reliable prediction and generalization of the experimental results. The input variables of the datasets were load, TiO₂/CeO₂ level, and test fuel composition, while the output variables consist of BTE, BSFC, CO, HC, CO₂, NO_x, and smoke.

The following regression analyses were performed:

Support vector regression (SVR).
Random forest (RF).
Gradient boosting regressor (GBR).
XGBoost regressor.
CatBoost regressor.
Ridge regression.

The dataset was split into 70% training and 30% testing samples. Goodness-of-fit statistics model fit were tested with R², RMSE, MAE, MSE, and actual vs. fitted plot. The optimal model was selected as the greatest R² and lowest error metrics.

Hyperparameter tuning strategy

Given the limited size of the experimentally generated dataset, hyperparameter tuning was performed using a conservative and controlled approach to avoid overfitting. Independently of each ML model, a number of hyperparameters were tuned to develop the best-performing model in slightly wider ranges due to top values referenced in the literature for ML treatment of engines. To purposely mitigate model complexity parameters including tree depth, number of estimators, regularization strength, and kernel parameters were controlled. For each model, the best-fitting hyperparameters were chosen that minimized prediction error without losing the capacity to generalize. It serves to guarantee a fair comparison of modelling approaches, as well as reliable and reproducible prediction performance when small-sample conditions prevail.

Model validation and error analysis

An extensive model validation framework was also put into place to validate that the machine learning models built from the scarce sample dataset obtained in the laboratory are robust and reliable. Besides the typical train-test separation, the most common method to validate the results was Leave-One-Out Cross-Validation (LOOCV). In LOCCV, one is used for validation, and others are for training. That is applied to all available data points. This allows every data point to be used once to evaluate model generalization performance which yields an unbiased assessment and the highest utilization of the data. When it comes to internal combustion engines, experimental datasets can be so small that both the approach and the model are beneficial in this case. Examined the prediction performance metrics averaged from each iteration to evaluate model performance overall.

Additionally, a residual analysis was conducted to examine the distribution and nature of prediction errors. In particular, the study explored the discrepancies between experimental results and predictions based on machine learning for some of the key output parameters. Residual plots were used to assess systematic bias and whether errors were evenly distributed across the operational range of values. It demonstrates the random spread of residuals around zero, indicating that the ML models picked up the baseline physical relationships and avoided overfitting. Both LOOCV and residual diagnostics are used in conjunction to ensure the output metrics for an ML model are mathematically valid, contextually appropriate, and not reliant on a single data split. Such a validation framework increases the confidence in the predictive and generalized capability of the proposed machine learning models for CI engine performance and emissions. In addition, model sensitivity analysis was performed to assess the relative influence of input parameters on engine output predictions.

Uncertainty analysis

A quantification of uncertainty was conducted to assess the confidence and reproducibility of the experimental data. For each measured parameter, its associated uncertainty was calculated based on the manufacturer specifications of the measuring instruments. The root-sum-square (RSS) method was used to calculate the overall uncertainty of derived parameters and is expressed as follows:

U is the overall uncertainty, while u_i are the individual measurement uncertainties. Using this method, the determined uncertainty in BTE and BSFC was ± 1.8% and ± 1.6%, respectively. For emission measurements, CO and CO₂ were (± 2.0%), HC was (± 3.0%), NO_x were (± 2.5%), and smoke opacity was (± 2.0%). These uncertainty levels are of the same order as those reported in shake-down salt cylinder validation studies on comparable single-cylinder diesel engines, further confirming the validity of experimental results.

Result and discussion

Performance characteristics

Figure 2 shows the variation of brake thermal efficiency (BTE) with engine load for Mahua biodiesel–diesel blends enriched with TiO₂ and CeO₂ nanoparticles. BTE of all the tested fuels indicates an unambiguous manner in showing an increasing trend with a load. This is largely ascribed to the increased combustion efficiency at higher loads, where the in-cylinder temperature is high enough to promote fuel oxidation but the relative contribution of frictional, pumping and heat transfer losses is reduced. Compared to the B20 blend, the lower density (825.8 kg/m³), lower viscosity (3.06 mm²/s), and higher calorific value (44.23 MJ/kg) of neat diesel (D100) shape the better atomisation and promote a faster combustion that results in higher BTE. The B20 blend exhibited a slightly lower BTE, as the viscosity of Mahua biodiesel is higher and has a lower heating value (39.38 MJ/kg), which leads to retarded fuel–air mixing and incomplete combustion. On the other hand, the dispersion of TiO₂ and CeO₂ nanoparticles considerably increases BTE at all loads due to its catalytic oxidation and oxygen availability during combustion. Out of the fuels performed, B20T50 could maximise BTE to near 30.5% at full load, which is far above neat diesel operation. The enhanced performance is due to the excellent oxygen release and catalytic performance of TiO₂ nanoparticles. There is a small decrease in BTE at higher concentrations of nanoparticles (75 ppm) due to agglomeration or enhanced viscosity. At full load, B20T50 improved BTE by approximately 6–8% compared to neat diesel and by about 9–10% compared to the base B20 blend. A similar trend is observed for CeO₂ composites; B20C50 presents BTE values close to B20T50 due to the oxygen storage and redox behaviour (Ce⁴⁺/Ce³⁺) of CeO₂.

Fig. 2 — Brake thermal efficiency (%) vs. engine load for various blends (%).

The increasing share of part-load losses, mainly friction loss, pumping work and heat transfer losses, has a dominant effect on brake thermal efficiency. Between 0 and 25% load (light load), the brake output is small, so even though the mechanical and pumping losses remain essentially constant, lower BTE values (14–17% for diesel operation) are apparent. More than 40–45% of the released fuel energy is spent on overcoming friction and heat losses under these operating conditions, making thermal utilisation poor. With engine load moving to the medium and high (75–100% load) range, the contribution of these losses is significantly smaller, and BTE increases to the level of approximately 28–32% (for diesel) and 30–34% (for nanoparticle-enhanced Mahua biodiesel blends). Further enhancement in BTE through reduced incomplete oxidation losses is achieved by adding TiO₂ and CeO₂ nanoparticles that increase the combustion efficiency. At higher loads, the BTE of NBC and DBCB20 blends enhanced by ~ 6–8% compared to the neat diesel due to enhanced heat release rates and improved utilisation of oxygenated fuel components. This quantitative analysis validates that the BTE increase as observed is primarily due to the decreased relative effect of part-load losses and the enhanced combustion properties of nanoparticle-assisted biodiesel fuels.

Overall, the optimum concentration of nanoparticles would seem to be in the region of 50 ppm for both TiO₂ and CeO₂. The performance trend can be seen as follows: The performance trend shows that Diesel < B20 < B20T25 < B20T50 ≈ B20C50 > B20T75 > B20C75, confirming that the addition of nanoparticles leads to better thermal efficiency.

The effect of engine load on brake-specific fuel consumption (BSFC) for all tested fuels is plotted in Fig. 3. This typical trend of reduction in BSFC with respect to increasing load is the characteristic feature of CI engines. The combination of incomplete combustion and the relatively larger heat transfer and frictional losses that occur at low loads leads to higher values of BSFC. At higher engine loads, the combustion is more efficient, and friction losses form a smaller percentage of the total output power, resulting in decreased specific fuel consumption. Neat diesel (D100) shows the least BSFC over all load conditions due to its higher calorific value (44.23 MJ/kg) and its lower viscosity (3.06 mm²/s), which aids in finer atomisation and further faster combustion. It employs a B20 blend, indicating a little bit higher BSFC than diesel; that is, Mahua biodiesel has a higher density and viscosity and a lower heating value (39.38 MJ/kg), thus maintaining the same brake power, but a bit higher fuel flow rate is required. Contrasting with this, BSFC was decreased significantly upon using different nanoparticles at the 75% load condition in comparison with the B20 blend due to higher effective combustion performance. Of the nanoparticle-doped fuels, B20T50 provides the lowest BSFC, as it reaches an optimum concentration of nanoparticles that promotes atomisation and oxygen availability during the combustion process. The increase of nanoparticle loading up to 75 ppm shows a slight change in BSFC because there will be agglomeration, and dispersion is not optimised compared to previous cases. At maximum load, BSFC for B20T50 was reduced by approximately 6–7% compared to B20 and by 4–5% compared to neat diesel. This same general pattern holds true for the performance of CeO₂-based blends, where B20C50 demonstrates the best fuel economy stemming from the unique oxygen storage and redox properties of CeO₂.

Emission characteristics

The HC emissions as a function of engine load are illustrated in Fig. 4 for Mahua biodiesel–diesel blends doped with TiO₂ and CeO₂ nanoparticles. Unburned hydrocarbons originate from incomplete combustion caused by wall quenching, flame extinction, and poor oxidation in fuel-rich regions. HC emissions for all fuels tested decrease with increasing load, likely due to the higher in-cylinder temperatures at higher loads, which improve fuel vaporisation, which assists in the more complete oxidation of unburned hydrocarbons. For the HC emissions at low-load conditions, it seems that high HC formation comes from the incomplete combustion, the wall quenching and locally fuel-rich regions. D100 shows relatively higher HC emissions throughout the load range due to insufficient oxidation in the cooler parts of the combustion chamber with neat diesel. Assuming that the Mahua biodiesel has some amount of oxygen in it, the B20 blend demonstrates a slight decrease in HC emissions compared to diesel due to improved combustion efficiency. Nevertheless, the high viscosity and low volatility of the biodiesel resulted in localised incomplete combustion and restricted the level of HC reduction. Furthermore, at all loads, the addition of TiO₂ and CeO₂ nanoparticles into the B20 blend greatly reduces HC emissions. These metal oxides act as oxidation catalyst carriers and regulate oxygen to promote oxidation reactions. B20T50 has the lowest HC emissions among the various blends, likely due to the optimal concentration of nanoparticles added in the initial three blends. When using a higher dosage of nanoparticles (75 ppm), HC emissions actually increase slightly as a consequence of an agglomeration phenomenon that reduces the overall efficiency of dispersion. HC emissions were reduced by approximately 20–25% for B20T50 compared to diesel at higher loads. CeO₂-based fuels show somewhat similar trends, albeit the lowest HC emissions (B20C50) may also relate to redox behaviour (Ce⁴⁺/Ce³⁺) with oxidation of unburned species.

The CO emissions of Mahua biodiesel–diesel blends containing TiO₂ and CeO₂ nanoparticles against engine load are shown in Fig. 5. Carbon monoxide is formed due to incomplete oxidation of fuel under locally rich or low-temperature combustion conditions. CO emissions drop with higher loads for all fuels tested, reflecting an increased combustion efficiency and higher in-cylinder temperatures at high loads that favour CO oxidation to CO₂. Low-load conditions lead to incomplete combustion and availability of insufficient oxygen, resulting in higher CO formation. Although neat diesel (D100) shows lower CO emissions in general, the CO emissions from D100 in the lower load conditions are considerably higher because of the fuelling in fuel-rich areas with unsatisfactory oxidation in cooler areas of the combustion chamber. A slight reduction of CO emissions for the B20 blend with respect to diesel is due to the fact that Mahua biodiesel has considerable oxygen that favours improved oxidation reactions. However, there might still have the limitation of low-load atomisation due to the low volatility and high viscosity of biodiesel, which impedes the full combustion of fuel. Comparison between various operating conditions of the large decrease in CO emissions due to addition of TiO₂ and CeO₂ nanoparticles to B20 blend at (A) B20 blend (B) B20 blend + 30 ppm TiO₂ (C) B20 blend + 30 ppm CeO₂. These metal oxide nanoparticles act as oxidation catalysts and an oxygen buffer to oxidize CO to CO₂ during combustion. For the tested blends, B20T50 has the lowest CO emission, indicating that the nanoparticles have reached the optimal concentration for oxidation. At a 75 ppm dosage of nanoparticles, a very small increase in CO emissions occurs, and this is likely due to agglomeration and reduced dispersion. CO emissions decreased by nearly 25–30% for nanoparticle-enhanced blends relative to D100. For CeO₂ blends, a similar trend is observed, with B20C50 showing the lowest CO emissions due to the redox behaviour (Ce⁴⁺/Ce³⁺), which appears to promote the oxidation kinetics.

The fluctuation of CO₂ emissions with respect to engine load for various Mahua biodiesel–diesel blends loaded with TiO₂ and CeO₂ nanoparticles is shown in Fig. 6. Carbon dioxide represents the final oxidation product of carbonaceous fuels and serves as an indicator of combustion completeness. For all the fuels, the CO₂ emissions also seem to favourably increase along with the load, which is a direct indication of the better combustion completeness at higher in-cylinder temperatures. At higher loads, improved oxidation interactions also encourage the oxidation of intermediate species like CO and unburnt hydrocarbons to CO₂. At part-load conditions, neat diesel (D100) shows relatively lower CO₂ emissions because of incomplete oxidation due to locally fuel-rich regions and lower combustion temperatures. CO₂ emissions with B20 are slightly higher than diesel, which can be explained by the presence of oxygen in Mahua biodiesel, leading to more complete oxidation of carbon-based species. Even though biodiesel viscosity is higher, which can also restrict oxidation efficiency at low loads. When TiO₂ and CeO₂ nanoparticles were added to a B20 blend, the CO₂ emissions increased remarkably under all operating conditions, showing the more complete combustion. The catalytic and oxygen-buffering behaviour of these nanoparticles enable oxidation. The highest CO₂ emissions were produced in B20T50 and B20C50, suggesting that 50 ppm of nanoparticles was the most effectively oxidised among the various fuels studied. However, at higher dosages (75 ppm) the CO₂ rise is negligible, most probably owing to agglomeration effects saturating the optimal value of effective catalytic activity. CO₂ emissions increased by about 5–8% for B20T50 compared to diesel, indicating improved oxidation completeness. Nanoparticle-assisted biodiesel blends have confirmed enhanced combustion effectiveness in an increase in the other mechanisms of CO₂ emissions and a decreased mechanism of incomplete combustion mechanisms.

Fig. 6 — Carbon dioxide (Vol %) vs. Engine load for various blends (%).

The NO_x formation is predominantly governed by the thermal (Zeldovich) mechanism and increases with higher peak combustion temperature and oxygen availability. Variation of NOₓ emissions with engine load for TiO₂ & CeO₂ doped Mahua biodiesel–diesel blends (Fig. 7) NO_x emissions rise with higher engine load, and this increase occurs for all tested fuels, a response largely dictated by greater in-cylinder temperatures and higher nitrogen residence time at exhaust- and combustion-temperature levels in high-load operation. They are conductive in inducing the thermal NO_x formation through the Zeldovich mechanism. D100 displays lower NOₓ emissions than both nano and non-nanoparticle blends at higher loads due to relatively lower peak combustion temperature. The moderate increase of NO_x emission for the B20 blends compared to diesel due to the relatively high cetane number and oxygen content of Mahua biodiesel can help to enhance the combustion reaction and increase combustion temperatures, leading to NO_x formation. At medium and high loads, NO_x emissions are multiplied when loading increases upon the addition of TiO₂ and CeO₂ nanoparticles. This gain is correlated with the efficiency in combustion, enhancement of oxidation reaction, and higher rate of heat release values from catalytic activity of nanoparticles. The B20T50 and B20C50 present the highest NO_x emissions compared to the other concoctions, which is consistent with the combustion characteristics. For higher concentrations of nanoparticles (75 ppm), NO_x increases only marginally, which is likely due to agglomeration preventing further increase of temperature. NO_x emissions increased by approximately 8–12% for nanoparticle-doped blends at full load. In conclusion, the NO_x trend observed truly reflects that the addition of nanoparticles (combustion process/occurring at higher in-cylinder temperatures) affects NO_x formation.

Fig. 7 — Nitrogen oxide (ppm) vs. engine load for various blends (%).

Variation of smoke opacity with load for TiO₂ and CeO₂ nanoparticles added to Mahua biodiesel–diesel blends (Fig. 8). Smoke formation occurs under locally fuel-rich diffusion combustion conditions due to soot nucleation and growth. Consistently for all fuels tested, the increase in smoke opacity with increasing load is caused by larger amounts of fuel injected and locally fuel-rich diffusion combustion zones at higher loads. In contrast, the smoke production is highly dependent on the types of fuel and the addition of nanoparticles. This is because neat diesel (D100) has the highest smoke opacity of all fuels studied at all loads due to the lack of inherent oxygen and the strong tendency to form soot precursors under fuel-rich diffusion-controlled combustion conditions. The oxygen content in Mahua biodiesel contributes to better soot oxidation during the late combustion phase, resulting in a significant reduction in smoke opacity of the B20 blend compared to diesel. However, the increased viscosity of the biodiesel limits spray atomisation, thus slightly hampering additional smoke reduction. At all loads, the dosage of TiO₂ and CeO₂ nanoparticles into the B20 blend are associated with a significant reduction in smoke opacity. The catalytic activity and promoted oxygen supply of these nanoparticles can increase soot oxidation, inhibit soot nucleation and promote carbonaceous particle burnout. B20T50 and B20C50 yield the lowest potential smoke opacity among all blends, which suggests a perfect nanoparticle mixing ratio. At high nanoparticle loading (75 ppm), a greater reduction in smoke is observed, but the reduction is small due to agglomeration to reduce effective catalytic surface area. Smoke opacity was reduced by approximately 35–40% compared to diesel at high load. Overall, using nanoparticles in biodiesel combustion helps to boost the oxidation of soot and therefore, reduces the amount of smoke emitted.

It was found from the experiment that TiO2 and CeO2 nanoparticles enhanced the engine performance as well as reduced emissions behaviour of Mahua biodiesel–diesel blends. The BTE rose with load, except for higher in which B20T50 and B20C50 at highest value due to better combustion. Similarly, BSFC dropped as evidence of improved fuel economy. The added nanoparticles significantly decreased HC and CO emissions, but slightly increased the CO₂ level indicating better combustion. While NOₓ emission slightly increased with the higher concentration, CeO₂ inhibited effectively this worsening. The soot opacity reduced significantly for both types of nanoparticles; 50 ppm resulted in the best system performance and emissions behaviour.

Combustion characteristics

The variation of in-cylinder pressure with crank angle for Mahua biodiesel–diesel blends enriched with TiO₂ and CeO₂ nanoparticles at full load is shown in Fig. 9. The initial rapid combustion process of neat diesel (D100) during the premixed phase gives it a higher calorific value and volatility, manifesting as a relatively high peak pressure. Compared with B100, the peak pressure of the B20 blend steadily decreased in small amounts, which could be attributed to the high viscosity and low heating value of Mahua biodiesel that delays fuel–air mixing and lowers the premixed combustion intensity. The incorporation of different TiO₂ and CeO₂ nanoparticles into the B20 blend has a significant effect on the peak in-cylinder pressure. The reason for this enhancement is said to be due to the better combustion kinetics and better atomisation of the fuel, as well as the promotion of faster oxidation reactions through the presence of the catalytic effect from the nanoparticles. B20T50 and B20C50 among the tested fuels show the highest peak pressures, which indicates that they have better heat release near top dead centre. At 75 ppm, peak pressure is marginally lower than that at 50 ppm, which is most probably a consequence of aggregation of the nanoparticles, which restricts dispersion and retards catalytic effectiveness. In conclusion, the results of the ICP together support the previous research stating that the addition of nanoparticles enhances the combustion intensity and the pressure rise.

Fig. 9 — In-cylinder pressure rise (bar) vs. crank angle (deg.).

D100 < B20 < B20T25 < B20C25 < B20T50 ≈ B20C50 > B20T75 > B20C75.

This proves that 50 ppm is the optimal dosage of nanoparticle for improving the combustion efficiency of Mahua biodiesel–diesel blends.

The heat release rate profiles for the eight tested fuels at full load are presented in Fig. 10. Due to fast evaporation and ignition right after injection from the nozzle, neat diesel presents a very pronounced premixed combustion peak. The lower premixed peak of B20 can be explained, as biodiesel has a high viscosity leading to lower fuel availability (due to delayed atomisation) during the premixed stage. The premixed heat release rate of the B20 blend is greatly improved when TiO₂ and CeO₂ nanoparticles are added. Such behaviour is linked to better ignition property, higher access of reactive oxygen species and catalytic promoting oxidation reactions. Among the blends investigated in this work, the combustion phasing within B20T50 and B20C50 is characterised by the highest values of premixed Heat Release Rate (HRR) peak, suggesting that energy is released at a higher rate compared to other blends. The premixed HRR peak is slightly lower at the higher nanoparticle concentration (75 ppm), which indicates that agglomeration may inhibit interaction between the nanoparticles and the fuel droplets. For the nanoparticle-doped fuels, the increase of HRR is responsible for the improvements in BTE and the decrease of BSFC, CO, HC and smoke emissions, and it contributes to the increase of NOₓ emissions associated with higher temperatures of combustion.

Machine learning

Machine learning model performance evaluation

While experimental studies lay the groundwork for understanding engine performance, the performance and emission characteristics associated with combined changes in engine load, biodiesel fraction and nanoparticle additives are highly coupled and nonlinear. Towards this end, it is difficult to create robust predictive relationships using more traditional empirical models, especially with multiple outputs. Machine learning techniques provide a data-driven method for this capture of complex relationships, without restrictive assumptions regarding functional form. In this study, machine learning is utilised to assess the performance of a host of regression models in estimating the performance and emission parameters of CI engines from experimentally generated data. The ML framework is meant to augment the experimental analysis by enabling accurate multi-output prediction, enhanced generalisation across all operating conditions, and aiding future studies of engine optimisation and calibration, particularly when experimental data are scarce.

The measured data for machine learning modelling were 45 in total: nine fuel blends and five engine loads. All data points were obtained by averaging repeated experimental results to improve the reproducibility and reliability. Due to the empirical statistical nature of engine testing, the size of our dataset is susceptible to a small-sample regime (common in engine-based ML studies). The purpose of the ML analysis is hence to compare models and make predictions that generalize within a controlled experimental context, not large-scale empirically based learning.

The predictive capability of six machine learning regression algorithms [Ridge Regression, SVR (Support Vector Regression), Random Forest, Gradient Boosting, XGBoost and CatBoost] was compared using the four standard statistical measures: R² (coefficient of determination), RMSE (Root Mean Square Error), MAE (Mean Absolute Error) and MSE (mean square error). These parameters together demonstrate the fit, adequacy and general performance of each model in predicting engine performance and emission response for TiO₂/CeO₂ nanoparticles with Mahua biodiesel–diesel blends. The overall Model Comparison is shown in Table 3.

Table 3.

Overall model comparison.

Model	R²↑	RMSE ↓	MAE ↓	MSE ↓	Performance rating
Ridge regression	0.8359	34.03	13.52	1158.3	Moderate
SVR	0.6168	57.2	14.22	3271.79	Poor
Random forest	0.9016	10.74	3.51	115.45	Excellent
Gradient boosting	0.8319	10.3	4.04	106.06	Moderate
XGBoost	0.9265 (Highest)	10.59	3.93	112.35	Best Overall
CatBoost	0.9247	49.96	8.88	2495.6	Poor (overfitting / unstable)

Open in a new tab

Significant values are in bold.

The overall prediction accuracies indicate that XGBoost outperformed all the other methods in general (R² = 0.9265), which represents a powerful ability for modelling of complex nonlinear relationships between blend composition, engine load and produced output responses. Random Forest also achieved good results (R² = 0.9016), and it had low results for RMSE and MAE too, so this confirms its strength and reduction of prediction errors. These two ensemble-based models take advantage of decision trees with randomness and multiple features and are better at capturing variations in the combustion process and emissions³⁰.

Table 4 shows the Parameter-Wise Best Model Summary for each engine output parameters. From the Table 4, it is interesting that the performance of ridge regression and gradient boosting was also moderate (R² = 0.8359 and R² = 0.8319, respectively). Ridge regression attenuated overfitting, but it did not have the nonlinear flexibility to account for combustion-dependent changes. Meanwhile, in the specific output case, such as NOₓ prediction, gradient boosting demonstrated good performance with the highest R² (0.9970). This indicates that gradient boosting is sensitive to temperature-dependent reaction kinetics parameters.

Table 4.

Parameter-wise best model summary.

Output parameter	Best Model	R²	Remark
BTE	SVR	0.9933	Extremely accurate for efficiency prediction
BSFC	XGBoost	0.9806	Most accurate fuel consumption predictor
HC	SVR / CatBoost	0.9298 / 0.8892	SVR best for unburned hydrocarbons
CO	CatBoost	0.9076	Best for low-concentration pollutant behaviour
NOₓ	Gradient Boosting	0.9970 (Highest)	Best at capturing combustion temperature influence
Smoke	SVR	0.9982 (Highest)	Excellent at tracking soot reduction trends
CO₂	XGBoost	0.9969	Very high correlation with combustion completeness

Open in a new tab

The performance of the Support Vector Regression (SVR) model was mixed. Also, with lower overall R² (0.6168), SVR was the most accurate method for predicting BTE (R² = 0.9933), smoke (R² = 0.9982) and CO₂ (R² = 0.9954). This shows that SVR is successful at modelling nonlinear correlation between the efficiency-related and combustion-completeness indices. Its relatively high RMSE and MSE in other parameters also suggest a sensitivity to data rescaling/scaling and variability. CatBoost had inconsistent performance, presenting high accuracy on the CO prediction part of its model (R² = 0.9076); however, it was much higher for both RMSE and MSE in total predictions due to an overfit or feature distribution sensibility. This implies that CatBoost may need a bigger data set or more variety in the categorical feature to give stable output²⁸.

To sum up, the comparative results show that XGBoost is the best model in terms of general prediction; SVR has better prediction performance for BTE, CO₂ and smoke; and gradient boosting predicts NOₓ more accurately. Accordingly, one should seek a hybrid or output-specific model selection instead of a one-model-fits-all solution.

Actual vs. predicted

Brake thermal efficiency

The actual vs. predicted plots for BTE are in shown Fig. 11. Of all the models tested, SVR shows the best prediction results, with an R² value of 0.9933, which indicates a nearly perfect correlation between BTE predicted and BTE actual values. The near-overlap between data points and the ideal-fit line demonstrates a capability of the model in representing non-linear combustion efficiency with particle-enhanced fuel blends. XGBoost and Random Forest also present good prediction results, with R² values of 0.9545 and 0.9559, suggesting their good generalisation capacity to learn combustion-relevant patterns. The prediction error of the Ridge Regression and the Gradient Boosting technique is slightly higher; the accuracy of CatBoost is quite acceptable, but we can see that it has a little deviation in high load conditions. The higher BTEs seen experimentally in Pant et al.‘s model of B20 with TiO₂ and CeO₂ additions have been captured correctly by all models, indicating that the role of catalytic oxygen supply, promoting micro-explosion activity and better fuel–air mixing, is properly predicted for nanoparticle-enhanced blends. In general, the comparison shows that ML techniques can accurately predict BTE, with SVR being the best model for this response variable.

Brake specific fuel consumption

As shown in BSFC Fig. 12, the BSFC prediction of XGBoost is relatively more accurate (R²=0.9806) and better with both RMSE and MAE than other model groups. This advantage is because it better captures the inverse association between combustion efficiency and fuel consumption. Both Random Forest and Gradient Boosting have high correlation between predicted response and true BSFC, which indicates that the ensemble data captures trends in the fuel usage mechanism. CatBoost provides fair accuracy; Ridge Regression is acceptable with slightly higher variability at mid-level load. The performance of SVR, on the other hand, is poor: R² is negative (this is not possible). This outcome indicates that it does not fit trends of BSFC as a result of the complicated correlation between calorific value, ignition delay and heat release characteristics. Empirically, the BSFC is seen to reduce with nanoparticle addition (especially at 50 ppm), which is mainly due to better combustion efficiency and decreased unburned fuel losses. This behaviour is well clarified by the best models, which strengthens that the higher atomisation and catalytic oxidation advanced by nanoparticles leads to a more effective fuel utilisation. Therefore, XGBoost is the best model for BSFC prediction with higher accuracy and robustness in a wide load range.

Hydrocarbon emissions

The model-wise data of actual vs. predicted HC also indicate considerable variation, as shown in Fig. 13. These two models (SVR and CatBoost) present the best prediction capability, with R² values of 0.9298 and 0.8892, respectively, which means a good fit with experimental tendencies to actual emissions. These kinematic models successfully account for the decrease of the HC emissions obtained with the addition of TiO₂ and CeO₂ nanoparticles as a function of improved oxidation rates and higher-level combustion completeness. On the other hand, ridge regression, random forest, and gradient boosting have higher values of RMSE and MAE, showing that it is more difficult to model the presence of unburned hydrocarbons, which become more sensitive. HC emissions are reported to decrease experimentally when burning nanoparticulate-doped fuels as a result of an increased oxygen scavenging opportunity and catalytic surface oxidation rates for partially burned hydrocarbons. The models that did perform well must have some way to recognize and embed these chemical-kinetic improvements, especially at high loads. However, lower-performing models are likely not sensitive to such transient and spatial generative factors for HC since transient combustion inefficiency could not be reproduced. In general, SVR is found to be the best model for HC prediction because it can handle nonlinear combustion and emission variations over different engine operating conditions.

Carbon monoxide emissions

The CO estimation graphs in Fig. 14 illustrate the highest accuracy with CatBoost, observing an R² of 0.9076 and the lowest overall (RMSE; MAE). CatBoost’s gradient boosting on decision trees such that it can accurately learn the small fluctuation of CO emissions due to incomplete combustion at low loads and better oxidation at high loads. XGBoost and Random Forest are also moderately accurate but have higher variability in predictions. Meanwhile, Ridge Regression and Gradient Boosting behave with a medium level of performance; the SVR’s poor prediction of CO can be seen in a negative R² value not adequately capturing the formation pattern. The reduction of CO is due to better oxidation activity with additional nanoparticles and the formation of a higher number of radicals that aid in the conversion of CO into CO₂. Both CatBoost and XGBoost capture these patterns, specifically the rapid fall in CO at medium to high loads. The less successful models may have been confounded by CO susceptibility to air–fuel ratios and local temperature gradients. So CatBoost is the best predictive model for CO, particularly for low-concentration pollutant prediction.

Carbon dioxide emissions

The plots of CO₂ prediction (Fig. 15) are presented, which indicates that the XGBoost and SVR provide high accuracy in terms of R² values as 0.9969 and 0.9954, respectively. Such models correctly capture the increasing trend of CO₂ emissions with load, arising from enhanced combustion completeness as a consequence of oxygen enrichment by nanoparticles. Random Forest and Gradient Boosting also behave well, following relatively tightly the ideal-fit line, while Ridge Regression and CatBoost exhibit bigger variations between lower and higher concentrations of CO₂. On the experimental side, the increasing CO₂ production is a clear indication of more complete combustion, meaning better oxidation behaviour for species containing carbon. The high performance of XGBoost + SVR indicates that the direct combustion terms are learned effectively. Given that CO₂ is highly correlated with combustion stoichiometry and heat release efficiency, the most successful “models” are those that accurately learn correlations between fuel–air mixing and combustion temperature. We report the results, which evidently validate that XGBoost is marginally better and maintains the steadiest fit for all points with less spread in the prediction. Thus, CO₂ can be well predicted by the XGBoost and SVR methods, demonstrating their powerful model ability for combustion completeness indicators.

Nitrogen oxides

For the NO_x predictor, as shown in Fig. 16, gradient boosting performs best among these five models with the maximum R² = 0.9970 and minimum RMSE and MAE. This reveals that the combustion temperature-induced formation of nitrogen oxides can be well modelled by gradient boosting. Models with XGBoost and Random Forest also predict very well with R² = 0.9966 which suggests that they capture thermal NOx formation dominated by in-cylinder temperature and oxygen concentration perfectly fine. On the contrary, both SVR and CatBoost achieve worse performance, which means they are incompetent to approximate steep gradients of NO_x variation at high engine loads. NOₓ generation is extremely sensitive to ignition delay, burn rate, and peak cylinder temperature that are significantly controlled by the catalytic ability of TiO₂ and CeO₂ NPs. Of course, the better models were able to reproduce this strong relationship, whereas poorer ones had difficulty in predicting sharp inclines at full load. Accordingly, gradient boosting is the optimal NOₓ prediction model due to its ideal trade-off between accuracy, robustness, and emission trend representation.

Smoke opacity

The smoke results seen in Fig. 17 for the predictions reveal that SVR fits the best by far (an R² = 0.9982, so it is in almost perfect agreement with the experiment). This gives the possibility to estimate quite accurately the effects of increased combustion/soot oxidation because of CeO₂ and TiO₂ nanoparticle addition. Also, XGBoost and Random Forest demonstrate reasonably high prediction accuracy with a little more spread-out prediction. Meanwhile, Ridge Regression and Gradient Boosting indicate slightly accurate predictions, while CatBoost presents the worst performance in smoke prediction, particularly at low smoke level ranges. Regarding the underlying mechanism of soot reduction, this is mainly because, experimentally, smoke suppression mostly arises from better atomisation and combustion efficiency as well as from active soot oxidation by CeO₂ oxygen storage. The excellent performance of SVR indicates it captures well the nonlinear and localised soot formation dynamics depending on nanoparticle concentration and load. Thus, SVR is the optimal method for predicting smoke emission, and in particular when studying fuels dominated by micro-combustion and catalytic soot oxidation.

Machine learning model validation using LOOCV (XGBoost model)

The results of the Leave-One-Out cross-validation (LOOCV) for the prediction of Brake Thermal Efficiency (BTE), with the LOOCV using the XGBoost model, are illustrated in Fig. 18. The fitted data looked very good across the whole operating range, as shown in the plot of measured vs. expected, where most of the data will fall near the ideal parity line. The model as a whole is very high (R² = 0.9933), so it explains approximately 99.3% of the experimental variability in BTE. The prediction error is very low: RMSE: 0.4293% and MAE: 0.3167%. This indicates that the model is capable of generalisation, even with minimal examples. Figure 18 shows residuals that are scattered in a random and symmetrical manner about the zero-error reference line. This implies that there is no observable pattern towards either low BTE levels or high BTE levels. This indicates no systematic bias; in other words, the model does not overestimate BTE at high loads and underestimate BTE at low loads. Also, the histogram of the residual distribution aligns with the normality test of the residual distribution, adhering closely to a normal distribution and centred at zero. The majority of the residuals are within ± 0.5 percentage points. XGBoost shows promising performance such that brake thermal performance of engine load, fuel oxygenation, and adding nanoparticles to improve combustion efficiency are well correlated with each other.

Fig. 18 — *LOOCV-based actual vs. predicted BTE*, *residual analysis*, *residual distribution*, *and performance metrics using the XGBoost model.*

Results of the XGBoost model for Brake Specific Fuel Consumption (BSFC) LOOCV test are presented in Fig. 19. The plot of residual BSFC vs. predicted BSFC shows a strong linear characteristic, with the predictions closely matched to the optimal fit line over the entire range of BSFC. The model explains 99.6% of the variation in BSFC with a high free coefficient of determination (R² = 0.9960). The RMSE = 0.0044 kg/kWh and MAE = 0.0031 kg/kWh results indicate a very tight clustering and that these predictions are highly consistent and accurate. The graph of residuals versus predicted values demonstrates that the residuals are randomly distributed around the zero line, without any obvious clusters or trends in either direction. This ensures that the prediction will still be correct irrespective of the workload that the system is performing. The mode of the residual distribution histogram exhibits symmetric and closely grouped data centred at zero. About 95% of the residuals from the average are inside of ± 0.01 kg/kWh. This reflects the model’s knowledge of the inverse correlation between BSFC and BTE, primarily attributed to the improved combustion of Bach mixes with addition of nanoadditives. The XGBoost model verified precise and unbiased prediction of BSFC under leave-one-out cross-validation results. This indicates that it will be appropriate for CI engine fuel consumption analysis.

Carbon monoxide (CO) emissions predictions (LOOCV) of the XGBoost model (Fig. 20). The graph overlaying observed and anticipated CO is communicative in that it shows high amounts of consistency, as evidenced through most of the data points remaining close to the line of ideal parity. The R² value from the model is 0.9711, which is quite good. That is, it accounts for over 97% of the variation in CO emissions. With an RMSE of 0.0050% and an MAE of 0.0038%, it can be said the model predicts incomplete combustion behaviour quite well. In the residual vs. fitted plot, we can see that residuals are randomly spread over the zero line and there is no clear pattern in the predicted CO range. The increased variability observed at higher CO levels reflects the sensitivity of CO production to local mixing of fuel and air and the combustion process itself at low load conditions. From the histogram of the residual distribution, we can see that most of the residuals are in the range of ± 0.005%. As a result, these predictions must be stable and unscrewed. The XGBoost model is seen as working reasonably well to indicate how varying conditions of fuel oxygen and nanoparticles can be used, as possible control variables, to enhance oxidation on CO emissions.

The LOOCV test using the XGBoost model as a predictor for hydrocarbon (HC) emissions is presented in Fig. 21. A plot of actual versus predicted values reveals excellent linear correlation, evidenced by an R² of 0.9675. In other words, the predictions are highly predictive for the entire HC emission range. The wider range and the nature of HC emissions (being more sensitive) justify the model having an RMSE of 2.8996 ppm and an MAE of 2.1273 ppm. The residue plot is plotted against the predicted values, and one can see that the residuals are randomly scattered about the zero line. How does it mean that there is no general over- or under-prediction? We have the histogram of the distribution of residuals that is similar to symmetrical and centred around the 0. The spread is a bit wider than that forecasted by the CO, however. This finding is consistent with expectation since the HC emissions are more dispersed due to significant influences from wall quenching, flame extinction and fuel spray characteristics. The challenges indicated reveal that the LOOCV outcomes show that the XGBoost model is accurately capable of finding essential combustion-related factors influencing HC production during NP-enhanced biodiesel combustion.

The test results of the XGBoost model of predicting CO₂ using the LOOCV have been presented in Fig. 22. The difference between the actual and expected values of CO₂ represented in the graph shows that the two values are very close to one another throughout the entire operating range. The predicted values too lie very close to the ideal parity line. The high coefficient of determination (R² = 0.9981) thus indicates that the model captures almost 99.8% of the variation in the CO₂ emission changes observed during the experiments. It can see that the performances of the predictions are quite good by looking at the calculated error metrics (RMSE = 0.1414 vol % and MAE = 0.0978 vol %), which shows that the model performs well for various engine load and fuel blend conditions. Residuals vs. Expected Values ( Fig. 22) This plot shows that the residuals are homogeneously distributed across the zero-error line. In other words, the projected range of CO₂ shows little systematic bias, if any at all. The histogram of the residual distribution indicates that almost all the residuals are between ± 0.2 vol %, which implies that the forecasts are stable and not skewed. The results reveal that the XGBoost model performs well in explaining the impact of high oxidation reactions, higher fuel oxygen levels, and combustion-supporting nanoparticles on CO₂ emissions. The present work demonstrated a strong and reliable model with respect to predicting CO₂ emissions from CI engines operating on Mahua biodiesel doped with nanoparticles as validated through the LOOCV study.

Results are shown in Fig. 23, in which the LOOCV test of predicted nitrogen oxides (NOx) emissions from the XGBoost model is shown. The plot that compares the actual and expected NOx has a perfect fit (R² value of 0.9980). That is almost all the variability of the experiment explained! Error metrics (RMSE = 23.4461 ppm, MAE = 17.6007 ppm) 4 show low values compared to the overall NOx level. That suggests that the forecasts are quite accurate. Residual diagnostics reveal a roughly symmetric distribution around zero, with a slight increase in variance for high levels of NOx in high-load situations. This result indicates that temperature does not always have a direct effect on thermal NOx production, making the relationship more complex. Most of the residual’s approach zero, which indicates no systematic bias, as observed from the histogram of residual distribution. Visible from the LOOCV results, the XGBoost model is capable of illustrating how the interrelations of temperature within the cylinder, available oxygen, and the impact of nanoparticles on combustion all act in concert to reduce NOx emissions.

The smoke opacity predictions of the XGBoost model from the LOOCV method are illustrated in Fig. 24. This, as expected, such a high R² value (R²=0.9949) from the graph that represents how the actual values are linearly related with the predicted values, confirms we are making good predictions. An evaluation of the model predictions on how soot is developing within the various operating conditions shows that the predictions are highly accurate (RMSE = 1.5172 HSU, MAE = 1.2023 HSU). The residuals in the residual-versus-predicted plot are uniformly distributed around zero with no striking structure. There is no evident bias in the predictions along the smoke spectrum. This could be due to the catalytic nature of nanoparticles, which may temporarily alter the behaviour of how soot oxidises at these elevated levels of smoke. The residual histogram is centred on zero, which means the predictions were closer to the real values over time. The LOOCV results suggest that the XGBoost model explains the complex interactions between fuel composition, nanoparticle additives and soot production in CI engines well.

Conclusion

Present study analysed the performance, emission and combustion behaviour of Mahua biodiesel-diesel blends with the addition of TiO₂ and CeO₂ nanoparticles, for which a machine learning model was also used to forecast engine outputs. Following observations are made.

The experimental study validated that Mahua biodiesel-diesel blends supplemented with TiO₂ and CeO₂ nanoparticles have a substantial impact on engine efficiency, emissions, and combustion characteristics. The brake thermal efficiency (BTE) improved with engine load for all the fuels, and the nanoparticle-doped blends outperformed the base B20 fuel. Integrating B20T50 brought about the highest brake thermal efficiency (BTE) of around 30.5% at full load, a 6–8% enhancement in comparison with neat diesel.
Load-dependent BSFC All the fuels achieved a decline in BSFC with elevated load. The presence of these nanoparticles led to a decline in BSFC by a maximum of 6–7% with respect to the B20 blend, where the B20C50 and B20T50 reported the minimum values of BSFC because of combustion enhancement and better oxidation kinetics.
Emission analysis revealed significant reductions in products indicating incomplete combustion. Relative to diesel operation, CO and HC emissions were approximately 25–30% and 20–25% lower for nanoparticle-enhanced blends, respectively, while smoke opacity was reduced by up to 35–40%, confirming improved soot oxidation. The moderate increase in CO₂ emissions suggests a greater degree of combustion completeness.
At higher loads, nanoparticle-assisted blends produced roughly 8–12% more NO_x emissions than the corresponding level of the base fuel, which can be associated with higher in-cylinder temperature and better combustion heat release according to mechanisms of thermal NO_x formation.
The benefits of nanoparticle-doped fuels were also observed with combustion analysis, where the peak in-cylinder pressure was higher and the premixed heat release rates were higher than for the baseline fuels. Positive effects on performance and decreases in BSFC and soot-associated emissions were attributed to higher peak pressures and HRR from B20T50 and B20C50, leading to the performance peaks.
Machine learning analysis revealed that the XGBoost was the best model in predicting CI engine outcomes of interest. The out-of-bag prediction for performance parameters produced high R² values (> 0.99) for performance parameters and > 0.97 for emissions, with low RMSE and unbiased residual distributions according to the Leave-One-Out Cross-Validation.

In summary, the results suggest that a lower concentration of nanoparticles is more beneficial, and a 50-ppm concentration of TiO₂ and CeO₂ nanoparticles provides a better balance in terms of improvement in efficiency and emission control. Thus, nanoparticle-assisted Mahua biodiesel can be considered as a potential alternative fuel for CI engines.

Future work

The study could be extended in the future by testing other dual or hybrid combinations of nanoparticles and optimising injection timings and EGR strategies to reduce NOx while maintaining efficiency. Furthermore, real-time embedded ML prediction and CFD-based combustion simulation could be coupled for more enhanced predictivity and less need for heavy engine testing.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1^{(29.9KB, xlsx)}

Abbreviations

CI: Compression ignition
BTE: Brake thermal efficiency
BSFC: Brake specific fuel consumption
HRR: Heat release rate
CA: Crank angle
ppm: Parts per million
TiO₂: Titanium dioxide nanoparticles
CeO₂: Cerium oxide nanoparticles
MME: MahZua methyl ester (Mahua Biodiesel)
B0: 100% Diesel
B20: 20% Biodiesel + 80% Diesel
B100: 100% Biodiesel
CO: Carbon monoxide
CO₂: Carbon dioxide
HC: Hydrocarbons
NOx: Oxides of nitrogen
HSU: Hartridge smoke unit
ML: Machine learning
SVR: Support vector regression
RF: Random forest regressor
GBR: Gradient boosting regressor
XGBoost: Extreme gradient boosting
CatBoost: Categorical gradient boosting
RMSE: Root mean square error
MAE: Mean absolute error
MSE: Mean square error
R²: Coefficient of determination
DAQ: Data acquisition system

Author contributions

**Durga Venkatesh Janaki: ** Conceptualization, Methodology, Experimental Investigation, Data Curation, Writing – Original Draft Preparation.**P. S. Ranjit: ** Resources, Fuel Preparation Support, Validation, Writing – Review & Editing.**B. Balakrishna1:** Machine Learning Model Development, Data Analysis, Visualization, Software Implementation.**Durga Venkatesh Janaki: ** Supervision, Project Administration, Technical Guidance, Final Manuscript Review and Approval.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1.Padmanaban, J., Kandaswamy, D. A., Joshua, P. J. T. & Natarajan, A. Statistical analysis of engine fueled with two identical lower aromatic biofuel blends at various injection pressure. Energy Sour. Part A Recover. Utilization Environ. Eff.46, 7720–7735 (2024). [Google Scholar]
2.Kshatriya, A. S. et al. Investigations into the combined effect of Mahua biodiesel blends and biogas in a dual fuel engine. Energies (Basel)15, (2022).
3.Kapilan, N. & Prabhu, S. Studies on Mahua oil biodiesel as the sustainable fuel in compression ignition engine. Acta Technol. Agricult.. 25, 85–91 (2022). [Google Scholar]
4.Natarajan, A., Kandasamy, A., Perumal Venkatesan, E. & Saleel, C. A. Experimental investigation on the effect of graphene oxide in higher alcohol blends and optimization of injection timing using an ANN method. ACS Omega. 8, 41339–41355 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Simhadri, K., Rao, P. S. & Paswan, M. Improving the combustion and emission performance of a diesel engine with TiO2 nanoparticle blended Mahua biodiesel at different injection pressures. Inter. J. Thermofluids21, (2024).
6.Jit Sarma, C. et al. Improving the combustion and emission performance of a diesel engine powered with Mahua biodiesel and TiO2 nanoparticles additive. Alexandria Eng. J.72, 387–398 (2023). [Google Scholar]
7.Elumalai, P. V. et al. Artificial neural networks model for predicting the behavior of different injection pressure characteristics powered by blend of biofuel-nano emulsion. Energy Sci. Eng10, (2022).
8.Joshua, P. J. T., Kandasamy, A., Venkatesan, E. P. & Saleel, C. A. Experimental study on sustainability involving the Pugh matrix on emission values of High-Temperature air in the premixed charged compression ignition engine. ACS Omega. 8, 41243–41257 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Soudagar, M. E. M. et al. Study of diesel engine characteristics by adding nanosized zinc oxide and diethyl ether additives in Mahua biodiesel–diesel fuel blend. Sci. Rep.10, (2020). [DOI] [PMC free article] [PubMed]
10.Aalam, C. S. & Saravanan, C. G. Effects of nano metal oxide blended Mahua biodiesel on CRDI diesel engine. Ain. Shams Eng. J.8, 689–696 (2017). [Google Scholar]
11.Janaki, D. V., Ranjit, P. S. & Balakrishna, B. Impact of magnesium oxide nanoparticles and hydrogen enrichment on CI engine performance with Mahua oil biodiesel using machine learning. Heat Mass. Transfer/Waerme- Und Stoffuebertragung61, (2025).
12.Sayyed, S., Das, R. K. & Kulkarni, K. Performance assessment of multiple biodiesel blended diesel engine and nox modeling using ANN. Case Stud. Therm. Eng.28, (2021).
13.Dugala, N. S., Goindi, G. S. & Sharma, A. Experimental investigations on the performance and emissions characteristics of dual biodiesel blends on a varying compression ratio diesel engine. SN Appl. Sci.3, (2021).
14.Pakalapati, J., Surakasi, R., Pakalapati, R. & Babu, K. Characterization and Evaluation of Electrical Properties of Corn Oil Biodiesel Blended with TiO2 Nano Additives. vol. 13 (2023).
15.Earnest, Y. A., Raghavan, J., George Roy, A. & Koshy, C. P. R. Study of engine performance and emission characteristics of diesel engine using cerium oxide nanoparticles blended orange Peel oil Methyl ester. Energy Nexus8, (2022).
16.Gad, M. S., Abdel Aziz, M. M. & Kayed, H. Impact of different nano additives on performance, combustion, emissions and exergetic analysis of a diesel engine using waste cooking oil biodiesel. Propuls. Power Res.11, 209–223 (2022). [Google Scholar]
17.Khan, T. M. Y. Direct transesterification for biodiesel production and testing the engine for performance and emissions run on biodiesel-diesel-nano blends. Nanomaterials11, 1–20 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Jain, A. et al. Impact of titanium dioxide (TiO2) nanoparticles addition in Eichhornia crassipes biodiesel used to fuel compression ignition engine at variable injection pressure. Case Stud. Therm. Eng.49, (2023).
19.Yusuf, A. A., Inambao, F. L. & Ampah, J. D. The effect of biodiesel and CeO2 nanoparticle blends on CRDI diesel engine: A special focus on combustion, particle number, PM2.5 species, organic compound and gaseous emissions. J. King Saud Univ. - Eng. Sci.10.1016/j.jksues.2021.12.003 (2022). [Google Scholar]
20.Kumar, S. & Dinesha, P. Combined influence of fuel injection strategy and nanoparticle additives on the performance and emission characteristics of a biodiesel fueled engine. Cogent Eng.8, (2021).
21.Madhuri, R. J., Naik, J. V. & Roa, V. Effects of Aloe Vera biodiesel blends with TiO2 on the emission and performance characteristics of a DI diesel engine. Biotechnol. J. Int.27, 1–7 (2023). [Google Scholar]
22.Do, Q. H., Lo, S. K. & Chen, J. F. A comparative study of machine learning techniques in prediction of exhaust emissions and performance of a diesel engine fuelled with biodiesel blends. Nat. Environ. Pollut. Technol.20, 865–874 (2021). [Google Scholar]
23.Awogbemi, O., Kallon, D. V. & Von Application of machine learning technologies in biodiesel production process—A review. Front. Energy Res. vol. 11 10.3389/fenrg.2023.1122638 (2023).
24.Magesh, N. et al. Experimental Investigation and Prediction of Performance, Combustion, and Emission Features of a Diesel Engine Fuelled with Pumpkin-Maize Biodiesel using Different Machine Learning Algorithms. Math. Probl. Eng. (2022).
25.Şahin, S. & Torun, A. Comparison of engine performance and emission values of biodiesel obtained from waste pumpkin seeds with machine learning. Agriculture (Switzerland)14, (2024).
26.Ramachandran, E. et al. Prediction of RCCI combustion fueled with CNG and algal biodiesel to sustain efficient diesel engines using machine learning techniques. Case Stud. Therm. Eng.51, (2023).
27.Zandie, M., Ng, H. K., Gan, S., Muhamad Said, M. F. & Cheng, X. Multi-input multi-output machine learning predictive model for engine performance and stability, emissions, combustion and ignition characteristics of diesel-biodiesel-gasoline blends. Energy262, (2023).
28.Şahin, S. Comparison of machine learning algorithms for predicting diesel/biodiesel/iso-pentanol blend engine performance and emissions. Heliyon9, (2023). [DOI] [PMC free article] [PubMed]
29.Ramalingam, K. et al. An evaluation of maximizing production and usage of biofuel by machine learning and experimental approach. Sci. Rep.15, 33265 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Karunamurthy, K. et al. Prediction of IC engine performance and emission parameters using machine learning: A review. J. Therm. Anal. Calorim.148, 3155–3177 (2023). [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Material 1^{(29.9KB, xlsx)}

Data Availability Statement

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

[CR1] 1.Padmanaban, J., Kandaswamy, D. A., Joshua, P. J. T. & Natarajan, A. Statistical analysis of engine fueled with two identical lower aromatic biofuel blends at various injection pressure. Energy Sour. Part A Recover. Utilization Environ. Eff.46, 7720–7735 (2024). [Google Scholar]

[CR2] 2.Kshatriya, A. S. et al. Investigations into the combined effect of Mahua biodiesel blends and biogas in a dual fuel engine. Energies (Basel)15, (2022).

[CR3] 3.Kapilan, N. & Prabhu, S. Studies on Mahua oil biodiesel as the sustainable fuel in compression ignition engine. Acta Technol. Agricult.. 25, 85–91 (2022). [Google Scholar]

[CR4] 4.Natarajan, A., Kandasamy, A., Perumal Venkatesan, E. & Saleel, C. A. Experimental investigation on the effect of graphene oxide in higher alcohol blends and optimization of injection timing using an ANN method. ACS Omega. 8, 41339–41355 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Simhadri, K., Rao, P. S. & Paswan, M. Improving the combustion and emission performance of a diesel engine with TiO2 nanoparticle blended Mahua biodiesel at different injection pressures. Inter. J. Thermofluids21, (2024).

[CR6] 6.Jit Sarma, C. et al. Improving the combustion and emission performance of a diesel engine powered with Mahua biodiesel and TiO2 nanoparticles additive. Alexandria Eng. J.72, 387–398 (2023). [Google Scholar]

[CR7] 7.Elumalai, P. V. et al. Artificial neural networks model for predicting the behavior of different injection pressure characteristics powered by blend of biofuel-nano emulsion. Energy Sci. Eng10, (2022).

[CR8] 8.Joshua, P. J. T., Kandasamy, A., Venkatesan, E. P. & Saleel, C. A. Experimental study on sustainability involving the Pugh matrix on emission values of High-Temperature air in the premixed charged compression ignition engine. ACS Omega. 8, 41243–41257 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Soudagar, M. E. M. et al. Study of diesel engine characteristics by adding nanosized zinc oxide and diethyl ether additives in Mahua biodiesel–diesel fuel blend. Sci. Rep.10, (2020). [DOI] [PMC free article] [PubMed]

[CR10] 10.Aalam, C. S. & Saravanan, C. G. Effects of nano metal oxide blended Mahua biodiesel on CRDI diesel engine. Ain. Shams Eng. J.8, 689–696 (2017). [Google Scholar]

[CR11] 11.Janaki, D. V., Ranjit, P. S. & Balakrishna, B. Impact of magnesium oxide nanoparticles and hydrogen enrichment on CI engine performance with Mahua oil biodiesel using machine learning. Heat Mass. Transfer/Waerme- Und Stoffuebertragung61, (2025).

[CR12] 12.Sayyed, S., Das, R. K. & Kulkarni, K. Performance assessment of multiple biodiesel blended diesel engine and nox modeling using ANN. Case Stud. Therm. Eng.28, (2021).

[CR13] 13.Dugala, N. S., Goindi, G. S. & Sharma, A. Experimental investigations on the performance and emissions characteristics of dual biodiesel blends on a varying compression ratio diesel engine. SN Appl. Sci.3, (2021).

[CR14] 14.Pakalapati, J., Surakasi, R., Pakalapati, R. & Babu, K. Characterization and Evaluation of Electrical Properties of Corn Oil Biodiesel Blended with TiO2 Nano Additives. vol. 13 (2023).

[CR15] 15.Earnest, Y. A., Raghavan, J., George Roy, A. & Koshy, C. P. R. Study of engine performance and emission characteristics of diesel engine using cerium oxide nanoparticles blended orange Peel oil Methyl ester. Energy Nexus8, (2022).

[CR16] 16.Gad, M. S., Abdel Aziz, M. M. & Kayed, H. Impact of different nano additives on performance, combustion, emissions and exergetic analysis of a diesel engine using waste cooking oil biodiesel. Propuls. Power Res.11, 209–223 (2022). [Google Scholar]

[CR17] 17.Khan, T. M. Y. Direct transesterification for biodiesel production and testing the engine for performance and emissions run on biodiesel-diesel-nano blends. Nanomaterials11, 1–20 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Jain, A. et al. Impact of titanium dioxide (TiO2) nanoparticles addition in Eichhornia crassipes biodiesel used to fuel compression ignition engine at variable injection pressure. Case Stud. Therm. Eng.49, (2023).

[CR19] 19.Yusuf, A. A., Inambao, F. L. & Ampah, J. D. The effect of biodiesel and CeO2 nanoparticle blends on CRDI diesel engine: A special focus on combustion, particle number, PM2.5 species, organic compound and gaseous emissions. J. King Saud Univ. - Eng. Sci.10.1016/j.jksues.2021.12.003 (2022). [Google Scholar]

[CR20] 20.Kumar, S. & Dinesha, P. Combined influence of fuel injection strategy and nanoparticle additives on the performance and emission characteristics of a biodiesel fueled engine. Cogent Eng.8, (2021).

[CR21] 21.Madhuri, R. J., Naik, J. V. & Roa, V. Effects of Aloe Vera biodiesel blends with TiO2 on the emission and performance characteristics of a DI diesel engine. Biotechnol. J. Int.27, 1–7 (2023). [Google Scholar]

[CR22] 22.Do, Q. H., Lo, S. K. & Chen, J. F. A comparative study of machine learning techniques in prediction of exhaust emissions and performance of a diesel engine fuelled with biodiesel blends. Nat. Environ. Pollut. Technol.20, 865–874 (2021). [Google Scholar]

[CR23] 23.Awogbemi, O., Kallon, D. V. & Von Application of machine learning technologies in biodiesel production process—A review. Front. Energy Res. vol. 11 10.3389/fenrg.2023.1122638 (2023).

[CR24] 24.Magesh, N. et al. Experimental Investigation and Prediction of Performance, Combustion, and Emission Features of a Diesel Engine Fuelled with Pumpkin-Maize Biodiesel using Different Machine Learning Algorithms. Math. Probl. Eng. (2022).

[CR25] 25.Şahin, S. & Torun, A. Comparison of engine performance and emission values of biodiesel obtained from waste pumpkin seeds with machine learning. Agriculture (Switzerland)14, (2024).

[CR26] 26.Ramachandran, E. et al. Prediction of RCCI combustion fueled with CNG and algal biodiesel to sustain efficient diesel engines using machine learning techniques. Case Stud. Therm. Eng.51, (2023).

[CR27] 27.Zandie, M., Ng, H. K., Gan, S., Muhamad Said, M. F. & Cheng, X. Multi-input multi-output machine learning predictive model for engine performance and stability, emissions, combustion and ignition characteristics of diesel-biodiesel-gasoline blends. Energy262, (2023).

[CR28] 28.Şahin, S. Comparison of machine learning algorithms for predicting diesel/biodiesel/iso-pentanol blend engine performance and emissions. Heliyon9, (2023). [DOI] [PMC free article] [PubMed]

[CR29] 29.Ramalingam, K. et al. An evaluation of maximizing production and usage of biofuel by machine learning and experimental approach. Sci. Rep.15, 33265 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR30] 30.Karunamurthy, K. et al. Prediction of IC engine performance and emission parameters using machine learning: A review. J. Therm. Anal. Calorim.148, 3155–3177 (2023). [Google Scholar]

PERMALINK

Performance emission and combustion characteristics of TIO₂ and CEO₂ nanoparticle enhanced Mahua biodiesel diesel blends using experimental and machine learning approaches

Venkatesh Janaki

P S Ranjit

B Balakrishna

Abstract

Supplementary Information

Introduction

Methodology

Fuel preparation

Table 1.

Fig. 1.

Engine test setup

Table 2.

Experimental procedure

Machine learning-based prediction

Hyperparameter tuning strategy

Model validation and error analysis

Uncertainty analysis

Result and discussion

Performance characteristics

Fig. 2.

Fig. 3.

Emission characteristics

Fig. 4.

Fig. 5.

Fig. 6.

Fig. 7.

Fig. 8.

Combustion characteristics

Fig. 9.

Fig. 10.

Machine learning

Machine learning model performance evaluation

Table 3.

Table 4.

Actual vs. predicted

Brake thermal efficiency

Fig. 11.

Brake specific fuel consumption

Fig. 12.

Hydrocarbon emissions

Fig. 13.

Carbon monoxide emissions

Fig. 14.

Carbon dioxide emissions

Fig. 15.

Nitrogen oxides

Fig. 16.

Smoke opacity

Fig. 17.

Machine learning model validation using LOOCV (XGBoost model)

Fig. 18.

Fig. 19.

Fig. 20.

Fig. 21.

Fig. 22.

Fig. 23.

Fig. 24.

Conclusion

Future work

Supplementary Information

Abbreviations

Author contributions

Data availability

Declarations

Competing interests

Footnotes

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases