Determination of Acidity of Edible Oils for Renewable Fuels Using Experimental and Digitally Blended Mid-Infrared Spectra

Abstract

Renewable fuels produced from animal- and plant-based edible oils have emerged as an alternative to oil and natural gas. Burgeoning interest in renewables can be attributed to the rapid depletion of fossil fuels caused by the global energy demand and the environmental advantages of renewables, specifically reduced emissions of greenhouse gases. An important property of the feedstock that is crucial for the conversion of edible oils to renewable fuels is the total acid number (TAN), as even a small increase in TAN for the feedstock can lead to corrosion of the catalyst in the refining process. Currently, the TAN is determined by potentiometric titration, which is time-consuming, expensive, and requires the preparation of reagents. As part of an effort to promote the use of renewable fuels, a partial least squares regression method with orthogonal signal correction to remove spectral information related to the sample background was developed to determine the TAN from the mid-infrared (IR) spectra of the feedstock. Digitally blended mid-IR spectral data were generated to fill in regions of the PLS calibration where there were very few samples. By combining experimental and digitally blended mid-IR spectral data to ensure adequate sample representation in all regions of the spectra–property calibration and better understand the spectra–property relationship through the identification of sample outliers in the original data that can be difficult to detect because of swamping, a PLS regression model for TAN (R² = 0.992, cross-validated root mean square error = 0.468, and bias = 0.0036) has been developed from 118 experimental and digitally blended mid-IR spectra of commercial feedstock. Thus, feedstock whose TAN value is too high for refining can be flagged using the proposed mid-IR method, which is faster and easier to use than the current titrimetric method.

1. Introduction

Edible oils are food substances obtained from plants and animal sources. They are usually liquids at room temperature and consist mainly of triglycerides, which are esters formed from the condensation reaction between glycerol and saturated, monounsaturated, and polyunsaturated fatty acids. Tropical edible oils such as palm oil and coconut oil are solids at room temperature because they contain high amounts of short-chain triglycerides and saturated fatty acids [1,2]. The amount and type of fatty acids in edible oils are influenced by the specific variety of the edible oil [3,4,5]. For example, safflower oil has more polyunsaturated fatty acids than palm oil, coconut oil, soybean oil, peanut oil, canola oil, or flaxseed oil.

Edible oils are also the major components in the feedstock used to produce renewable fuels in a variety of industries, including transportation and agriculture. Renewable fuels [6,7] produced from edible oils have emerged as an alternative to oil, natural gas, coal, and other fossil fuels. Examples of edible oils that are used as feedstock for producing renewable fuels include soybean oil, peanut oil, canola oil, rapeseed oil, palm oil, cotton seed oil, coconut oil, safflower oil, and flaxseed oil [8,9,10,11,12]. Currently, 95% of renewable fuels are produced from edible oils that are available on a large scale from agriculture [13]. Approximately 7% of the plant-based edible oils harvested annually are used to produce biodiesel [14]. Most research on renewable fuels is focused on producing biodiesel from plant-based edible oils [15,16,17].

Burgeoning interest in renewable fuels can be attributed to the rapid depletion of fossil fuels caused by the increasing global energy demand and the environmental advantages of renewable fuels, specifically reduced emissions of greenhouse gases. Other advantages of renewables, for example, biodiesel, are their higher combustion efficiency and lower sulfur and aromatic content [18,19]. Biodiesel is also safer as its flashpoint is 423° K compared to 350° K for diesel [20]. Biodiesel also has a higher cetane number [21] than diesel. Cetane number is a measure of the readiness of fuel to auto-ignite when injected into an engine.

Edible oils used as feedstock have been analyzed for a variety of chemical and physical properties, including volatile matter, moisture, fixed carbon, ash, and inorganic and organic elemental content. The four properties of feedstock that are crucial for the conversion of edible oils to renewable fuels are viscosity, density, iodine value, and total acid number (TAN). The feedstock used to produce renewable fuels should be neither too dense nor too viscous, as thick material requires a larger amount of work to move it through the pipeline. Furthermore, the viscosity and density of edible oils can influence the performance of renewable fuel in fuel injection systems, with higher viscosity and density of the fuel leading to poorer atomization and inefficient combustion, resulting in a build-up of deposits on the combustion chamber wall. The iodine value defines the amount of olefin in the feedstock, and, therefore, the amount of hydrogen consumed during hydro-treating. Finally, the TAN for the feedstock is also an important parameter, as even a small increase in TAN for the feedstock above a critical threshold value can lead to corrosion of the catalyst in the refining process.

In this study, the viscosity, density, and iodine value of the feedstocks did not create problems in the refining of renewable fuels. To obviate the effects of higher TAN values, it was necessary to blend several feedstocks prior to refining. As part of a broader effort to standardize the feedstock used in the refining of renewable fuels, the present work focuses on the development of a secondary reference method to determine TAN based on coupling mid-infrared (IR) spectroscopy with partial least squares (PLS) regression to produce a method that is faster, less expensive, easier to use, and can be performed on site compared to the current method for TAN based on an acid-base (potentiometric) titration [22,23]; this represents a paradigm shift in problem solving. Previously, PLS regression has been used in conjunction with the near-infrared region for quantitative analysis. The prediction of protein content in wheat replacing the time-consuming and hazardous Kjeldahl method [24] and the determination of the octane number of gasoline [25] are two examples of a similar paradigm shift in solving important analysis problems.

2. Materials and Methods

Fourier-transform infrared (FTIR) absorbance spectra (4000 cm⁻¹ to 400 cm⁻¹) of 45 samples of feedstock used to produce renewable fuels were collected at 4 cm⁻¹ resolution at 64 scans each with Happ Genzel apodization using an iS50 Thermo-Nicolet FTIR spectrometer equipped with a diamond attenuated total reflection (ATR) accessory and a DTGS detector. Feedstock used to produce renewables were purchased on the commodities market, and limited information was provided about the edible oil type, composition or the processing history. Each feedstock sample was placed on the diamond ATR crystal via a disposable pipette, and the mid-IR spectrum was measured. A representative IR spectrum of a feedstock sample is shown in Figure 1. The most intense absorption bands in the FTIR spectrum are observed at 2925 cm⁻¹ (asymmetric -C-H stretching of -CH2-), 2854 cm⁻¹ (symmetric -C-H stretching of -CH2-), 1746 cm⁻¹ (-C = O stretching of ester) and 1163 cm⁻¹ (-C-O stretching and -CH2- bending). The spectral region between 2200 and 2000 cm⁻¹, which corresponds to the absorbance by the diamond ATR crystal, was excluded from the analysis. Our previous studies on edible oils have shown that absorbance in this region is due solely to absorption by the diamond ATR crystal used to collect the mid-IR spectra [26].

Digitally blended spectral data were generated as part of this study to augment the training set of 45 feedstock samples. Digital blending was performed by combining unprocessed FTIR spectra of real samples to obtain spectra that are representative of samples with a proscribed TAN value (see Figure 2). To obtain a digital blended spectrum representing the IR spectrum of a sample with a TAN value of 8.75, the IR spectrum of a sample with a TAN value of 8.2 is averaged with the IR spectrum of a sample with a TAN value of 9.3. Gaussian distributed noise is added to the IR spectrum of each digital blend to homogenize the data. For each spectrum, noise is only added to the regions that contain IR bands. For a training set of digitally blended IR spectra, the largest absorbance value is identified at each wavelength and one thousandth of this value is multiplied by Gaussian distributed random noise that has a mean of zero and standard deviation of one. If the largest absorbance value is less than or equal to zero, noise is not added to the digitally blended spectrum at that wavelength. Figure 3 compares a digitally blended IR spectrum (TAN is 8.75) to a measured IR spectrum (TAN is 8.71) for the region 1800–1600 cm⁻¹. PLS calibrations for the TAN were developed using only this spectral region as it contains the carbonyl stretch of the carboxylic acid group of fatty acids, which is the source of acidity in the feedstock. PLS was selected because it is considered the gold standard for linear multivariate calibrations.

PLS calibrations [27] for TAN using experimental and/or digitally blended data from the FTIR spectra of feedstocks were developed using UNSCRAMBLER 11 (Camo Analytics). For each calibration, the spectra were preprocessed using orthogonal signal correction (OSC) [28] followed by mean centering to improve both the quality and performance of the model. The number of latent variables for each PLS model was determined using cross validation [29]. Several figures of merit [30] were computed for each partial least squares (PLS) regression model including root mean square error of calibration (RMSEC), standard error of calibration (SEC), bias, root mean square error of cross validation (RMSECV), and standard error of cross validation (SECV).

Mid-IR spectra are often preprocessed to remove systematic noise such as baseline variation and multiplicative scatter effects using first and second derivatives or multiplicative scatter correction [31]. However, these methods may also remove information from the spectra about the response variable. Better results for the PLS calibration of the mid-IR spectra of the feedstock were obtained when OSC, which removes features from the data unrelated to TAN, was employed. When these features are removed before the spectrum is analyzed by PLS, the performance of the calibration model is less impacted by changes in the chemical composition of the background sample matrix, which was another reason for preprocessing the spectral data with OSC prior to PLS.

3. Results and Discussion

Figure 4 summarizes the results of a PLS calibration model developed from the mid-IR spectra of the 45 feedstock samples using a single latent variable. Figures of merit for this PLS calibration (RMSEC, SEC, RMSECV, SECV, R², and bias) are summarized in

Table 1. Figures of merit for PLS calibration of the forty-five feedstock samples.

Figure of Merit	Calibration	Cross Validation
Slope	0.957	0.946
Offset	0.1859	0.2114
Correlation	0.978	0.976
R2	0.957	0.956
Root Mean Square Error	1.13	1.18
Standard Error	1.146	1.195
Bias	−4.2969 × 10−7	−0.0200969

. Both the fitted and cross-validated estimates of TAN exhibited low bias. For cross validation (i.e., jackknifing), the data set was divided into 45 training set prediction set pairs. Each training set consisted of 44 samples, and the prediction set consisted of only 1 sample. Each sample was in the prediction set only once. PLS calibration models developed from the 44 samples in the training set were used to predict the TAN for the sample in the corresponding prediction set. The cross validation set results are summarized in Table 1 for the entire sample cohort and

Table 2. Cross-validation results for the forty-five feedstock samples.

SID 1	Actual	Predicted	Deviation	SID 1	Actual	Predicted	Deviation
16	3.7	6.79	3.09	64	10.85	9.43	1.42
17	9.3	9.29	0.01	66	0.02	0.00	0.02
18	8.71	10.03	1.32	67	0.02	0.00	0.02
44	0.03	0.00	0.03	68	1.11	1.69	0.58
45	11.41	9.97	1.44	69	1.11	1.35	0.24
46	3.76	3.48	0.28	70	1.81	0.75	1.06
47	0.08	0.00	0.00	71	1.89	1.65	0.24
48	0.36	0.76	0.40	72	0.32	0.19	0.13
49	0.04	0.00	0.04	73	0.32	0.81	0.49
50	0.06	0.004	0.056	75	8.45	7.48	0.97
51	0.05	0.00	0.05	76	8.45	8.13	0.32
52	0.1	1.76	1.66	77	11.09	9.06	2.03
53	12.39	9.26	3.13	79	4.45	6.48	2.03
54	13.64	13.38	0.26	80	18.17	17.50	0.67
55	13.64	13.68	0.04	81	18.17	18.69	0.52
56	0.03	0.00	0.03	82	1.67	0.03	1.64
57	13.29	12.22	1.07	83	1.67	2.15	0.48
58	0.03	0.00	0.03	84	0.05	0.00	0.05
59	0.14	0.55	0.41	85	0.05	0.24	0.19
60	8.2	9.02	0.82	86	0.03	0.23	0.20
61	1.15	3.61	2.46	87	0.03	0.38	0.35
62	0.72	1.56	0.84	88	0.02	0.27	0.25
63	3.98	6.50	2.52

¹ SID is the sample identification number.

for each sample. The correlation with TAN is good, and the differences between fitted and cross-validated predictions for R², root mean square error, and standard error do not indicate overfitting by PLS.

To strengthen the calibration, 103 digitally blended IR spectra were generated using the IR spectra of the 45 feedstock samples. In some cases, the feedstock samples used to generate blended IR spectra were selected to fill in regions of the calibration (see Figure 4) where there were only a few samples (e.g., TAN values between two and four, five and eight, and fourteen and eighteen.) In other cases, the spectra of the feedstock samples used to generate digitally blended spectra were selected to reproduce samples that lie in regions of the calibration that are well represented (see Figure 4, e.g., a TAN between zero and two). By using this set of digitally blended spectra for calibration, a PLS regression model for TAN can be developed that spans a wide range of TAN values and is well represented in all regions of the calibration.

As a first step towards developing a digital training set, the PLS calibration developed from the 45 feedstock samples (see Figure 4 and Table 1 and Table 2) was used to predict the TAN values of the digitally blended data.

Table 3. Prediction results for digitally blended experimental data.

SID 1	Actual	Predicted	Deviation	SID 1	Actual	Predicted	Deviation
1	0.06	0.00	0.6	53	2.04	4.0636	2.0236
2	0.1	0.24	0.14	54	2.06	3.4673	1.4073
3	12.35	11.23	1.12	55	2.17	3.6244	1.4544
4	4.46	5.266	0.806	56	2.35	3.9773	1.6273
5	1.105	0.9378	0.1672	57	2.545	4.0707	1.5257
6	8.75	9.1665	0.4165	58	2.825	3.2572	0.4322
7	8.075	6.381	1.694	59	2.895	3.6143	0.7193
8	0.06	0.9642	0.9042	60	2.935	4.0599	1.1249
9	0.04	0.00	0.6658	61	4.075	6.6229	2.5479
10	11.31	12.235	0.04	62	4.105	4.9669	0.8619
11	0.05	0.00	0.05	63	4.215	6.4834	2.2684
12	11.13	9.9241	1.2059	64	2.0825	2.1122	2.9707 × 10−2
13	4.66	4.7785	0.1185	65	2.1175	2.2907	0.1732
14	4.7	5.4875	0.7875	66	2.1375	2.5135	0.376
15	4.72	4.8912	0.1712	67	12.365	11.3466	1.0184
16	4.83	5.0483	0.2183	68	7.93	8.3313	0.4013
17	5.01	5.4012	0.3912	69	11.045	10.5806	0.4644
18	5.435	4.9767	0.4583	70	15.73	15.1351	0.5949
19	5.475	5.6858	0.2108	71	15.905	15.6496	0.2554
20	5.495	5.0895	0.4055	72	15.905	15.807	9.8046 × 10−2
21	5.605	5.2466	0.3584	73	15.73	15.5135	0.2165
22	5.785	5.5995	0.1855	74	15.905	16.028	0.123
23	5.715	5.2025	0.5125	75	15.905	16.1854	0.2804
24	5.755	5.9116	0.1566	76	10.295	9.1862	1.1088
25	5.775	5.3153	0.4597	77	10.55	9.6694	0.8806
26	5.885	5.4724	0.4126	78	15.28	13.6492	1.6308
27	6.065	5.8253	0.2397	79	15.28	14.0277	1.2523
28	6.205	4.7981	1.4069	80	9.92	9.0089	0.9111
29	6.245	5.5072	0.7378	81	9.99	9.366	0.624
30	6.265	4.9109	1.3541	82	10.03	9.8116	0.2184
31	6.375	5.068	1.307	83	9.92	9.3874	0.5326
32	6.555	5.4209	1.1341	84	9.99	9.7444	0.2456
33	6.655	6.284	0.371	85	10.03	10.19	0.16
34	6.695	6.993	0.298	86	14.917	14.1943	0.7224
35	6.715	6.3967	0.3183	87	15.15	14.8803	0.2697
36	6.825	6.5538	0.2712	88	15.15	15.0902	5.9833 × 10−2
37	7.005	6.9068	9.8247 × 10−2	89	14.917	14.4466	0.4701
38	6.83	6.7985	3.1494 × 10−2	90	15.15	15.1326	1.7357 × 10−2
39	6.87	7.5075	0.6375	91	15.15	15.3425	0.1925
40	6.89	6.9112	2.1247 × 10−2	92	16.543	16.0758	0.4675
41	7	7.0684	6.8352 × 10−2	93	16.66	16.4188	0.2412
42	7.18	7.4213	0.2413	94	16.66	16.5238	0.1363
43	2.235	3.384	1.149	95	16.543	16.5804	3.7096 × 10−2
44	2.275	4.093	1.818	96	16.66	16.9235	0.2635
45	2.295	3.4967	1.2017	97	16.66	17.0284	0.3684
46	2.405	3.6538	1.2488	98	12.07	11.0056	1.0645
47	2.585	4.0067	1.4217	99	12.19	10.8321	1.3579
48	2.78	4.1001	1.3201	100	12.35	11.2314	1.1186
49	3.06	3.2866	0.2266	101	12.245	11.6775	0.5676
50	3.13	3.6437	0.5137	102	12.365	11.504	0.861
51	3.17	4.0893	0.9193	103	12.525	11.9033	0.6217
52	2	3.3545	1.3545

¹ Sample identification number.

summarizes the results of the PLS calibration for predicting the TAN values of the digitally blended spectral data. Of the 103 digitally blended spectra, the difference between the corresponding TAN value as predicted by the PLS calibration and the value expected for the digitally blended data (i.e., the deviation) exceeded a user-determined critical threshold value, which is ±1 or greater, for 29 digitally blended IR spectra. (Differences of ±1 unit or greater are significant.) An examination of the samples used to generate these 29 digitally blended spectra (see

Table 4. Identity of samples used to construct the 29 digitally blended spectra.

SID 1	Real Samples	SID	Real Samples	SID 1	Real Samples
3	45, 57	46	48, 79	61	16, 79
7	46, 53	47	62, 79	63	63, 79
12	45, 64	48	68, 79	67	54, 77
28	53, 88	52	63, 88	76	53, 60
30	53, 59	53	52, 63	78	53, 80
31	48, 53	54	59, 63	79	53, 81
32	53, 62	55	48, 63	98	57, 64
43	79, 88	56	62, 63	99	57, 77
44	52, 79	57	63, 68	100	45, 57
45	59, 79	60	63, 71

¹ Sample identification number.

) revealed that spectra generated from samples whose sample identification (SID) numbers were 45, 53, 57, 63, and 79 (see Table 2) are problematic in terms of their TAN predictions. The provenance of these five samples is unique compared to the other forty samples in the cohort, as these five samples consisted of edible oils blended with used cooking oil that was contaminated with spices and/or alcohol, depending upon the part of the world from where they were purchased. Furthermore, digital adducts of these five samples do not appear to follow a linear additive model based on their poor PLS fits. As our approach for digital blending assumes that all data follow a linear additive model [32], the 29 digital adducts of these five samples were deemed unsuitable for inclusion in the calibration set.

Figure 5 summarizes the results of a one-component PLS calibration model developed from the remaining seventy-four digitally blended spectra. The other twenty-nine digitally blended spectra discussed in the previous paragraph were not included in this model because of their poor TAN fit. Figures of merit for this PLS calibration are summarized in

Table 5. Figures of merit for PLS calibration of the digitally blended experimental data.

Figure of Merit	Calibration	Cross Validation
Slope	0.992	0.991
Offset	0.0653	0.0783
Correlation	0.996	0.996
R2	0.992	0.992
Root Mean Square Error	0.454	0.468
Standard Error	0.457	0.471
Bias	−4.4215 × 10−7	0.0036208

. Using the digitally blended data, the slope and R² of the calibration line (for both fitted and cross-validated) are effectively one, the root mean square error (for both fitted and cross-validated) has been reduced by 50% (see Table 1 versus Table 5), and there is also a reduction of one order of magnitude for the bias associated with cross validation (see Table 1 versus Table 5).

The PLS calibration developed from the 74 digitally blended spectra was used to predict the TAN values of the 45 feedstock samples. Figure 6 shows a plot of the predicted versus actual values.

Table 6. Prediction results for experimental spectra.

SID 1	Actual	Predicted	Deviation	SID 1	Actual	Predicted	Deviation
16	3.7	6.75	3.05	64	10.9	9.5	1.4
17	9.3	9.29	0.01	66	0.02	0.00	0.02
18	8.71	9.77	1.06	67	0.02	0.00	0.02
44	0.03	0.00	0.03	68	1.11	1.89	0.78
45	11.4	10.1	1.30	69	1.11	1.45	0.34
46	3.76	3.28	0.48	70	1.81	0.79	1.02
47	0.08	0.00	0.08	71	1.89	1.82	0.07
48	0.36	0.75	0.39	72	0.32	0.24	0.08
49	0.04	0.00	0.04	73	0.32	0.84	0.52
50	0.06	0.00	0.06	75	8.45	7.63	0.8159
51	0.05	0.03	0.02	76	8.45	8.32	0.1251
52	0.1	1.71	1.61	77	11.1	9.64	1.46
53	12.4	9.29	3.11	79	4.45	6.46	2.01
54	13.6	13.3	0.3	80	18.2	18.4	0.2
55	13.6	13.7	0.1	81	18.2	19	0.8
56	0.03	0.00	0.03	82	1.67	0.00	1.67
57	13.3	12.4	0.9	83	1.67	2.29	0.62
58	0.03	0.00	0.03	84	0.05	0.00	0.05
59	0.14	0.54	0.40	85	0.05	0.38	0.33
60	8.2	8.84	0.64	86	0.03	0.36	0.33
61	1.15	3.29	2.14	87	0.03	0.41	0.38
62	0.72	1.29	0.57	88	0.02	0.37	0.35
63	3.98	6.27	2.29

¹ Sample identification number.

summarizes the results of the PLS calibration for predicting the 45 TAN values of the original spectral data. The R2 for predicted TAN values of the 45 feedstock samples using the PLS model developed from digital data was 0.9625 (see Figure 6), which is larger than the R2 (0.957) for the fitted values computed from the PLS model for TAN that was developed using these same 45 feedstock samples as a training set (see Table 1). Clearly, the PLS calibration developed from digitally blended spectral data can provide reasonable predictions of TAN for actual feedstock samples.

Figure 7 shows the results of the PLS calibration for the data cohort of 74 digitally blended spectra and 44 experimental spectra. Sample 53 (an experimental mid-IR spectrum) was deleted from the original data cohort of 74 digitally blended spectra and 45 experimental spectra because it was flagged as an outlier by PLS. The 118 spectra cover almost the entire range of the calibration.

Table 7. Figures of merit for PLS calibration of experimental and digitally blended data.

Figure of Merit	Calibration	Cross Validation
Slope	0.983	0.982
Offset	0.115	0.126
Correlation	0.991	0.991
R2	0.983	0.982
Root Mean Square Error	0.7324	0.7427
Standard Error	0.7355	0.7458
Bias	−1.0574 × 10−7	0.0021452

summarizes the figures of merit for this calibration. The root mean square error of calibration for the TAN is 0.7324 compared to a root mean square error of calibration of 1.13 for the PLS model developed from the experimental data (see Table 1). Clearly, there is benefit in combining digital data with experimental data for determining the TAN from mid-IR spectra which demonstrates the advantages of using digital data to enhance PLS multivariate calibrations. Although there was no independent test set to validate this model, an independent test set was not necessary as the feedstock used to produce the renewables for the pilot plant also served as standards for the PLS calibration of the TAN. In practice, different feedstock samples were combined to produce renewable fuels, and the sample calibration set for the PLS model includes samples from the same lots of the raw materials used to produce these fuels. By combining digitally blended spectral data with experimental spectral data, the PLS calibration model obtained was superior to the calibration model obtained using only experimental data (compare Table 7 to Table 1).

4. Conclusions

FTIR spectra of renewable feedstocks with proscribed acid property values were successfully generated using digitally blended experimental data obtained from the mid-IR spectra of renewables with known TAN values. This approach for generating additional data for PLS calibrations has several advantages that include providing digitally blended data for regions of the spectra–property calibration where the number of samples is sparse or nonexistent and greater understanding of the spectra–property correlation through identification of sample outliers in the original data that are often difficult to detect using traditional PLS diagnostics. The results of this study also demonstrate the advantages of variable selection to focus on the informative regions in the spectrum and the use of OSC for extracting signatures indicative of acidity from the mid-IR spectra of the feedstock of the renewables. In future studies, baseline correction will be investigated to ensure high-quality PLS calibration models for TAN, iodine value, viscosity and density.