Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2024 Apr 5.
Published in final edited form as: J Hazard Mater. 2023 Jan 16;447:130806. doi: 10.1016/j.jhazmat.2023.130806

Capillary Flow Velocity Profile Analysis on Paper-Based Microfluidic Chips for Screening Oil Types Using Machine Learning

Soo Chung a,b, Andrew Loh c, Christian M Jennings d,†, Katelyn Sosnowski d, Sung Yong Ha c, Un Hyuk Yim c,*, Jeong-Yeol Yoon a,d,*
PMCID: PMC9940998  NIHMSID: NIHMS1867198  PMID: 36680906

Abstract

We conceived a novel approach to screen oil types on a wax-printed paper-based microfluidic platform. Various oil samples spontaneously flowed through a micrometer-scale channel via capillary action while their components were filtered and partitioned. The resulting capillary flow velocity profile fluctuated during the flow, which was used to screen oil types. Raspberry Pi camera captured the video clips, and a custom Python code analyzed them to obtain the capillary flow velocity profiles. 106 velocity profiles (each with 125 frames for 5 s) were recorded from various oil samples to build a training database. Principal component analysis (PCA), support vector machine (SVM), and linear discriminant analysis (LDA) were used to classify the oil types into heavy-to-medium crude, light crude, marine fuel, lubricant, and diesel oils. The second-order polynomial SVM model with PCA as a pre-processing step showed the highest accuracy: 90% in classifying crude oils and 81% in classifying non-crude oils. The assay took less than 30 s from the sample to answer, with 5 s of the capillary action-driven flow. This simple and effective assay will allow rapid preliminary screening of oil types, enable early tracking, and reduce the number of suspect samples to be analyzed by laboratory fingerprinting analysis.

Keywords: oil spill, Raspberry Pi, paper microfluidic chip, capillary action, SVM

Graphical Abstract

graphic file with name nihms-1867198-f0001.jpg

1. Introduction

Marine oil spills have frequently occurred worldwide from ship’s operational discharge, accidents, intentional dumping, etc. (Gaines et al., 2006; Boehmer-Christiansen, 2008; Yim et al., 2012). While large-scale oil spill accidents are relatively rare, small-scale oil spills continue to occur almost daily (Yim et al., 2012). Smaller oil spills caused by leakage or discharge from ships are often inconspicuous, and it is difficult to obtain information if the responsible ship remains unknown (Shankar et al., 2020; Suneel et al., 2019). On average, about 260 oil spills have been reported annually in South Korea for the last 10 years (Korea Coast Guard, 2021). A preliminary screening tool will benefit the coast guard field personnel for early tracking of these ships to prevent further spillage and enforce cleanup responsibility. Rapid pre-screening of oil types will significantly help the enforcement personnel to narrow down the responsible ships. The current gold standard is collecting spilled samples and suspected sources, followed by analyzing them in a remote laboratory in a fingerprinting manner, which takes substantial time and resources. Such oil fingerprinting analyses can be classified into spectrometry and chromatography. A pre-screening step of classifying oil types will save the reagents and reduce the turnaround time in the following precision laboratory analysis in a tiered approach. Only the suspect samples can be analyzed in the laboratory in a fingerprinting manner. In addition, the classification of crude vs. non-crude oils and sub-classification of non-crude oils can help identify the ships responsible for smaller, routine oil spills, as described above. For example, only diesel oils should be collected from the ships and compared with the suspect sample if it is classified as diesel oil.

Oils contain aromatic hydrocarbon that can be identified by UV spectrometry or Fourier transform infrared (FTIR) spectrometry (Permanyer et al., 2005; Mirnaghi et al., 2018). Our laboratory has recently developed a portable in-situ fluorescence spectrometer device that could classify into three to four oil types, e.g., crude, heavy fuel, light fuel, and lubricant oils (Bills et al., 2020; Sosnowski et al., 2022). Spectrometric analyses, however, have a limitation in that they rely only on aromatic hydrocarbons. More accurate identification can be made by gas chromatography-mass spectrometry (GC-MS) or gas chromatography-flame ionization detector (GC-FID), which has widely been used for oil fingerprinting analysis (Riley et al., 2018; Juahir et al., 2017). The characteristic make-up of different oil types would yield a unique chromatogram, which can be analyzed via a fingerprinting approach, including machine learning classifications (Riley et al., 2018; Juahir et al., 2017). However, chromatography methods require benchtop equipment and, subsequently, a laboratory environment, which is not appropriate to be implemented in situ in the field.

In this work, we aim to conceive a novel approach to pre-screen oil types that are based on neither spectrometry nor chromatography. Instead, we chose a microfluidic platform, where a liquid sample flows through a micrometer-scale channel to induce chemical or biological reactions, incubation, separation, and detection. Specifically, paper substrates (cellulose or nitrocellulose fibers) have become increasingly popular, referred to as paper-based microfluidic chips or microfluidic paper-based analytic devices (μPAD) (Klug et al., 2018; Kaarj et al., 2018; Yetisen et al., 2013). Liquid flow is spontaneous via capillary action (also known as wicking), unlike the conventional silicone-based microfluidic chips. Therefore, there is no need for a high-voltage electroosmotic flow or an external/internal pump, which can be costly or challenging to operate (Pantoja et al., 2004). In addition, capillary action occurs based on intermolecular forces between the solution and surrounding porous material surfaces and is not affected by outside gravity, inertia, or any external forces (Klug et al., 2018; Nia and Jessen, 2015; Kim, 2000). As such, this method may be well-suited for the on-ship and/or field-based analyses. Paper-based microfluidic chips have primarily been used for sensing various chemicals, proteins, bacteria, viruses, etc., using specific receptor molecules, including enzymes, antibodies, aptamers, nucleic acids, etc. However, since paper substrates can also filter and partition the mixtures in various types of samples (Songjaroen et al., 2012; Santhiago et al., 2014), paper-based microfluidic chips can be utilized as a platform for paper chromatography. Furthermore, since oil samples are also mixtures of multiple hydrocarbons, we can expect the separation and partitioning of each component as the oil sample flows through the paper fibers via capillary action.

In this work, we measure the capillary flow velocity through the paper microfluidic channel length to create a capillary flow velocity profile of various oil samples. As the oils flow through the paper fibers, one oil component can be retained earlier in the flow, another element is retained later in the flow, and others are not retained. The capillary flow velocity is described in a modified Lucas-Washburn (L-W) equation (Klug et al., 2018):

l2t=RγLGcosθ2μ

Where l is flow distance, t is time, R is the pore size (corresponding to the capillary radius), γLG is the surface tension at the liquid-gas interface (also known as the wetting front), θ is the liquid contact angle, and μ is the liquid viscosity. (Note that this equation does not contain any parameters related to gravitational forces.) Assuming the pore size (R) remains the same, the capillary flow rate (l2/t) is a function of interfacial properties (γLG and θ) and viscosity (μ) of a liquid. The interfacial properties are related to the capillary pressure, and the viscosity is related to the inertial/viscous force. If both interfacial properties and viscosity remain the same throughout the flow, we can lump the right side of Eq. 1 into a single constant K. The flow distance (l) can be plotted against time (t) as:

l=Kt

The flow velocity (dl/dt) becomes:

dldt=K2t

The resulting flow velocity profile should show a monotonous decrease.

However, different oil components interact with the cellulose fibers at varying extents and are subsequently retained at different channel lengths. Such retention should alter interfacial properties (γLG and θ) and/or viscosity (μ) of a liquid, leading to fluctuations in K. As shown later in this work, the flow velocity profiles did not decrease monotonously – they fluctuated up and down as the oil flowed through the paper channels. While it is impossible to pinpoint each fluctuation, we can treat the entire velocity profile as a multi-dimensional data set. A supervised machine learning classification can be used to identify unknown oil samples by collecting a good number of training data sets with known identities of oil types.

Interactions between each oil component and the paper fibers and subsequent retention should alter the interfacial properties at the wetting front (i.e., liquid-gas = oil-empty pore interface) and the viscosity of the flowing oil sample. Such interactions and retention should arise from the molecular properties of each oil component and the paper fibers, including charge, polarity, hydrophobicity, etc. These interactions and retentions should be altered throughout the capillary action from the inlet to the end of each paper microfluidic channel.

Before we collect the training and test data sets, we need to ensure that oil samples can flow through the paper pores spontaneously via capillary action. Both cellulose and nitrocellulose fibers are strongly hydrophilic and allow water to flow spontaneously, referred to as “wicking.” However, since most oil samples are hydrocarbons and subsequently hydrophobic, they would not flow through paper pores. Many paper-based microfluidic chips have recently been fabricated with a wax printer (Yetisen et al., 2013). Wax is printed on a paper substrate and later melted on a hot plate to make the wax fill the depth of a paper substrate. This process creates a hydrophobic barrier on a paper-based microfluidic chip. Water samples flow in between these barriers where the wax is not printed. We can use this hydrophobic barrier to allow oil samples to flow, while the area with no wax printing does not allow oils to flow. We also need to ensure sufficient porosity of the wax-printed and wax-melted paper to allow capillary action to the oil samples and set up the wax amount. In addition, the wax was injected from a printer at a constant volume, minimizing the variations in the chips’ porosity.

In this work, we used a Raspberry Pi camera to capture a video clip of the capillary action of various oil samples through the paper-based microfluidic chip. The capillary flow velocity (dl/dt) profile was obtained throughout the flow using a custom Python code. Experiments were repeated with varying oil samples to build a training database. Firstly, principal component analysis (PCA) was used to reduce the high-dimensional data (i.e., the capillary flow velocity profile over time) into a smaller dimension. Then, identical oil types were clustered together to classify different oil types of test samples (Stravers et al., 2019; Imamura et al., 2017). Each principal component was correlated to the material properties of various oil samples. Secondly, support vector machine (SVM) and linear discriminant analysis (LDA) were used as supervised machine learning to classify the oil types. These two machine learning techniques can provide affordable predictive performance with a limited sample size and should fit well with the data we acquired (Pisner and Schnyer, 2020; Cheng et al., 2009; Golub et al., 1999; Armstrong et al., 2002). Initially, the data were used without reducing their dimensionality, using all 125 frames or the first 10 or 25 frames. Later, PCA was used as a pre-processing step to reduce the dimensionality while using 10, 5, or 3 principal components (PCs).

This work investigates the capillary flow velocity profile analysis toward successful oil type classification. It also seeks to evaluate the Raspberry Pi-based paper microfluidic chromatography as a novel, handheld, auxiliary method of oil type classification.

2. Materials and Methods

2.1. Paper-Based Microfluidic Chips

Paper-based microfluidic chips were fabricated using a wax printer (ColorQube8580; Xerox Corporation, Norwalk, CT, USA). Chips were designed using SolidWorks (Dassault Systèmes; Vélizy-Villacoublay, France), and each channel was 1.25 mm wide and 25 mm long. The chip design was printed with solid ink on cellulose chromatography paper (Grade 2 CHR; thickness = 0.18 mm; flow rate = 115 mm / 30 min; catalog no. 3002-917; GE Healthcare, Maidstone, Kent, UK). Each chip was cut out and placed on a hot plate at 100°C for 1 min to melt the wax through the paper’s depth while evenly coating fibers. Unlike other paper-based microfluidic chips, the inside of the channel was filled with wax and thus hydrophobic, allowing oil to flow through the chip.

2.2. Oil Samples

All oil samples were collected by the Korea Institute of Ocean Science and Technology (KIOST). A total of 30 different oil sample types were used in this work. Thirteen were crude oil samples from Middle East Asia (Persian Gulf, Saudi Arabia, and Iran), Southeast Asia (Malaysia and Philippines), and Australia (e.g., Pyrenees). Five were marine fuel oils (MFOs), six were diesel oils, and six were lubricant oils. All oil samples are diluted at a ratio of 10:1 with octane (Sigma-Aldrich) and stored at 4°C. A total of 106 oil samples (3–5 samples per each oil type) were analyzed on paper-based microfluidic chips, each time using a different chip, generating 106 flow velocity profiles. Density was measured with a density meter and viscosity with a viscometer. Saturate, aromatic, resin, and asphaltene contents were measured by thin layer chromatography with flame ionization (TLC/FID; Iatroscan Mk IV; Iatroscan Labs Inc., Tokyo, Japan), following the method described by (Bills et al., 2020).

2.3. The Flow Velocity Profile Monitoring Device

The flow velocity was monitored from the paper-based microfluidic chip during the capillary action using a custom-made device. The device consisted of a Raspberry Pi microcontroller (Raspberry Pi Foundation, Cambridge, UK) and a Pi camera (Raspberry Pi Foundation). The device housing was designed using SolidWorks and 3D-printed using the Ultimaker 3 (Ultimaker B.V.; Cambridge, MA, USA), as shown in Figure 1. The device dimensions are 100 mm wide, 120 mm long, and 80 mm high.

Figure 1.

Figure 1.

Open in a new tab

A flow velocity profile monitoring device. The paper-based microfluidic chip is dipped into the oil sample to induce capillary action through the paper chip. The Pi camera captures a video clip of this capillary action. A Raspberry Pi microcontroller processes the video clip to create the capillary flow velocity profile, i.e., the plot of the flow velocity over time.

2.4. The Flow Velocity Profile Analysis

The paper-based microfluidic chip was coated with hydrophobic wax to allow the oil to flow through the chip. First, the diluted oil sample was loaded on the slide glass tray using a 0.5-mL microtube. Next, the inlet of a paper-based microfluidic chip (attached to a microscope glass slide) was dipped into the slide glass tray by manually sliding down the microscope glass side, as shown in Figure 1. This action initiated the capillary action, during which the oil sample flowed rapidly through the channel within a few seconds. Finally, a Pi camera took a real-time video clip and sent it to the Raspberry Pi microcontroller. The color intensity changed as the oil sample flowed through the paper chip, which was used to identify the flow distance. The video clip was split into frames (25 frames/s), and the flow distance (l) was measured using a custom Python code. The flow distance changes per time frame were then evaluated as the flow velocity (dl/dt). The flow velocity was assessed for 5 s, i.e., a total of 125 frames.

2.5. Machine Learning Classifications

Principal component analysis (PCA) was performed using Python code from the sci-kit learn library (https://scikit-learn.org/stable/). PCA is a non-supervised classification tool that classifies multivariate data into a reduced dimension. All 125 frames of flow velocity profiles were used for PCA. The first two principal components (PC1 and PC2) were initially used to create a PCA score plot and classify oil samples into several categories.

Support vector machine (SVM) and linear discriminant analysis (LDA) were performed using MATLAB version R2022a (The MathWorks, Inc.; Natick, MA, USA) using fitececoc (https://www.mathworks.com/help/stats/fitcecoc.html) and fitcdiscr functions https://www.mathworks.com/help/stats/fitcdiscr.html) (Supplementary Code S1). SVM and LDA are supervised classification tools that characterize or separate two or more classes of objects or events using a linear or non-linear combination of features. Initially, all 125 frames (5 s) of data were used without any pre-processing for SVM and LDA. The first 25 or 10 frames of data were also used without pre-processing for SVM and LDA since the most fluctuations could be observed in the first 1 s. Next, PCA was used as a pre-processing step for SVM and LDA. All 125 frames were used for PCA, and the first 10 PCs were used for SVM and LDA. Similarly, the first 5 or the first 3 PCs were also tested for SVM and LDA.

3. Results and Discussion

3.1. Flow Velocity Profiles

Initially, we started the experiments using undiluted oil samples. However, they were too viscous to flow through the paper chips. Therefore, oil samples were diluted at 5:1, 10:1, and 20:1. While dichloromethane (DCM) was initially used to dilute the oil samples (Bills et al., 2020), they melted the pipette tips and were unusable. Therefore, we switched to octane to dilute the oil samples. The most distinctions in flow velocities could be found with 10:1 dilution, while 20:1 flowed too fast and 5:1 too slow. Therefore, we used 10:1 dilution with octane as described in the Materials and Methods.

The flow length (l) from the inlet was evaluated from the image of each time frame. The increase in the flow length was divided by the time increment to obtain the flow velocity. The flow velocity was plotted against time to yield a flow velocity profile. The total number of profiles was 106 from 30 different oil sample types. Each profile was obtained from 125 frames (over 5 s), thus 125 dimensions of data. Paper microfluidic chips were used once and then discarded. Figure 3 shows 106 flow velocity profiles, categorized into heavy-to-medium crude oil, light crude oil, marine fuel oil (MFOs), lubricant oil, and diesel oil. A significant change in flow velocity was observed within the first 1 s. It took less than 5 s for the sample to reach the end of the channel. Since oil samples are mixtures CHNSO compounds, chromatographic separation occurred, as shown in Figure 2. Such separation caused numerous fluctuations in the flow velocity profile, contrary to the monotonous decrease expected from the Lucas-Washburn equation.

Figure 3.

Figure 3.

Open in a new tab

106 flow velocity profiles of heavy-to-medium crude oils (18), light crude oils (31), marine fuel oils (17), lubricant oils (22), and diesel oils (18) on the cellulose paper-based microfluidic chips.

Figure 2.

Figure 2.

Open in a new tab

Still images at 1 s (top) and 3 s (bottom) of the capillary action with various oil samples. Pyrenees (Australia) and Iranian are crude oil samples. MFO is marine fuel oil. The separation of oil samples is indicated with a red arrow.

The heavy vs. light classification was made using the American Petroleum Institute gravity (API gravity). The API gravity is related to the specific gravity (SG) by API = 141.5 / SG – 131.5 (Pabón and de Souza Filho, 2019; Demirbas et al., 2015). In (Pabón and de Souza Filho, 2019), crude oils are categorized into API < 22.3° as heavy, 22.3°–31.1° as medium, and >31.1° as light. In (Demirbas et al., 2015), they are categorized into API < 22° as heavy, 22°–38° as intermediate, and >38° as light. Considering the global API averages of all crude oils being 33.0°–33.5° (Demirbas et al., 2015), we have decided to categorize the crude oils into heavy-to-medium vs. light using the cut-off API of 34.5° = 0.8524 g/cm3, the average of (Pabón and de Souza Filho, 2019) and (Demirbas et al., 2015). Among thirteen crude oil sample types used in this study, five types were heavy-to-medium, and eight were light, which was a good segmentation of all crude oil sample types. This work did not aim to evaluate API to classify crude oil types. It was used to pre-divide all crude oils into two groups, and the capillary flow velocity profiles classified them into these two categories based on the training data set.

3.2. Oil Type Classification by PCA

The flow velocity profiles were used as input variables for PCA, and the principal components (PCs) were obtained. PC1 represented 30.6% of data for all 106 velocity profiles, and PC2 represented 8.4%. The score plots of PC1 and PC2 are shown in Figure 4. While the other PCs were also evaluated, along with PC1 and PC2, no meaningful classifications could be observed. This result can be correlated to the large fluctuations in the flow velocity in the first 1 s out of 5 s. Each symbol represents a single oil sample (and, subsequently, a single flow velocity profile). Each oil category’s averages and standard deviations – heavy-to-medium crude, light crude, MFO, lubricant, and diesel oils – are also shown as colored ellipses within the score plots. Separation could be observed between heavy-to-medium vs. crude oils, although the ellipses overlapped substantially. MFO, lubricant, and diesel oils were separated more distinctively. When all five oil categories were plotted in a single PCA score plot, the crude oils overlapped with the other three oil categories (MFO, lubricant, and diesel oils). This result can be explained by the varying characteristics of different crude oil samples, primarily differentiated by their origins. Crude oil spills are rare but significant in magnitude, while the other three oils’ spills are more frequent but small in volume (ITOPF Limited, 2021; Korea Coast Guard, 2021). Therefore, we have separated the oil type classifications into two scenarios: 1) identification of crude oil spills into heavy-to-medium vs. light categories, and 2) identification of routine fuel/lubricant oil spills into MFO vs. lubricant vs. diesel categories.

Figure 4.

Figure 4.

Open in a new tab

The PCA score plots (PC1 and PC2) of 106 flow velocity profiles from paper-based microfluidic chips (top). The ellipses represent the averages and standard deviations of each oil category. Density and viscosity were plotted against PC1 (bottom).

Both PC1 and PC2 were plotted against the oils’ properties: density, viscosity (mPa-s), saturate content (%), resin content (%), aromatic content (%), and asphaltene content (%). The material properties of all oil sample types are summarized in Supplementary Table S1. As expected, both density and viscosity were negatively correlated to PC1 for crude oils (heavy-to-medium vs. light) and non-crude oils (MFO, lubricant, and diesel), as shown in Figure 4 at the bottom. For crude oils, weak correlations could also be found for density-PC2, aromatic-PC1, and aromatic-PC2, as shown in Supplementary Figure S1. No further correlations could be found for saturate-PC1, resin-PC1, asphaltene-PC1, viscosity-PC2, saturate-PC2, resin-PC2, and asphaltene-PC2. For non-crude oils, weak correlations could be found for resin-PC1, aromatic-PC1, and asphaltene-PC1, as shown in Supplementary Figure S2. No further correlations could be found between saturate and PC1, and all properties and PC2.

To summarize, the material properties were correlated to the flow velocity profile data, which enabled the separation of oil categories in the PCA score plot. Nonetheless, more data (not just PC1 and PC2) should be used to identify the oil categories successfully. Furthermore, supervised learning can also contribute to better classification of oil types, which is explained in the following section.

3.3. Oil Type Classification by SVM and LDA

To better classify the oil categories, we utilized supervised machine learning techniques – SVM and LDA. We randomly selected 75% of the data as a training set and attempted to classify the remaining 25% data. Since there were 49 crude oil profile data, we used 36 of them for training and 13 for the test (5 were heavy-to-medium crude oils and 8 were light crude oils). Likewise, 57 non-crude oil profile data were split into 43 for training and 14 for the test (4 were MFOs, 5 were lubricant oils, and 5 were diesel oils). Three different random splits were used – identified as split 1, split 2, and split 3, respectively. For SVM, linear, second-order polynomial (poly 2), and third-order polynomial (poly 3) models were used. For both SVM and LDA, all 125 frames of data were initially used without pre-processing step (i.e., PCA). For comparison, all 125 frames of data were pre-processed with PCA, and the first ten PCs (PC1 through PC10) were used. The results are summarized in Figures 5A and 5B. Overall, the second-order polynomial SVM model showed the best accuracies in classifying the crude oils (85% with and without PCA) and the non-crude oils (81% with PCA). Therefore, we fixed the model to the second-order polynomial (poly 2) SVM.

Figure 5.

Figure 5.

Open in a new tab

SVM and LDA classifications of crude (A) and non-crude oils (B) with linear, second-order polynomial (poly 2), and third-order polynomial (poly 3) models, with and without using PCA. 106 profiles were randomly split into 75:25 training/test sets. Poly 2 SVM model was further optimized using the first 25 or 10 frames (without PCA) or the first 3 or 5 PCs (C and D). Confusion matrices are shown for the best-performing model.

We also tested high-order polynomial SVM models (e.g., fifth and tenth order) and RBF kernel SVM model. While the training model showed higher classification accuracy than the second-order polynomial model, the accuracies of the test sets (three sets of random splits) were significantly inferior, indicating these models were overfitted. Therefore, we did not evaluate these models further.

We then optimized how much data should be used in classifying the oil types using the SVM poly 2 model. In the cases without PCA, we compared using all 125 frames, the first 25 frames, and the first 10 frames. In the cases with PCA (all 125 frames were fed into PCA), we compared using PC1 through PC10, PC1 through PC5, and PC1 through PC3. The results are summarized in Figures 5C and 5D. For the crude oils, 90% accuracy could be achieved using PC1 through PC3. For the non-crude oils, 81% accuracy could be achieved using PC1 through PC10. For these optimum cases, the confusion matrices of three different random test sets are also shown at the bottom of Figure 5. The confusion matrices of all other SVM and LDA classifications are shown in Supplementary Figures S3 through S6.

Additionally, we isolated 16 profiles as an independent validation test set. The remaining 90 profiles were randomly split into 75% training and 25% test sets, and the same SVM and LDA techniques were used. The results are summarized in Supplementary Figure S7. The second-order polynomial SVM models showed comparable accuracies to the results shown in Figure 5. Accuracies were 90% for classifying crude oils (previously 90%) and 74% for classifying non-crude oils (previously 81%). The latter’s compromise can be attributed to the smaller amount of data.

4. Conclusion

We developed an easy-to-use, low-cost, and rapid assay for identifying the oil type into five categories: heavy-to-medium crude oils, light crude oils, marine fuel oils, lubricant oils, and diesel oils. The oil samples were introduced to the wax-printed paper-based microfluidic chips and spontaneously flowed via capillary action. The flow velocity fluctuated during the capillary action due to the active chromatographic separation of the oil components. A Raspberry Pi camera captured a video clip of this capillary action, and a custom Python code generated the flow velocity profile. The resulting profiles were fed into the PCA, SVM, and LDA models and the oil types were classified into five categories. PCA could classify crude oils into two categories (heavy-to-medium vs. light) and the other oils into three (marine fuel, lubricant, vs. diesel oils), although unable to classify all five categories together. Various models of SVM and LDA were also evaluated in classifying the crude oils and non-crude oils. The flow velocity profiles were randomly split into three training/test set combinations at a 75:25 ratio. The second-order polynomial SVM model with PCA as a pre-processing step showed the best accuracies: 90% in classifying heavy-to-medium vs. light crude oils using three PCs (PC1 through PC3) and 81% in classifying marine fuel, lubricant, vs. diesel oils using ten PCs (PC1 through PC10). The assay took less than 30 s from the sample to answer, with 5 s of the capillary action-driven flow. A limitation of this work is the use of a relatively small number of oil samples and types and the non-use of “weathered” samples. A larger number of training data sets, including weathered samples, could strengthen the training model in the follow-up studies. In addition, a clear limitation exists since this method cannot precisely predict the exact type of oil or evaluate the chemical compositions (such as SARA). Nonetheless, this simple and effective assay will allow a rapid pre-screening of oils before using laboratory equipment in a tiered approach. Another limitation is that the oil type classification may not lead to an optimum remediation action. However, we may be able to use the oil type information to narrow down the responsible ships in the field.

Supplementary Material

1

Environmental Implication.

Recently, the number of oil spill incidents at sea has increased worldwide. The type of oil spill is crucial to trace its origin and plan remedial action. Such identification should be made at the point of sample collection to locate the ship and the contamination source. The current gold standard for monitoring ocean oil spills is laboratory-based gas chromatography. In this work, we conceived a novel field-ready approach (< 30 s) to identifying different oil types on a wax-printed paper microfluidic platform by collecting capillary flow velocity profiles. Oil types were successfully classified using various machine learning-based classification methods.

Highlights.

  • A novel approach to screen oil types on a paper microfluidic platform

  • Raspberry Pi camera acquired capillary flow velocity profiles of diverse oil samples

  • Various machine learning based classifications were tested, including PCA, SVM, and LDA

  • 90% accuracy in classifying crude oil samples and 81% in non-crude oil samples

  • < 30 s from the sample to answer without the need for laboratory equipment

Acknowledgments

This research was a part of the project titled “Development of Advanced Oil Fingerprinting System (PN67490),” funded by the Korea Coast Guard, Republic of Korea. The authors also acknowledge “Monitoring of Source and Behavior of the Particulate Matter at Busan Seaport Area (NRF-2019 2019M1A2A210395512),” funded by the Ministry of Education through the National Research Foundation, Korea. In addition, C.M.J. acknowledges the Environmental Health Sciences Transformative Research Undergraduate Experience (EHS-TRUE) at the University of Arizona, funded by U.S. National Institutes of Health – National Institute of Environmental Health Sciences (NIH-NIEHS) grant number R25ES025494. K.S. acknowledges the Computational and Mathematical Modeling of Biomedical Systems Training Grant from the U.S. National Institutes of Health – National Institute of General Medical Sciences (NIH-NIGMS) grant number GM 132008. Finally, the authors thank Ms. Bai Hei at the University of Arizona for assistance in data analyses.

Footnotes

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

CRediT Authorship Contribution Statement

Soo Chung: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Visualization. Andrew Loh: Resources, Data curation. Christian M. Jennings: Methodology, Investigation. Katelyn Sosnowski: Methodology, Software, Formal analysis, Investigation. Sung Yong Ha: Resources. Un Hyuk Yim: Conceptualization, Methodology, Supervision, Project administration, Funding acquisition. Jeong-Yeol Yoon: Conceptualization, Methodology, Formal analysis, Investigation, Data curation, Writing – review & editing, Visualization, Supervision, Project administration, Funding acquisition.

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

References

  1. Armstrong SA, Staunton JE, Silverman LB, Pieters R, den Boer ML, Minden MD, Sallan SE, Lander ES, Golub TR, Korsmeyer SJ, 2002. MLL translocations specify a distinct gene expression profile that distinguishes a unique leukemia. Nat. Genet. 30, 41–47. 10.1038/ng765. [DOI] [PubMed] [Google Scholar]
  2. Bills MV, Loh A, Sosnowski K, Nguyen BT, Ha SY, Yim UH, Yoon J-Y, 2020. Handheld UV fluorescence spectrophotometer device for the classification and analysis of petroleum oil samples. Biosens. Bioelectron. 159, 112193. 10.1016/j.bios.2020.112193. [DOI] [PubMed] [Google Scholar]
  3. Boehmer-Christiansen S, 2008. Estimates of oil entering the marine environment from sea-based activities, reports and studies No. 75. Energy Environ. 19, 765. 10.1260/095830508784815955. [DOI] [Google Scholar]
  4. Cheng J, Yu D, Tang J, Yang Y, 2009. Application of SVM and SVD technique based on EMD to the fault diagnosis of the rotating machinery. Shock. Vib. 16, 89–98. 10.3233/SAV-2009-0457. [DOI] [Google Scholar]
  5. Demirbas A, Alidrisi H, Balubaid MA, 2015. API gravity, surfur content, and desulfurization of crude oil. Petroleum Sci. Technol. 33, 93–101. 10.1080/10916466.2014.950383. [DOI] [Google Scholar]
  6. Gaines RB, Hall GJ, Frysinger GS, Gronlund WR, Juaire KL, 2006. Chemometric determination of target compounds used to fingerprint unweathered diesel fuels. Environ.Forensics 7, 77–87. 10.1080/15275920500506949. [DOI] [Google Scholar]
  7. Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP, Coller H, Loh ML, Downing JR, Caligiuri MA, Bloomfield CD, 1999. Molecular classification of cancer: Class discovery and class prediction by gene expression monitoring. Science 286, 531–537. 10.1126/science.286.5439.531. [DOI] [PubMed] [Google Scholar]
  8. Imamura R, Murata N, Shimanouchi T, Yamashita K, Fukuzawa M, Noda M, 2017. A label-free fluorescent array sensor utilizing liposome encapsulating calcein for discriminating target proteins by principal component analysis. Sensors 17, 163. 10.3390/s17071630. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. ITOPF Limited, 2021. Oil Tanker Spill Statistics 2021. ITOPF Limited, London. [Google Scholar]
  10. Juahir H, Ismail A, Mohamed SB, Toriman ME, Kassim AM, Zain SM, Ahmad WKW, Wah WK, Zali MA, Retnam A, Taib MZM, Mokhtar M, 2017. Improving oil classification quality from oil spill fingerprint beyond six sigma approach. Mar. Pollut. Bull. 120, 322–332. 10.1016/j.marpolbul.2017.04.032. [DOI] [PubMed] [Google Scholar]
  11. Kaarj K, Akarapipad P, Yoon J-Y, 2018. Simpler, faster, and sensitive Zika Virus assay using smartphone detection of loop-mediated isothermal amplification on paper microfluidic chips. Sci. Rep. 8, 12438. 10.1038/s41598-018-30797-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Kim C-J, 2000. Microfluidics using the surface tension force in microscale. Proc. SPIE 4177, 18–24. 10.1117/12.395667. [DOI] [Google Scholar]
  13. Klug KE, Reynolds KA, Yoon J-Y, 2018. A capillary flow dynamics-based sensing modality for direct environmental pathogen monitoring. Chem. Eur. J. 24, 6025–6029. 10.1002/chem.201800085. [DOI] [PubMed] [Google Scholar]
  14. Korea Coast Guard, 2021. 2021. Korea Coast Guard Annual Report. Korea Coast Guard, Incheon. https://www.kcg.go.kr/ebook/whitebook/2020/index.html [Google Scholar]
  15. Mirnaghi FS, Soucy N, Hollebone BP, Brown CE, 2018. Rapid fingerprinting of spilled petroleum products using fluorescence spectroscopy coupled with parallel factor and principal component analysis. Chemosphere 208, 185–195. 10.1016/j.chemosphere.2018.05.111. [DOI] [PubMed] [Google Scholar]
  16. Nia SF, Jessen K, 2015. Theoretical analysis of capillary rise in porous media. Transp. Porous Media 110, 141–155. 10.1007/s11242-015-0562-1. [DOI] [Google Scholar]
  17. Pabón REC, de Souza Filho CR, 2019. Crude oil spectral signatures and empirical models to derive API gravity. Fuel, 237, 1119–1131. 10.1016/j.fuel.2018.09.098. [DOI] [Google Scholar]
  18. Pantoja R, Nagarah JM, Starace DM, Melosh NA, Blunck R, Bezanilla F, Heath JR, 2004. Silicon chip-based patch-clamp electrodes integrated with PDMS microfluidics. Biosens. Bioelectron. 20, 509–517. 10.1016/j.bios.2004.02.020. [DOI] [PubMed] [Google Scholar]
  19. Permanyer A, Douifi L, Dupuy N, Lahcini A, Kister J, 2005. FTIR and SUVF spectroscopy as an alternative method in reservoir studies. Application to western Mediterranean oils. Fuel 84, 159–168. 10.1016/j.fuel.2004.06.027. [DOI] [Google Scholar]
  20. Pisner DA, Schnyer DM, 2020. Chapter 6 - Supprt vectpr machine. In: Mechelli A, Vieira S (Eds.), Machine Learning: Methods and Applications to Brain Disorders. Academic Press, London/San Digeo/Cambridge/Oxford, pp. 101–121. 10.1016/B978-0-12-815739-8.00006-7. [DOI] [Google Scholar]
  21. Riley BJ, Lennard C, Fuller S, Spikmans V, 2018. Pyrolysis-GC-MS analysis of crude and heavy fuel oil asphaltenes for application in oil fingerprinting. Environ. Forensics 19, 14–26. 10.1080/15275922.2017.1408163. [DOI] [Google Scholar]
  22. Santhiago M, Henry CS, Kubota LT, 2014. Low cost, simple three dimensional electrochemical paper-based analytical device for determination of p-nitrophenol. Electrochim. Acta 130, 771–777. 10.1016/j.electacta.2014.03.109. [DOI] [Google Scholar]
  23. Shankar R, Jung JH, Loh A, An JG, Ha SY, Yim UH, 2020. Environmental significance of lubricant oil: A systematic study of photooxidation and its consequences. Water Res. 168, 115183. 10.1016/j.watres.2019.115183. [DOI] [PubMed] [Google Scholar]
  24. Songjaroen T, Dungchai W, Chailapakul O; Henry CS, Laiwattanapaisal W, 2012. Blood separation on microfluidic paper-based analytical devices. Lab Chip, 12, 3392–3398. 10.1039/C2LC21299D. [DOI] [PubMed] [Google Scholar]
  25. Sosnowski K, Loh A, Zubler AV, Shir H, Ha SY, Yim UH, Yoon J-Y, 2022. Machine learning techniques for chemical and type analysis of ocean oil samples via handheld spectrophotometer device. Biosens. Bioelectron. X 10, 100128. 10.1016/j.biosx.2022.100128. [DOI] [Google Scholar]
  26. Stravers CS, Gool EL, van Leeuwen TG, Aalders MCG, van Dam A, 2019. Multiplex body fluid identification using surface plasmon resonance imaging with principal component analysis. Sens. Actuat. B Chem. 283, 355–362. 10.1016/j.snb.2018.12.050. [DOI] [Google Scholar]
  27. Suneel V, Rao VT, Suresh G, Chaudhary A, Vethamony P, Ratheesh R, 2019. Oil pollution in the eastern Arabian sea from invisible sources: A multi-technique approach. Mar. Pollut. Bull. 146, 683–695. 10.1016/j.marpolbul.2019.07.015. [DOI] [PubMed] [Google Scholar]
  28. Yetisen AK, Akram MS, Lowe CR, 2013. Paper-based microfluidic point-of-care diagnostic devices. Lab Chip 13, 2210. 10.1039/c3lc50169h. [DOI] [PubMed] [Google Scholar]
  29. Yim UH, Kim M, Ha SY, Kim S, Shim WJ, 2012. Oil spill environmental forensics: The Hebei spirit oil spill case. Environ. Sci. Technol. 46, 6431–6437. 10.1021/es3004156. [DOI] [PubMed] [Google Scholar]

Associated Data

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

Supplementary Materials

1
NIHMS1867198-supplement-1.pdf (1.2MB, pdf)

RESOURCES