Current benchtop instruments with edge filters providing Raman spectra down to 50 cm

 can be used to study the crystallization of PHBHx (polyhydroxybutyrate-hexanoate) without the complexity of larger instruments. By collecting the averaged signal from spherulites, the effects of orientation can be averaged out, enabling the use of multivariate techniques to compare samples. These polymers are being commercialized because they are created by fermentation and are biodegradable, making them an ideal replacement for petroleum-derived polymers that do not biodegrade. In addition, it is understood how to control the polymer’s physical properties by controlling the molecular weight and the percent of sidechains which, in turn, controls the crystallinity. To compare spectra of different samples, it is necessary to remove the effects of polymer chain orientation, which averaging the spectrum of a spherulite will enable.

Although a polymer physicist is well aware of the formation of spherulites in semicrystalline polymers, it is an amazing experience to see them for the first time in an optical microscope with crossed-polar illumination. Figure 1 shows a spherulite in a film of PHBHx polyhydroxybutyrate-13% hexanoate that had been annealed for 36 h at 70 

C. Our understanding of the structure that accounts for what is observed is that the crystallites are nucleated at the center of these features and grow radially along the axis of maximum crystal growth. The sizes of the spherulites will depend on a variety of factors (including the number of nucleation sites, the structure of the molecules, and the cooling rate or annealing temperature), and can range between micrometers to centimeters. One might assume that the radial direction corresponds to the chain axes, but, in fact, it is often a different axis. In PHBHx, it is the 

-axis that is the axis of optimum growth (1), and 

 is also the axis of the unusual interaction between one of the methyl protons on one chain in the unit cell and the carbonyl in the second chain in the unit cell (2).

The spherulites are actually a three-dimensional (3D) spherical structure comprised of lamellae that align along the radial axes. Lamellae are polymer chains that fold back and forth on each other in a direction normal to the growth direction. Lamellae define the crystalline regions, with the disordered chains between the lamellae being amorphous. For those readers who are unfamiliar with these concepts, I recommend looking at the Wikipedia article on polymer spherulites. In the case of PHBHx, the lamellae actually twist around the growth axis (3).

These concepts should affect the Raman polarization behavior. In the instrument being used, which is the XploRA Plus, the laser (532 nm) polarization is east-west (EW) at the microscope stage. Therefore, in an EW line scan through the center, the laser will be polarized along the 

-axis. In a north-south (NS) line scan, the laser polarization will be normal to 

-axis would not be unique in this system because the lamellae are rotating around the growth axis (3). If we do a line scan at 45

 to NS or EW, we expect to see a spectrum that is an average.

Figure 2 shows profiles through the center nucleation site (as accurately as could be determined) for the three bands that have been associated with crystallinity. The phonon at 80 cm

 represents beating of the two chains in the unit cell. The carbonyl band is actually several overlapping bands where the frequency downshifts in the crystalline state. The CH band at 3000 cm

 becomes a doublet in the crystalline form, and it has been interpreted to be a splitting of the band at 3000 cm

 when the chains in the unit cell interact (4). Except for the carbonyl profiles in the 0% Hx sample, the diagonal profiles (not shown) are approximately equal to the average of the vertical and horizontal profiles. We believe that the reason for this discrepancy is the large increase in intensity of the crystalline component in the vertical profile of the 0% sample, as shown in Figure 3.

Figure 3 shows the spectra (from top to bottom) of narrow horizontal and vertical slices of the low crystallinity samples, and then horizontal and vertical slices of the high crystallinity samples. To validate the conjecture in the previous paragraph, I show the carbonyl region expanded in Figure 4. Note that the spectra of the low crystallinity material at the top shows only approximately a 30% increase in intensity in the vertical slice, but in the spectra of the high crystallinity material at the bottom, there is a very large increase in the crystalline component for the vertical profile. In addition, I have placed arrows to indicate some subtle features buried in the broad amorphous bands.

I am currently recording line profiles with a higher groove density grating and more closely spaced data points. Under these conditions, we are clearly able to correlate the bands in the micrograph to differences in the Raman spectra. Figure 6 shows one such scan, revealing the variation in the intensity of the carbonyl band in a horizontal cut through the center of a spherulite of the pure PHB sample. What is curious is that the 0% sample does not show the banding in the micrograph recorded between crossed polars. By systematically measuring radial slices of spherulites, we should be able to understand what the role of polymer orientation has to do with the bands seen in the micrographs.

This project started because I wanted to compare spectra of different PHBHx samples with different concentrations of Hx. Because I need to eliminate polarization and orientation effects on relative intensities to compare samples, it occurred to me that mapping a spherulite and calculating the average spectrum would eliminate these effects. This idea stunningly opened up new possibilities. Not only can I see intensity ripples as a function of radial distance, but by comparing spectra between horizontal and vertical cuts I can start analyzing the polarization and orientation behavior. Clearly, there is a lot of information coming from these measurements that we will be able to exploit in a more complete analysis of the spectra of this material.

(1) Liu, C.; Noda, I.; Martin, D. C.; Chase, D. B.; Ni, C.; Rabolt, J. F. Growth of Anisotropic Single Crystals of a Random Copolymer, Poly[(R)-3-hydroxybutyrate-co-(R)-3-hydroxyhexanoate] Driven by Cooperative –CH···O H-Bonding. 

, 111–118. DOI: 10.1016/j.polymer.2018.08.046.

(2) Sato, H.; Dybal, J.; Murakami, R.; Noda, I.; Ozaki, Y. Infrared and Raman Spectroscopy and Quantum Chemistry Calcula- tion Studies of C–H···O Hydrogen Bondings and Thermal Behavior of Biodegradable Polyhydroxyalkanoate. 

, 744–747, 35-46, DOI: 10.1016/j.molstruc.2004.10.069

(4) Adar, F.; Street, R.; Noda, I. Isothermal Crystallization of Polyhydroxyalkanoate (PHA) Utilizing Raman Spectroscopy to Follow Chain Packing As Well As Molecular Motion, 

Articles

Column: Molecular Spectroscopy Workbench

Studying the crystallinity of biodegradable polymers can be done without the complexity of larger instruments. Here’s how.

Measuring the Crystallinity of PHBHx with Varying Amounts of Sidechains on a Benchtop Instrument

Artificial intelligence (AI), and its subfield machine learning (ML), are major buzzwords in today’s technology world. In a two-part series, let’s begin to take a look under the hood and behind the scenes to see what AI is and its applicability to analytical chemistry and spectroscopy for future discussion and elaboration. What are the benefits and limitations, as well as the praises and detractions, of AI? How does AI relate to chemometrics?

As with any modeling technique, it might serve us well to remember Robert A. Heinlein’s reference: “There ain’t no such thing as a free lunch (TANSTAAFL)” (1), and supercalibratestatisticexpedientalgorithmadnauseous, a play-on word inspired by the wisdom of Mary Poppins (2), reminding us about the hype of new algorithms. The term was coined to mock the then-current practice of trying all-possible combinations of available algorithms and test statistics to automate the process of developing calibration models for near-infrared (NIR) analysis. It took a long time, but eventually, the NIR community recognized the futility of that methodology. Remember, if you are going to make mistakes, automation allows you to make them more quickly and repetitively. One must be cautioned when using powerful and automated regression-based methods to remember that modeling algorithms are only able to build an accurate predictive model based on the information they are presented. As Mark Twain once said, “There are three kinds of lies: lies, damn lies, and statistics” (3–5). And as Rasmus Bro, a leader in the field of chemometrics, has said:

“As most people are aware, there is currently a hype on machine learning and artificial intelligence (AI), and that is all fine and good. Part of the hype is that these methods are sometimes oversold, and so then people are relieved that now they don’t have to think about things, saying, ‘I don’t have to think about experimental design, sampling error, analytical quality, etc.’ And when that happens, those projects typically fail, because these new and excellent methods don’t actually replace the responsibility for you to actually know what you are doing” (6).

With all these precautionary tales in mind, we now move ahead with our initial overview discussion of AI and chemometrics.

Basically, AI and machine learning (ML), as they are used today, consist of applying signal processing and multiple layer neural networks to model a function or set of functions that has been loosely associated with the intelligent activity of living organisms. The same challenges for building successful models may be present in all modeling activities, including chemometrics; these may include sparse or limited data, noisy data, unknown limits to the potential data and predicted values, a lack of basic understanding of the nature of the data, and so forth. In chemometrics, it is generally accepted that a chemist or chemical engineer is deriving (or at least interpreting) a quantitative, qualitative, or classification model for a chemical system, and therefore has domain knowledge of the nature of the data and the expected model prediction results. 

, or fundamental knowledge of the system one is working with, is essential to avoid major errors, using computer modeling. That is why, if one is using AI to build a self-driving automobile, it would be very handy to have expert drivers involved during the modeling process. Without domain knowledge, one will undoubtedly see major errors in relying on models generated without such knowledge used as input information. These caveats also apply to modeling weather (where one asteroid or erupting volcano changes everything), the future economy (with the potential for wars and famines), and other unbounded or somewhat unknown domains.

For an “official” AI response, we asked the question, “How are machine learning and chemometrics related?” to the program Chat.OpenAI.com (7), an AI program in development. The AI answer is transcribed below. Note that as the AI program “learns,” its responses will change over time.

“Machine learning and chemometrics are both fields that involve the use of statistical and computational methods to analyze and interpret data. However, they are applied in different contexts and for different purposes. Machine learning is a subfield of artificial intelligence that focuses on the development of algorithms and models that can learn from and make predictions or decisions based on data. These algorithms and models are trained on a set of labeled data, which consists of input data and corresponding output data, and are then used to make predictions or decisions about new, unseen data. Machine learning methods are widely used in a variety of applications, including image and speech recognition, natural language processing, and predictive modeling.

Chemometrics, on the other hand, is the application of statistical and mathematical methods to the analysis of chemical data. It is used to extract meaningful information and knowledge from chemical data sets, such as those obtained from spectroscopic or chromatographic measurements. Chemometrics methods are used to analyze and interpret chemical data in order to gain insights into the underlying chemical processes and to make predictions 

about the properties and behavior of chemical substances.

Both machine learning and chemometrics involve the use of statistical and computational techniques to analyze and interpret data. However, while machine learning is generally applied to a wide range of data types and applications, chemometrics is specifically focused on the analysis of chemical data.”

As with many technical matters, it is useful to begin by discussing the various concepts involved with the subject of interest, and compose a list of (and a set of definitions for) those terms. If one is to develop AI technology and ML algorithms, it is helpful to know what 

 are. These concepts seem self-evident or intuitive to sentient beings; however, clear definitions of terms are helpful for classification and understanding. There have been many attempts to precisely define 

 from Merriam-Webster includes the following:

1) “...(T)he ability to learn or understand or to deal with new or trying situations: Reason; also: the skilled use of reason; and 2): the ability to apply knowledge to manipulate one’s environment or to think abstractly as measured by objective criteria (such as tests)” (8).

The definition of intelligence from Oxford Learner’s Dictionaries is

“...(T)he ability to learn, understand, and think in a logical way about things; the ability to do this well....” (9).

A definition of AI from the Oxford English Dictionary is as follows:

“(T)he capacity of computers or other machines to exhibit or simulate intelligent behavior; the field of study con- cerned with this. Abbreviated AI” (10).

The often-referred to Wikipedia source has a more detailed definition:

“AI is intelligence demonstrated by machines, as opposed to the natural intelligence displayed by animals including humans....”

Wikipedia notes that older definitions of AI referred to machines that mimic human cognitive skills, such as learning and problem solving. A more modern definition if AI uses phrases such as “machines that act rationally,” which somewhat muddies the waters in terms of a precise definition (11). We might summarize the various ideas about a definition of AI, for our purposes, as:

“(A)n automated technology using computers and algorithms that is capable of simulating learning, thought, creativity, rationality, and analysis; and that is able to perform routine or complex tasks that have been historically accomplished by humans.”

ML may then be defined as a subfield of AI where computers (machines) are programmed in such a way that they are able to self-iterate and optimize problem solving for modeling data in an automated way, using computational algorithms specifically designed for such purposes. Let us be frank here: the perceived intelligence or learning is a computer simulation using automation and optimization routines of computer algorithms—no more and no less. Once programmed with appropriate algorithms, the computer (machine) is capable of simulating problem solving and learning functions that are normally attributable to humans.

The foundation of learning AI is to understand the variety of ML (deep learning) algorithms, and then learn how to apply these ML algorithms for various applications. Next, one may design a user interface to interact with the AI programming. The main subfields involved include using the various models for defining an AI problem, learning how to design AI programs, deciding how humans will interact with the AI programs, and learning how to modify and teach an AI program. One may add here that it is also important that experts evaluate or check the output of an AI program to test its legitimacy.

It may surprise those analytical chemists completely unfamiliar with AI that many of the chemometrics (multivariate) algorithms they are familiar with are also classified as AI algorithms—these are algorithms that are routinely taught and used by AI programmers. We refer our readers to the following articles for a head start on this discussion (12–14). This previous series of articles provides definitions and a set of references for signal preprocessing, component analysis, quantitative (calibration) methods, qualitative (classification) methods, and programming platforms.

Table I shows a comparison of the chemometrics algorithms and the AI and ML algorithms as derived by the authors. Table II is a comparison table generated by an AI program (Chat.OpenAI.com) that compares AI and ML algorithms with chemometrics algorithms using its own AI terminology.

Figure 1 shows a basic flow chart of commonality between AI approaches and chemometric modeling. Note that any modeling system is only capable of accurately modeling the data it has been shown or data that follow a logical framework, such as game theory. Within this limited framework, modeling techniques are able to interpolate within its known boundaries and generally produce an accurate result, or at least, a result with a bounded estimate of error. If the modeling system has not seen the data, or if there is not a rational or known framework for potential results, then the results are merely an automated or computerized guess.

The concept of an open or closed model is helpful at this point, because we may imagine what could happen when we model a constrained data set and when the nature of the entire potential data set is not well understood. As Figure 2 illustrates, the limitation of models using powerful fitting algorithms is the “completeness” of the data used and the assumptions made about the nature of the data. Modeling “uncharted” data in science can be, and often is, problematic and unpredictable. Model prediction outside of a closed system is based on hypothesized assumptions about the nature of the data. If the nature of the data and the assumptions about the data are unknown or not clearly understood, then the ability of any modeling system to predict a result accurately is also unknown. Therefore, caution when using extrapolation in a modeling system is wise.

One can imagine that game playing is a good example of what a modeling system or AI could easily master, because the universe of possible game moves is well-known and constrained to a finite set of options. With the high speed of computation possible in today’s computers, a modeling system is able to compute all possible combinations of moves and describe the best option (or options) available based on probability, separate decision algorithms, or a competitor’s responses to its moves. It would seem that a well-written AI or computer system would be the master of game playing and associated tasks. A next level of complexity may arise in self-driving automobiles, but this activity is still fairly well-defined under normal circumstances. That being said, there are occasionally extreme circumstances, such as weather, unusual sky or light reflections, sudden objects appearing on the roadway, multiple car collisions, and so forth, that would be extremely difficult to model and negotiate.

AI computers would also seem to be able to master inventory, logistics, traffic control, basic war games, and prediction of basic human behaviors—all with relatively constrained circumstances. Even a form of pseudo-creativity would seem to be relatively straightforward with access to the Internet and computerized data processing. For new inventions, a computer system is capable of reviewing the prior patent art and literature, evaluating the known physical laws and basic science regarding the subject, and then “suggesting” potentially novel ideas not yet tried. However, issues of compassion, love, sacrifice, selflessness, and generosity would be extremely difficult to model; these are also factors that weigh in heavily for positive human decisions.

So let us get back to the task at hand, which is to discuss AI for applications in analytical spectroscopy. Potential uses would be for instrument alignment (instrument calibration) and diagnostics, instrument measurement conditions and parameter optimization, experimental design, signal preprocessing, quantitative and qualitative calibration, library searching, and automated imaging and image enhancement functions.

There are several steps in developing both AI and chemometric models that are basic and somewhat unchangeable. These would include the following steps, which can be accomplished either manually or in an automated fashion:

1) Define the problem that needs to be solved.

2) Define the type and amount of data required to train the model.

8) Repeat the process to optimize the model .

There are three main skill sets that are essential in learning how to work with AI. They are as follows:

1) Learn how to operate basic software tools (such as Python), language, and syntax, and the types of algorithms needed.

2) Learn the basics of machine learning algorithms: regression, k-nearest neighbors (KNN), support vector machine (SVM), artificial neural networks (ANNs), reinforcement learning methods, computer vision tools, and so forth.

3) Learn how to specifically solve the problems you are interested in by researching other examples and those researchers already problem solving in your area of interest.

In Part I of this two-part set of introductory articles on AI, we have attempted to give a very concise and top-level overview of AI, a technology topic that is now taking the world by storm. AI discussions are now ubiquitous across media, politics, business, academia, and especially in science and engineering. There is so much more to be said, but we have here brought an introduction to a topic we may continue to explore through additional articles and social media, such as in the “Analytically Speaking” podcast series, Episode 9, featuring Rasmus Bro as the guest speaker (6). In addition, in Part II of this written series, we plan on exploring actual applications of AI and ML to different vibrational spectroscopy methods, such as Raman, Fourier transform infrared (FT-IR), NIR, and UV–visible (UV-vis) spectroscopic techniques.

 (John Wiley & Sons, New York, NY, 1991), pp. ix.

 (University of California Press, Berkeley, CA, 2010), p. 228. (Attributed to Twain dictating this to his secretary in Florence, April, 1904, in “Notes on “Innocents Abroad”). 

http://gutenberg.net.au/ebooks02/0200551h.html

https://www.york.ac.uk/depts/maths/histstat/lies.htm

(5) Velleman, P.F. Truth, Damn Truth, and Statistics. 

https://doi.org/10.1080/10691898.2008.11889565

(6) Spectroscopy, Analytically Speaking Podcast, Episode 9, Feb. 1, 2023. 

https://www.spectroscopyonline.com/analytically-speaking-podcast 

(7) OpenAI, Build Next-Gen Apps With OpenAI’s Powerful Models. 

(8) Merriam-Webster Dictionary, Intelligence. 

https://www.merriam-webster.com/dictionary/ intelligence#:~:text=1%20%3A%20 the%20ability%20to%20learn,noun 

(9) Oxford Learners Dictionaries, Intelligence. 

https://www.oxfordlearnersdictionaries.com/us/definition/english/intelligence

(10) Oxford English Dictionary, Artificial intelligence, n. 

https://www.oed.com/viewdictionaryentry/Entry/271625

(11) Wikipedia, Artificial Intelligence. 

https://en.wikipedia.org/wiki/Artificial_intelligence

(12) Workman, Jr, J.; Mark, H. A Survey of Chemometric Methods Used in Spectroscopy, 

(13) Workman, Jr, J.; Mark, H. Survey of Key Descriptive References for Chemometric Methods Used for Spectroscopy: Part I, 

(14) Workman, Jr, J.; Mark, H. Survey of Key Descriptive References for Chemometric Methods Used for Spectroscopy: Part II, 

There were some errors in the October 2022 Chemometrics in Spectroscopy column that evaded detection until after the column was published (1). Therefore, we hereby publish these corrections:

 was given as 10000; this number contains one more zero than the correct amount. The correct expression for 8

In the first text column on page 20, under “Errors in Representation,” there were several instances of inserting the incorrect subscript to define the number system the specified number belongs to. For example, in the passage“. . . 2

 = 0.25 is too small...” it is clear that the digit 2 cannot belong to a binary number, since a binary number can only contain the digits 0 and 1, as we stated several times. The correct statement is “. . . 2

The same error occurred several times over the subsequent six lines. Instead of belaboring each occurrence of the error, here we simply present the corrected passage, starting with “for example”:

, which is again too large. We note that 0.0101

, which is still too large, while 0.01001

, is again too small. Finally, we are closing in on 0.3

In Table III, the last two entries under the column heading “Truncated Decimal Number” simply repeat the second-to-last entry instead of reflecting the continuing truncation of pi. The correct table entries are 3.141000000000000

(1) Mark, H.; Workman, Jr, J. Decimal Versus Binary Representation of Numbers in Computers. 

https://doi.org/10.56530/spectroscopy.mm1179p4

Columns: Chemometrics in Spectroscopy

Are you intrigued by artificial intelligence, but unsure what it really means for analytical chemistry? Read on.

Artificial Intelligence in Analytical Spectroscopy, Part I: Basic Concepts and Discussion

Automatic coal-rock recognition (ACRR) is of considerable theoretical and practical significance for unmanned coal mining. To the best of our knowledge, this is the first study to assess laser-induced breakdown spectroscopy (LIBS) combined with an artificial neural network (ANN) for automatic coal-rock recognition. Each sample in this study was subjected to LIBS testing and spectrum collection 20 times in the air, and the average value was taken as the LIBS data. Spectral data were optimized and dimensionality reduction was performed using partial least-squares discriminant analysis (PLS-DA). The 10 selected wavelength lines were used to construct a simplified spectral model (SSM). The ANN based on SSM was designed to classify the coal and rock. The results demonstrated that LIBS combined with an ANN has a high recognition accuracy rate, providing a rapid and accurate coal-rock recognition method for unmanned coal mining.

Intelligent coal mining and processing are important for sustainable and safe coal mining (1,2). Automatic coal-rock recognition (ACRR) is one of the key technologies for intelligent coal mining and processing. ACRR is the “eye” of unmanned coal mine equipment, which can replace human eye observation and reduce staff. It is an indispensable key link in the process of realizing the unmanned work surface. ACRR technology can provide a basis for the shearer to adjust the cutting height of the drum, which can reduce coal loss and increase resource recovery rate, reduce the content of the gangue in the falling coal, and improve the quality of raw coal. More importantly, ACRR can help picks to avoid rocks in time, greatly reducing the possibility of gas explosions caused by dangerous temperatures generated by picks cutting rocks (3,4). The process can significantly reduce the pick wear, extend the life of the device, and reduce the frequency of hardware replacement, which obviously decreases the costs. Sophisticated ACRR technology can eliminate the need for coal miners at the working face. In addition, ACRR also realizes the rapid separation of coal gangue in coal preparation plants. All in all, coal-rock recognition has very important theoretical and practical significance for intelligent coal mining.

To date, considerable research has been devoted to various coal-rock recognition methods (5). These techniques are categorized as passive-detection (PD) recognition and active-detection (AD) recognition, according to whether the technique depends on cutting rocks.

The PD–ACRRs include cutting force monitoring (6), thermal detection, vibration detection (7), and acoustic detection (8). The differences in physical quantity caused by cutting coals or rocks are used for recognition. In the recognition process of PD–ACRR, this kind of technique is likely to take effect only after shearer picks cutting rocks.

The AD–ACRRs include γ-ray detection (9), radar detection (10), and image analysis (11,12). The signal monitored by the active ACRR technique is caused by the inherent characteristics of coal and rock, rather than cutting the different substances. Therefore, AD–ACRR technology does not have to wait for “errors” (cutting rocks) to appear before “correcting” (avoiding cutting rocks). In theory, AD–ACRR technology can achieve “error-free,” unlike PD–ACRR.

However, currently available active ACRR techniques on the market have strict requirements. When coal rock’s radioactive elements are low, or when vermiculite is mixed into the coal seam, γ-ray detection cannot be used. Electromagnetic waves can be seriously attenuated in thick coal seams, and radar detection fails. The camera cannot obtain a clear image in the coal mining working face, so image analysis is often ineffective.

Some other active methods, such as X-ray fluorescence (XRF) and prompt gamma neutron activation analysis (PGNAA), have proven effective. But PGNAA machines are very bulky (minimum weight is 3200 kg), and the neutron source has potential health hazards and is subject to strict regulatory requirements.

In view of the current research status, a safer, low cost, high perfor

mance and active ACRR technique is still necessary.

Several laser-based methods have been recently developed to analyze coal rock (13–17). Laser-induced breakdown spectroscopy (LIBS) is a successful analytical technique for the determination of the elemental composition of materials (18–22). Each plasma radiation spectrum line corresponds to unique transitions in atoms, ions, or molecules. These are used as “fingerprints” to identify the sample structure (23,24). Therefore, LIBS can be used to classify and identify different substances. The advantages of LIBS include the nondestructive assessment of small samples, fast detection, no initial preparation, no radioactive source, and online and in situ analysis (17,25). This shows that the use of LIBS for ACRR is very promising. LIBS uses chemometric methods such as linear or parametric correlation, principal component analysis (PCA), and partial least-squares discriminant analysis (PLS-DA) to identify organic compounds (26,27). LIBS combined with artificial intelligence algorithms such as artificial neural networks (ANN) are very accurate (28). ANN with LIBS spectroscopy is used to classify and predict coal rock characteristics. ANN was recently developed for LIBS data processing for material identification (29–32). Hai Yan-he (33) presented the research of gas explosion mechanism, in accordance with the free radical mechanism of vapor reaction and the spectral theory of polyatomic molecules. As long as the laser pulse energy intensity is strictly controlled, the safety of using laser instruments in the coal mining working face can be guaranteed. Hai understands the conditions of laser safe application in coal mines, which provide theoretical support for LIBS application to the coal mine working face.

This paper describes coal-rock recognition using LIBS combined with ANN. PCA and PLS-DA were used to reduce the dimension of input space in ANN models while decreasing the training time and preserving, or even improving, model accuracy. ANN recognizes and determines the characteristics of different types of samples. This paper applies LIBS technology to coal rock recognition and makes a detailed demonstration.

LIBS spectra of “coal” and “rock” were collected, the spectral data were analyzed and processed comprehensively, and PLS-DA was used to optimize, process, and reduce the spectral data, thereby establishing the LIBS ACRR simple spectral model (SSM). Finally, we tried to use an artificial neural network to classify and predict the “coal” and “rock” to achieve the purpose of coal rock recognition. The results of the coal rock recognition can provide an operational basis for unmanned coal mining equipment. Figure 1 shows the workflow of the LIBS ACRR system.

A Q-switched Nd: YAG laser (ICE 450 laser power supply Quantel) with a 1064-nm wavelength, 7±2 ns pulse duration, and 40 mJ energy per pulse was used as an excitation source. The laser was operated in external trigger mode, and it was triggered by a digital delay generator for synchronization, with a repetition frequency of 1 Hz. The Nd:YAG laser beam was focused on the sample through a convex lens (focal length: 50 mm) at normal incidence. To avoid excessive ablation, the sample was mounted on a programmable translation stage. The laser was focused on a spot with a diameter of approximately 1 mm. The experiment was conducted in the laboratory.

An optical system collected the sample’s emitted light and transmitted it to the spectrometer (20-01-13-A Avantes) using optical fibers. The spectra were acquired in an air atmosphere. A schematic of the LIBS system and laboratory apparatus used in this study was shown in Figure 2. 

Because of the different elemental composition of coal and rock, the LIBS spectrum produced by them were distinctive. We used this set of LIBS experimental equipment to generate and collect spectra, and then realized the classification and recognition of coal rock by analyzing the spectra.

In this paper, four types of coal samples were used as “coal” in ACRR, while gangue samples were used as “rock.” The samples were provided by the China Coal Technology Engineering Group.

The four types of coal were raw coal, clean coal, medium coal, and slime coal. All coal samples were in the form of coal powder. 

 referred to coal that only excludes visible gangue without any other processing. After the coal was washed, the coal changed into high-quality coal suitable for special purposes, which was called 

 was an intermediate product after the recovery of clean coal and the removal of gangue during coal separation and processing. These three types of coal were used to simulate different qualities of coal. 

 generally referred to the semi-solid matter formed by coal powder containing water, which used to simulate the coal wetted by the water used to cool the shearer picks.

The impact of the laser pulse would oscillate the surface of the coal samples if the coal powders were not prepared as briquettes. Moreover, to ensure the stability and accuracy of the laser focus position, which directly affects the results of the LIBS experiment, the surface of the coal samples needs to be tight, uniform, and as flat as possible. Thus, all coal samples were sieved and compressed with 20 meshes at 10 Mp. The “rock” (gangue) was in its original stone form. The samples were shown in Figure 3.

In the LIBS experiment, each LIBS spectrum was acquired with 20 laser pulses at different sample points to reduce the data variations because of changes in experimental conditions. The emission spectra were recorded in a 240–480 nm wavelength range and contained intensity information of 6074 wavelengths. The total number of samples was 50, and the samplings were performed 20 times at different points on the sample surface as mentioned above. This meant that in the end all the data form a 6067 × 1000 data matrix. Therefore, it was necessary and important to establish a simplified spectral model (SSM) to quickly and accurately analyze the spectral data to ensure the feasibility of ACRR.

Dimension reduction and classification was the main purpose of the SSM. Partial least square discriminant analysis (PLS-DA) is a special kind of regression analysis used for classification problems. In the case of PLS-DA, its principal components are linear combinations of features; the number of these components gives the dimension of the transformed space. As a standard, the components are orthogonal to each other. Therefore, an appropriate number of feature components are selected to establish the SSM.

The variable importance in projection (VIP) is a popular measurement tool in the PLS-DA literature (34). The VIP quantifies the response of each variable summed over all components and categorical responses, enabling the measurement of the global effect of each wavelength. The VIP score of a variable is calculated as a weighted sum of the squared correlations between the PLS-DA components and the original variable. The weights correspond to the percentage variation explained by the PLS-DA component in the model (35,36).

The number of spectral lines of a simple spectral model (

) was the key issue. Figure 4 showed the SSM with different numbers of spectral lines. Spectrum lines of the SSM were selected according to the VIP value from high to low, and the number is small to large.

The most important advantages of ANNs are: simultaneous action in parallel structures because they are composed of nonlinear information processors; good adaptation with the surrounding environment; and accurate mapping of the input and output data (37). The ANN model consists of an input layer, one or more hidden layers, and an output layer. Each layer consists of a basic unit called a neuron or node. In biological structures, the units are connected by a synapse. The ANN synapse simulates the weight and bias, which connect each neuron to all neurons in a successive layer, as shown in Figure 5.

The equation 1 of each neuron output is expressed as reference (38):

 are input, weight, and bias, respectively. The best example of this function is called a 

Because of the differences in the coal and rock characteristics of each coal mining face, before each new face is put into use, a certain amount of coal and rock spectra need to be collected on the spot for the new face to establish SMM and train ANN. The ANN has two steps: 1) network learning with the reference samples’ spectra (training); and 2) identification of an unknown spectrum by assigning it to a reference material class (test). For the training process, the learning data input was divided into three parts: training (70%); self-validation (15%); and testing (15%). After training and self-validation, the models were validated using the test data. Unknown samples were identified after an optimized low error network was constructed. The identification process was based on the ANN’s ability to detect the degree of similarity between the new spectrum and each reference spectra used in the training process.

The training was performed by feeding many inputs into the network,and comparing the corresponding outputs with the targets by computing the error. The mean squared error (MSE) is the most common error measurement method in ANN training.

Tests were conducted to study the coal-rock recognition capabilities of LIBS and ANN. In the first step, LIBS tested 10 group samples in ambient air for PLS-DA analysis and training of the ANN. Each group included one raw coal sample, one medium coal sample, one clean coal sample, one slime coal sample, and one gangue sample. A total of 50 samples were tested. Each LIBS spectrum was acquired with 20 laser pulses at different sample points to reduce the data variations because of changes in experimental conditions. The delay time was approximately 1.28 μs.

The LIBS comparison spectra of one set of samples averaged (20 laser pulses at different sample points) obtained using the spectrometer are shown in Figure 6.

The PLS-DA score scatters plot of all 50 samples are shown in Figure 7. The elemental composition and content of the four types of coals are different, which leads to distinct LIBS spectra for each type. This distinction was reflected in the PLS-DA scatter plot, as shown in Figure 7. Although the coal and rock were significantly different, the types of coal analyzed also had their differences. The scattered points of different types of coal and rocks were gathered in different positions in the PLS-DA scatter plot. This conclusion is not only used to recognize the coal and rock, but to also provide a theoretical basis for the future use of LIBS in coal quality testing. The spectrum of gangue and coal had proven to be different in classification tests, and the degree of distinction can be used to accurately predict the classification of coal and rock. The variation of the first two principal components was 0.669 and 0.152, respectively. The square of the percentage of information retained from the original data, which is 

As mentioned above, a single spectrum was composed of 6067-dimensional vectors. Therefore, it was necessary and important to reduce the dimensionality by applying the simple spectral model.

) determines the effect and influence of SSM simplifying the original spectral data structure and the accuracy of the ANN based on SSM to identify coal and rock. The number 

 of SSM variables should increase from 1, and variables are selected one by one according to the VIP value of variables from high to low. 

 must enable the established SSM to retain the interpretation of the original data; that is, enable the ANN based on SSM to achieve good coal and rock recognition accuracy, and 

 must be as small as possible at the same time.

It is worth mentioning that when we classified and identified the coal and rock spectra, we considered not only the statistical significance, but also the physical significance. In this study, considering that coal was a very complex substance, including sulfides and oxides, we decided to use 10 characteristic wavelengths to construct the SSM under the balance of simplification and model accuracy. Since carbon is of obvious importance for the analysis of coal, and because carbon atoms have a large number of peripheral electrons that are difficult to ionize, their spectral intensity is affected. This may be the reason why the VIP value of carbon is not so high. So even if the VIP value of the carbon line (247.82 nm) is not in the top ten (VIP = 2.51), we decided to use it in SSM. The selected 10 wavelengths of SSM were listed in Table I. Taking the spectrum of a gangue sample and a clean coal sample as an example, the spectrum of 6074 wavelengths data was reduced to 10 wavelengths by SSM, as shown in Figure 8.

By using SSM, the data volume of single sample spectra decreased from 6074 wavelengths to 10 wavelengths, which was a 99.8% data reduction as compared to the raw data.

The simplified spectral model was used for the ANN input. The ANN classifier was designed by Matlab to have 10 input neurons (10 variables of SSM), 10 neurons in the hidden layer, and 2 neurons (coal/rock) in the output layer.

The ANN classifier evaluation results were shown in Figure 9. Figure 9a presented the training set, validation set, and test set error of each epoch. One epoch was equivalent to training once with all samples in the training set. As the ANN training times increased, the training error decreased. The training stopped when the best verification performance (MSE = 10

) of the ANN was achieved at the 73rd time (epoch), where there were 73 epochs in Figure 9a. The error curves of the three sets had a good correlation. The trend of the curve gradually decreased until a suitable target was reached. The convergence was reached at epoch 73, which may have occurred because the training sample size was small.

Figure 9b showed the correlation coefficients of the training set, the validation set, the test set, and all samples set. The 

 values of four sets are all greater than 0.97. The coal-rock interface was random during the actual cutting process. Therefore, we made a physical model of a simulated coal mining face, with randomly distributed coal-rock interfaces to verify the effectiveness of the ANN, as shown in Figure 10. The model is 1.7 meters long and 1.2 meters wide. The sampling points were separated by 0.1 meters, and the recognition result of the sampling points was used as the recognition result of the square where the sampling points were located. A total of 204 sampling points were evenly distributed in the area to be identified, and the coal and rock distribution of the simulated coal mining face would be described by these 204 squares. The results of coal and rock recognition at sampling points were shown in different colors. Black indicated that the recognition result of the sampling point was “coal,” and gray indicated that the recognition result of the sampling point was “rock.”

The results of the 204 test points are shown in Figure 11. The recognition results were analyzed combined with the coal-rock sample, as shown in Figure 12. The red box in the figure marked the sampling points that were clearly recognized incorrectly. Because such sampling points were close to the coal-rock interface, the amount of gangue in the coal increased significantly, which lead to a misjudgment. The blue box marked the sampling point for which the recognition result was suspicious. There were both coal and rock in the area, so the reason for determining it as rock may be that the sampling point did fall on the rock, or it may be that the sampling point fell on the coal but was misjudged as rock. Therefore, the number of sampling points for coal and rock recognition errors is 2–4, and the accuracy of sampling point recognition is 98.03% to 99.01%. The recognition results of coal-rock recognition were close to that of the actual interface. The experimental results proved that the new ACRR method presented in this paper is accurate and effective, and has the inherent advantages of LIBS, such as (but not limited to) speed, safety, in situ measurement, and low cost.

This study classified coal and rock using LIBS combined with PLS-DA and an ANN algorithm for ACRR. The classification model consisted of three steps. First, 50 samples were tested via LIBS in ambient air, and the emission spectra were recorded in a 240–480-nm wavelength range. We then used PLS-DA to analyze the spectra of all the samples to obtain 10 wavelengths to construct the SSM. Finally, we used ANN to classify the 50 samples based on the SSM. ACRR was achieved using this classification.

The results showed that the classification model’s accuracy was over 97%. Furthermore, the data volume decreased from 6074 wavelengths to 10 wavelengths, demonstrating that the SSM is accurate and less time-consuming than other methods. These experiments verified that automatic coal-rock recognition by LIBS combined with an artificial neural network can be used for the efficient, fast, and accurate classification of coal and rock. It is worth mentioning that, from the PLS-DA scatter plot (Figure 7), it can be seen that the difference not only exists between coal and rock, but also has a good distinction between different types of coal, providing a theoretical basis for the future use of LIBS in coal quality testing.

The authors acknowledge the financial support from the National High-tech Research and the Development Program of China under Grant 2014AA052602 and the National Basic Research Program of China under Grant 2014CB046306.

The authors declare that they have no conflict of interest.

(1) Wang, J. Development and Prospect on Fully Mechanized Mining in Chinese Coal Mines. 

(2) Yu, Z.; Zhu, S.; Wu, Y; et al. Study on the Structural Characteristics of the Overburden Under Thick Loose Layer and Thin-bed Rock for Safety of Mining Coal Seam. 

(3) Wang, H; Zhang, Q. Dynamic Identification of Coal-Rock Interface Based on Adaptive Weight Optimization and Multi-Sensor Information Fusion. 

(4) Li, W., et al. Coal and Coal Gangue Separation Based on Computer Vision. Paper presented at 2010 Fifth International Conference on Frontier of Computer Science and Technology 

(5) Li, Y.; Cheng, G.; Chen, X.; Liu, C. Coal-Rock Interface Recognition Based on Permutation Entropy of LMD and Supervised Kohonen Neural Network. 

(6) Yang, J. J.; Fu, S. C.; Jiang, H.; Zhao, X. Y.; Wu, M. Recognition of Cutting Hardness of Coal Rock Properties Based on Fuzzy Criteria. 

(7) Wang, B.; Wang, Z.; Zhu, S. Coal-Rock Interface Recognition Based on Time Series Analysis. Paper presented at 2010 International Conference on Computer Application and System Modeling (IC- CASM 2010) 

(8) Junkai, X.; Wang, Z.; Wanzhi, Z.; Yanpeng, H. Coal-Rock Interface Recognition Based on MFCC and Neural Network. 

International Journal of Signal Processing, Image Processing and Pattern Recognition

(9) Asfahani, J.; Borsaru, M. Low-Activity Spectrometric Gamma-Ray Logging Technique for Delineation of Coal/Rock Interfaces in Dry Blast Holes. 

(10) Rubin, L. A.; Fowler, J. C. Ground-Probing Radar for Delineation of Rock Feaures. 

(11) Sun, J; Su, B. Coal–Rock Interface Detection on the Basis of Image Texture Features. 

(12) Wu, Y. X.; Tian, Y. M. Method of Coal-Rock Image Feature Extraction and Recognition Based on Dictionary Learning. 

(13) Musazzi, S., et al. “Elemental Analysis of Coal by Means of the Laser Induced Breakdown Spectroscopy (LIBS) Technique.” Paper presented at IEEE Sensors Applications Symposium Proceedings. 

(14) Li, X.; Wang, Z.; Fu, Y.; Li, Z.; Liu, J.; Ni, W. Application of a Spectrum Standardization Method for Carbon Analysis in Coal Using Laser-Induced Breakdown Spectroscopy (LIBS). 

(15) Redoglio, D.; Golinelli, E.; Musazzi, S.; Perini, U., et al. A Large Depth of Field LIBS Measuring System for Elemental Analysis of Moving Samples of Raw Coal. 

(16) Liangying, Y., Jidong, L.; Wen, C.; ge, w.; Kai, S.; Wei, F. Analysis of Pulverized Coal by Laser-Induced Breakdown Spectroscopy. 

(18) El Haddad, J.; Canioni, L. ; Bousquet, B. Laser-Induced Breakdown Spectroscopy (LIBS), Part I: Review of Basic Diagnostics and Plasma–Particle Interactions: Still-Challenging Issues Within the Analytical Plasma Community. 

(19) Harmon, R. S.; Remus, J.; McMillan, N. j.; McManus, C.; Collins, L. et al. LIBS Analysis of Geomaterials: Geochemical Fingerprinting for the Rapid Analysis and Discrimination of Minerals. 

(20) Guo, G.; Niu, G.; Shi, Q.; Lin, Q.; Tian, D.; Duan, Y. Multi-Element Quantitative Analysis of Soils by Laser Induced Breakdown Spectroscopy (LIBS) Coupled with Univariate and Multivariate Regression Methods. 

(21) Farhadian, A. H.; Tehrani, M. K.; Keshavarz, M. H.; Darbani, S. M. R. Energetic Materials Identification by Laser-Induced Breakdown Spectroscopy Combined with Artificial Neural Network. 

(22) Galbács, G. A Critical Review of Recent Progress in Analytical Laser-Induced Breakdown Spectroscopy. 

(23) Miziolek, A. W.; Palleschi, V.; Schechter, I. 

 (Cambridge University Press, Cambridge, United Kingdom, 2006)

(24) Cremers, D. A., Multari, R. A.; Knight, A. K. 

Laser‐Induced Breakdown Spectroscopy, from Encyclopedia of Analytical Chemistry: Applications, Theory and Instrumentation

(26) Lasheras, R.; Bello-Gálavez, C.; Rodríguez-Celis, E. M.; Anzano, J. Discrimination of Organic Solid Materials by LIBS Using Methods of Correlation and Normalized Coordinates. 

(27) Gottfried, J.L.; Harmon, R.S.; De Lucia, Jr. F.C.; Miziolek, A.W. Multivariate Analysis of Laser-Induced Breakdown Spectroscopy Chemical Signatures for Geomaterial Classification. 

(28) Rezaei, F.; Karimi, P.; Tavassoli, S. Effect of Self-Absorption Correction on LIBS Measurements by Calibration Curve and Artificial Neural Network. 

(29) Koujelev, A; Sabsabi, M.; Motto-Ros, V.; Laville, S. Laser-Induced Breakdown Spectroscopy with Artificial Neural Network Processing for Material Identification. 

(30) Boueri, M.; Motto-Ros, V.; Lei, W-Q.; Zheng, L-J. et al. Identification of Polymer Materials Using Laser-Induced Breakdown Spectroscopy Combined with Artificial Neural Networks. 

(31) El Haddad, J.; Villot-Kadri, M.; Ismaël, A.; Gallou, G.; Michel, K.; Bruyère, D.; Laperche, V.; Canioni, L.; Bousquet, B. Artificial Neural Network for On-Site Quantitative Analysis of Soils Using Laser Induced Breakdown Spectroscopy, 

(32) Inakollu, P.; Philip, T.; Rai, A.K.; Yueh, F-Y.; Singh, J.P. A Comparative Study of Laser Induced Breakdown Spectroscopy Analysis for Element Concentrations in Aluminum Alloy Using Artificial Neural Networks and Calibration Methods. 

(33) He, Y. Application of Laser to Research of Gas Explosion Mechanism. 

(34) Galindo-Prieto, B.; Eriksson, L.; Trygg, J. Variable Influence on Projection (VIP) for OPLS Models and Its Applicability in Multivariate Time Series Analysis. 

(35) Peerbhay, K.Y.; Mutanga, O.; Ismail, R. Commercial Tree Species Discrimination using Airborne AISA Eagle Hyperspectral Imagery and Partial Least Squares Discriminant Analysis (PLS-DA) in KwaZulu–Natal, South Africa. 

Peer-Reviewed Research

An artificial neural network was combined with LIBS to provide a rapid and accurate coal-rock recognition method for unmanned coal mining.

Automatic Coal-Rock Recognition by Laser-Induced Breakdown Spectroscopy Combined with an Artificial Neural Network

The solution-cathode glow discharge (SCGD) is a low-power, atmospheric- pressure, ambient-atmosphere microplasma that is proving to be a proficient excitation source for atomic emission spectroscopy (AES). The analytical figures of merit of SCGD-AES experiments often compete with established, conventional approaches despite the fact that it is a simple, small, low-cost instrument. The operating principles of the SCGD are reviewed here, including experimental operating parameters, plasma conditions, analytical performance, matrix interferences, and application examples.

The solution-cathode glow discharge (SCGD) is a low-power, atmospheric-pressure, ambient-atmosphere microplasma that is proving to be a proficient excitation source for atomic emission spectroscopy (AES). The analytical figures of merit of SCGD-AES experiments often compete with established, conventional approaches despite the fact that it is a simple, small, low-cost instrument. The operating principles of the SCGD are reviewed here, including experimental operating parameters, plasma conditions, analytical performance, matrix interferences, and application examples.

A new class of miniaturized plasma sources that exploit the liquid–plasma interface of an atmospheric pressure glow discharge are gaining popularity as excitation sources for atomic emission spectroscopy (AES). Reports of glow discharge electrolysis date back to at least 1887, but it was not until 1993 that the first modern instrument for AES was developed by Cserfalvi and coworkers (1). The solution-cathode glow discharge (SCGD) experiment is much different from conventional plasma or flame-based AES strategies. In SCGD-AES, the plasma itself is responsible for sampling material directly from a liquid sample solution, as well as desolvating, vaporizing, atomizing, and exciting the constituent atoms for quantitation by emission spectroscopy. Since the introduction of SCGD, several decades of instrumental refinement and an improved fundamental understanding of the plasma has allowed it to achieve analytical performance comparable to traditional AES and atomic absorption approaches (2,3).

The potential advantages of using a miniaturized plasma, like the SCGD, in place of the argon inductively coupled plasma (ICP) or pre-mix flames were immediately recognized. The SCGD operates in the open ambient atmosphere and does not require compressed purified argon or flammable gases. The miniaturized plasma also requires only modest power (100 W) and uses simple electronic circuitry compared to the popular 1 kW radio-frequency ICP. Because the SCGD plasma directly samples the analytical solution, there is also no need for the nebulizers and spray chambers that are used with other methods. Moreover, because the liquid surface is constantly regenerated by a flowing stream, memory effects are nearly elimi- nated, and the plasma is naturally suited to coupling with chromatographic separations. The discharge predominantly produces neutral atomic emission lines, leading to a relatively uncluttered emission spectrum. Although the SCGD is certainly not a panacea to all limitations, these attributes make the SCGD attractive as a simple, low-cost AES excitation source with significant advantages best suited for applications like continuous, portable monitoring studies.

A schematic of a typical SCGD experiment is shown in Figure 1. The SCGD is an ambient atmosphere, atmospheric-pressure glow discharge formed between a metal pin-type anode electrode (tungsten, titanium, or molybdenum) and the liquid surface of a sample solution, which acts as the discharge cathode. The SCGD is sustained by a potential of 700–1000 V, requiring a power of approximately 100 W. The electrical circuit is completed by a graphite rod or a platinum wire placed within the liquid waste reservoir, and the circuit often includes a ballast resistor to prevent arc formation. A peristatic, or syringe, pump is used to introduce samples into the discharge capillary at a typical rate of 1–3 mL/min. As the liquid emerges from the capillary tip (0.4 mm i.d.), the liquid surface forms a cathode that is constantly replenished by the flowing stream. Excess liquid that is not sampled into the discharge overflows from the capillary tip into the waste reservoir.

Samples are typically prepared as acidic solutions (1% HNO

). Because the solution forms a part of the electrical circuit, maintaining constant solution conductivity is an important factor in ensuring a reproducible response. A variety of supporting electrolytes, detergents, and chemical additives have been reported as means of controlling solution conditions. Overall, most studies have shown that a low solution pH is the most important factor for optimum AES response in the SCGD, which makes acid digestion a particularly convenient sample pretreatment method. Atomic emission that originates in the SCGD is captured and analyzed in the typical manner, although many studies make use of portable fiber-optic-based spectrometers.

A large number of discharge geometries, solution conditions, and powering schemes have been reported, and several are critically compared in recent reviews (2–5). One popular modification of the SCGD is to simply reverse the applied voltage polarities, forming a solution-anode glow discharge (SAGD). The SAGD retains many of the benefits of the SCGD (low power, simple spectra, and no support gas required), but it has been found to provide enhanced sensitivity and limits of detection to select elements, including silver, bismuth, cadmium, mercury, indium, thallium, and zinc (6). The SAGD is sustained over the same current and voltage range as the SCGD and can use modestly acidified sample; thus, it can be sustained in the same instrumental arrangement as the SCGD.

A representative SCGD atomic emission spectrum is shown in Figure 2. The background emission from the SCGD plasma differs from argon-based plasmas such as the ICP. Significant broadband molecular background features exist in the range of 250–350 nm because of emission from NO, OH, and N

The SCGD ambient-atmosphere microplasma has emerged as an alternate excitation source for atomic emission spectroscopy that is able to perform admirably compared to established, conventional approaches—with lower cost.

February 2023

Measuring the Crystallinity of PHBHx with Varying Amounts of Sidechains on a Benchtop Instrument

Artificial Intelligence in Analytical Spectroscopy, Part I: Basic Concepts and Discussion

Automatic Coal-Rock Recognition by Laser-Induced Breakdown Spectroscopy Combined with an Artificial Neural Network

Solution-Based Glow Discharges for Atomic Emission Spectroscopy Come of Age