Computerized system validation (CSV) in good practice (GxP)-regulated laboratories is alleged to take ages and generate piles of paper to keep inspectors and auditors quiet. To the rescue is the draft FDA guidance on computer software assurance (CSA). But...is this really the case?

I wrote about the Food and Drug Administration (FDA)’s computer software assurance (CSA) in an earlier “Focus on Quality” column (1). I was not complimentary of the CSA, because it was “regulation” by presentation and publication, as the FDA had failed to issue the draft guidance, which I likened to waiting for Godot. Well, Godot has arrived! FDA issued the draft guidance on September 12th, with an effective date of September 13th (2). The Center for Devices and Radiological Health (CDRH) is the originator in consultation with Center for Drug Evaluation and Research (CDER), but only medical device good manufacturing practice (GMP) regulations are mentioned in the guidance.

It is good that the guidance is available so that claims of the more outspoken advocates can be verified. In this article, we explore if computerized system validation (CSV) is to be cast into the recycling bin of history with CSA ruling the roost, or whether CSA and CSV live together happily ever after, especially as the aim of both is that computerized systems are fit for their intended use. I discuss this guidance from the perspective of good practice (GXP)-regulated laboratories in the pharmaceutical industry.

In the interests of transparency, and to put my comments into perspective, let me outline my philosophy for implementing and validating laboratory computerized systems in the pharmaceutical industry: 

1. There is no point in implementing and validating a computerized system if you do not gain substantial business benefits, as noted by in Clause 5.6 of the 2016 World Health Organization (WHO) guidance (3).

 Otherwise, why would an organization bother wasting money on implementing a slow and error-prone hybrid approach? Deriving business benefits involves redesigning the business process, and using electronic signatures coupled with eliminating spreadsheet and calculator calculations and paper printouts. Interfacing the system with other informatics applications will leverage even greater process efficiencies. Yes, this takes more time, but the payback on the overall investment is immense.

2. Performed correctly, the first phase of a validation life cycle (selection) is investment protection and the second phase (implementation) is simply good software engineering practice. 

Documenting user requirements ensures you purchase and implement the right instrument and application (4,5). Documenting the configuration and settings provide you with the baseline for change control and ensures data integrity. It is also the ultimate backup in case of a major disaster, as you’ll be able to reinstall and reconfigure the application quickly.

3. Leverage a supplier’s software development to reduce user acceptance testing subject to an on-site or virtual assessment with a report (6).

 Many functions in laboratory systems are parameterized and not configured (including temperature and detector wavelength), and the business process does not change. If the supplier has developed and tested these functions adequately, why would you want to test them as noted by GAMP 5 (7,8)?

GAMP risk assessment is not the only methodology you can use (9). Furthermore, nobody can test any computerized system completely—not even a supplier. Users should focus on demonstrating that the system works as intended rather than finding software errors. See the big picture for the system: will it be connected to another laboratory informatics application in the future, or remain as a standalone system? Conversely, avoid adding “just in case” user requirements that you may never use. Finally, document your assumptions, exclusions, and limitations for your user acceptance testing (10).

5. Flexibility or scalability in your validation approach is essential, as one size does not fit all.

 You need to tailor your validation to the risk posed by each system and the records generated by it. As regulations, citations, guidance, and technology are changing, there are always better ways to perform the next validation. Maintaining a status quo CSV approach is not an option.

This philosophy above is based on process efficiency and investment protection, while regulations are only mentioned in point 5. If you approach CSV correctly, then business efficiencies pay for the system and the validation costs.

Before the release of the draft guidance, CSA was proposed as a better approach to demonstrating fitness for intended use by avoiding the generation of piles of paper to keep auditors and inspectors quiet. The aims were:

Focus on critical thinking rather than documentation and testing formalization;

Define the intended use of the software and determine the impact on product quality and patient safety;

Identify any functions within the software that impact product quality, patient safety, or data integrity, and determine the risk on them;

Leverage supplier development and documentation where possible; and

Define assurance activities through conducting scripted and unscripted testing activities based on risk.

Now that we have the draft guidance available, we can assess how these promises stack up with reality. Note that the maturity of an organization in interpreting intended use, risk assessment, and testing will impact if they streamline their current CSV approaches or not.

The title of the draft guidance is “Computer Software Assurance for Production and Quality System Software (2).“

The draft guidance consists of 25 pages. We find on page 4 that CSA is not a universal panacea, a change to the regulatory paradigm, an approach that turns the world upside down, or a revolution that replaces CSV—only that CSA is a supplement to an existing 20-year-old FDA guidance. What?! Despite of all the trumpeting of how superior CSA will be, the “General Principles of Software Validation (GPSV) (11),” issued in 2002, will be the master, and CSA only a slave. When finalized, the CSA guidance (11) simply replaces three pages of Section 6 with about 20 pages (see Figure 1). Green represents the sections of the GPSV that will remain effective, red represents the three sub-sections of that will be replaced by the CSA guidance, and yellow shows how Section 6 will look when the CSA guidance is finalized. This column will not consider the examples in Appendix A, as there is no direct impact on spectroscopic systems. Five points immediately arise:

1. Far from being a total replacement for CSV, CSV and CSA will coexist. However, the current Section 6 mentions both system validation and software validation—so what term is CSA replacing? CSV is required for medical device software validation under 21 

 11 (13). CSA will only be focused on lower risk production and automation applications. In February 2022, FDA proposed updating 21 

 820 to harmonize with ISO 13485-2016 (14). It will be interesting to see if the CSA guidance is compatible with the new regulation.

2. CSA does not cover the whole gamut of software covered by CDRH, but only the lower risk software involved in production and quality systems. However, production systems are hardly mentioned in this guidance.

3. If the CSA guidance is supposed to be a supplement to the GPSV and replace section 6, why not combine the two into a single guidance? The obvious advantage is that there would be a single guidance to read rather than two, one of which has a section that has been replaced but without notifying readers. Ozun Okutan has provided a detailed rebuttal of CSA in a blog on LinkedIn. There are existing medical device software standards that are used outside the US so why not use them (15)?

4. The focus is on software—a rather amorphous term. Does this mean application, a macro, or just “software?”

5. The FDA doesn’t make it clear whether sections 3–5 of the GPSV do or no not impact section 6, so it’s a dog’s dinner of a situation.

When I read the guidance for the first time, this was a shock, as some CSA advocates were talking about a paradigm shift. I was expecting a new approach; instead, the wheels appear to come off the CSA bandwagon, and the guidance is a half-baked anticlimax. Still, let’s soldier on, and see what other gems await us.

One confusing element of the CSA guidance is the terminology used. “Computerized system validation” includes elements of the software, computer hardware, IT infrastructure, any connected instrument, and the trained staff. Conversely, without any definition, “computer software assurance” focuses on software alone:

Computer software assurance is a risk-based approach for establishing and maintaining confidence that software is fit for its intended use.

Computer software assurance establishes and maintains that the software used in production or the quality system is in a state of control throughout its lifecycle (2).

Fine words, but meaningless. At least the definition of validation in the GPSV provides understanding (11).

The main body of the draft guidance (section V) is split into four main sections under the heading of CSA Risk Framework, shown in Figure 2.

The first stage is to define the software’s high level intended use, which is not about individual functions within the software. This is essential, because software that is used directly or indirectly as part of production or quality system must be validated under 21 

 820.70(i) (12), whereas the guidance does not apply for software that does not. Direct software is stated to be:

 Automating production processes, inspection, testing, or collection and processing of production data.

 Development tools that test or monitor software system or automated software test tools.

 Automating quality system processes, collection and processing of quality system data, or maintaining a quality record under 21 

Software usually not requiring validation is listed as general business management and supporting infrastructure, such as networking. For the latter, there is a potential clash if applied in pharma as the Principle of EU GMP Annex 11 requires IT infrastructure to be qualified (16).

As the General Principles guidance is still effective, don’t forget the sentence in Section 5.2.2: 

It is not possible to validate software without predetermined and documented software requirements

 (11). Your user requirements specification (URS), that I interpret as “intended use,” is the means to ensure the right system for the right job and protect your investment.

As CSV uses URS, and CSA uses intended use, there is a potential for conflict and misunderstanding. It would help greatly if the CSA guidance provided advice on how to document intended use. Is definition of intended use a user requirements specification, or a replacement for it? It could also include the use of a software engineering tool to manage requirements electronically.

Are these two terms interchangeable or different? It depends. If you are the 

, the two terms have different meanings. “Intended use” is applied to analytical instruments, but “intended purpose” refers to both analytical procedure and software, as they focus on the overall process. This is emerging during the revision of 

In the guidance, the FDA offers a binary choice here between “high process risk” and “not high process risk”—no fence sitting here—and identifies high process risk features as functions or operation automated by software. The five areas identified in the guidance are shown in Figure 3. Item 2 is the one area where laboratory systems may be involved—Process Analytical Technology (PAT), where spectroscopic systems integrated with process systems test drug content in tablets.

Laboratory-centric examples provided in this section to illustrate the risk-based approach are three enterprise resource planning (ERP) scenarios, a manufacturing execution system (MES), and an automated graphical user interface.

Determining the Appropriate Assurance Activities

Simply put, for high process risk functions, the higher the risk, the more validation, sorry assurance activities are required, and vice versa. Mentioned earlier in the guidance are the requirements to ensure that security, configuration settings, data storage, data transfer, and error trapping features must be set up and documented—interestingly, nothing about data integrity.

A major problem with the CSA guidance is that the word “testing” is mentioned 145 times, versus 106 times for quality. There is a potential danger of falling into the trap of trying to test quality into software rather than ensure requirements are properly defined.

One of the key CSA benefits advocated was streamlined test documentation. There are five types of assurance activity (as shown in Figure 4), which are mostly based on the general concepts of software testing in part 1 of 

Dynamic testing in which the tester’s actions are not prescribed by written instructions in a test case, and is outlined in 

 (18); it would be good if the FDA referenced the 2022 version to be current. The guidance sub-divides testing, as shown in Figure 4, into:

Ad-hoc testing unscripted practice that does not rely on large amounts of documentation;

Error-guessing testing based on a tester’s knowledge of past failures; and

Exploratory testing spontaneously designs and executes tests based on the tester’s relevant knowledge and experience of the test item.

As I noted in my earlier article, prototyping is performed without documentation (1), as well as evaluating new software in a development environment.

Dynamic testing in which the tester’s actions are prescribed by written instructions in a test case (18), which are divided into:

 A hybrid of scripted and unscripted testing scaled according to the risk passed by the system; and

 Scripted testing that includes links to traceability to requirements and auditability

These last two scripted testing approaches are not in the 

 29119-1 (18), but have been devised by FDA.

A problem with the guidance is extensively quoting a software and systems engineering standard 

 29119, where testers will have a fuller picture of software and may be testing the same modules or application repeatedly or in depth. What problems can you see from the testing options above?

1. To perform error-guessing and exploratory unscripted testing, you need a knowledge of the application you are testing. If it is new system and you have just received training, how much do you know of past failures or experience of the test item? Zero? Then these two test approaches have just bitten the dust.

 Annex 11 clause 4.4 states, “User requirements should be traceable throughout the life cycle (16).“ Therefore, can only robust scripted testing be applied in a regulated GMP pharmaceutical laboratory? So where is the much-vaunted improvement and benefit in testing claimed by the CSA community? There appears to be none in a pharmaceutical context.

Does this mean that a coach, horses, and a safari of migrating wildebeest have been driven though a key benefit of CSA? The risk is that there is an apparent disconnect between the guidance’s approach to streamlining testing and a user’s knowledge of the application coupled with pharma GMP regulatory requirements for network qualification and requirements traceability (16). Another problem is with the ISO standard quoted for the different testing approaches.

Q: When is a Consensus Standard Not a Consensus Standard?

FDA use some of the definitions for types of software testing from this ISO standard. Developing a consensus is important for ISO standards; however, there is a problem with ISO 29119-1 (18). The standard has been criticized as too simplistic, and does not represent a consensus for software testing (19,20). Wikipedia has a summary of the contro- versy, which has led at least two testing organizations to call for the standard to be withdrawn (21). Some of the reasons given for the withdrawal were it did not represent a consensus, it profited by selling standards, it was deemed too prescriptive, there was no need for testing standardization, and—wait for it—there was a heavy focus on documentation (21)! The draft CSA guidance appears to select some testing methodologies (and ignore others) without reason.

In Table I of the CSA guidance, we can see the level of documentation required. It is important to note that all types of assurance activities require documentation of the work performed, but obviously the extent and detail varies. Its good to see that the old saying “If its not written down, it’s a rumor” survives at the FDA.

 Requires intended use and risk determination to be documented, a summary description of the testing performed, any issues found and resolved, who performed the testing and when, and review and approval (as appropriate!), plus a conclusion statement.

 Requires intended use and risk determination to be documented, as well as the test objectives, expected results, pass/fail criteria, detailed report of testing, any issues found and resolved, who performed the testing and a conclusion statement, and also calls for review by a second person pre- and post-execution of the protocol.

An example of exploratory testing of a spreadsheet for collecting and presenting nonconformance data is discussed. A spreadsheet? You cannot be serious! Readers will know of my high regard for spreadsheets used in a regulated environment (22,23), so I’m surprised to see a spreadsheet example in an FDA guidance document. This raises serious issues of data integrity and governance, such as how to document and test requirements to show that the DI controls work. Perhaps this was a lack of critical thinking?

Completing this section is the following sentence:

As a least burdensome method, FDA recommends the use of electronic records, such as system logs, audit trails, and other data generated by the software, as opposed to paper documentation and screenshots, in establishing the record associated with the assurance activities (2).

This is best part of the whole guidance! I have no problem with this sentence, as I have used an application to self-document user acceptance testing for years (6). Why do more?

In this section, I’ll look at what was promised versus what was delivered, where have we seen this before, and summarize some of the inconsistencies that we’ve seen in the guidance.

Now let us compare the CSA approach promoted in presentations and publications before in the draft guidance was issued, as shown in Table I.

Let us look at the missing approaches in more detail:

 Proposed as a major component that was lacking in CSV. Is it mentioned directly in the CSA guidance? No. Missing in action.

 Even unscripted testing requires documentation as noted in the draft guidance

However, the impression given in some CSA presentations was that unscripted testing equated to undocumented testing.

Is it mentioned directly in the CSA guidance? No. Missing in action.

 discusses leveraging supplier involvement in sections 2.1.5, 7 and 8.3, plus appendix M2 for supplier assessments, as noted in my earlier publication on CSA (1). 21 

 11.10(a) on validation could also imply using suppliers for accuracy and reliability of electronic records (13).

Why supplier involvement was omitted in the guidance is probably due to the emphasis on 

 29119-1. The main emphasis of the FDA draft guidance appears to be on developing and testing software, and not on effective implementation of purchased applications.

 Fine, in principle, but it takes time to learn and use the automation tool, as well as validate and assure the software. Expect the first system to take longer to validate. As many lab systems are standalone how will the automation tool cope with this?

Articles

Column: Focus on Quality

Now that the draft FDA guidance on computer software assurance (CSA) is out, we can see what it is—and is not.

CSA: Much Ado About Nothing?

Fiber optics have undoubtedly changed the world of spectroscopy and communications, existing as a foundation for countless platforms we encounter and work with each day. The use of multiple fibers in a bundle further expands these capabilities in several ways, some of which are obvious, and others that are more concealed and clever. Starting as a Central Intelligence Agency (CIA) encoding project under the guise of “gastroscope development,” fiber-optic bundles have evolved into many uses across industries. In this column, we explore this intriguing history and where it has led to, including the use of modern processing techniques that allow three-dimensional (3D) interpretations of traditionally two-dimensional (2D) readings.

When it comes to candy, would you rather have a single big chocolate bar, or a pack of M&M’s? Perhaps you may prefer a box of assorted chocolates, or you are really eccentric and prefer Skittles or some fruit-flavored sweet. Whatever your preference, any industry that exists long enough develops various form-factor and flavor options, and the world of fiber optics is no different. Over time, folks have moved away from the traditional chocolate bars and into more diverse offerings, and this article aims to touch on one such genre of fiber optics known as 

Those without much direct experience with bundles may wonder how much there is to say about putting fibers next to each other, and those who work closely with bundle systems may wonder how a single article could capture any meaningful and complete point on the topic. Indeed, it is an area with a deep and intriguing history that has led it down multiple paths serving countless applications. In this column, we take a glance at some key highlights of the past, and some of the coolest implementations in the present and near future.

But first, back to the chocolate—who doesn’t love a classic chocolate bar? It has everything you love in a single, homogeneous, bulk form factor, and it satisfies chocoholic cravings at the most basic level. Standard fiber is very similar, giving you the ability to route light in a single, homogeneous, bulk form factor in a way that satisfies many applications just fine. Also, folks have been working with glass for a while, or at least since 2500 BC in Egypt and Mesopotamia, with Egyptians already carving glass heads by 1500 BC (1). Although proper fiber optics would come much later, even glass blowers back in the time of Jesus must have seen their glass rods guide light when putting them into a glowing furnace and looking at the ends.

But two things had to happen before glass fiber could become a common reality: One was a demonstration by British physicist John Tyndall in 1854 showing that light could travel through a curved stream of water (2), and the other was the Industrial Revolution, which made glass commonplace for windows and bottles (1). With the concept of single-core glass fiber established in the 1800s, but with no spectrometer to connect to, these fibers existed largely as a novelty during the Victorian Era, providing elegant decorations and demonstrations. But the eggheads were still tinkering, as they do, and they could see the flickers of potential these new “light pipes” had to offer.

Fast forward to the 1900s, specifically Nazi Germany, and we find a fellow named Heinrich Lamm, who assembled an awkward and primitive bundle of glass fiber, and he used this to show how the ‘V’ image of a light bulb filament could be carried around a bend (1). Not terribly exciting today, but extremely revolutionary at the time! Unfortunately, Lamm’s Jewish background forced him to flee Germany and his work, which then sat idle for over a decade. A general timeline of some of these key events is shown in Figure 1, along with the gradual improvement of glass fiber transmission as the years went on.

The end of World War II brought the beginning of the Cold War, and now information technologies were just as attractive as projectiles and explosives. Brian O’Brien had a clever idea for a fiber-bundle scrambler, which was immediately valuable to the CIA because the spy agency could use one half of the bundle to make ciphers that only the spy with the matching half could decode; the spy, in turn, could send messages back that only the spy masters could decode. Each pair of encoders would have a unique pattern, so loss of one scrambler would not break security for those using others. The CIA pushed this project forward under the guise of “gastroscopes,” which, indeed, had their own value for non-espionage purposes, but also had dramatically differ- ent requirements. Fibers in a gastroscope to be inserted down the esophagus need to be long, thin, and flexible, while fibers in a spy-device should be short, fat, and monolithic. Realizing this, a hungry MIT professor named Frederick Norton stole a clever idea from a Swiss roll cake. To be honest, there’s actually no proof of his relying on cake for the idea, but there’s also no proof against it, and the concept is quite similar. He would stack glass rods together, and then heat them to fuse and be twisted and stretched as desired; these could then be cut like a roll cake to create unique optical ciphers. Rather than cake, the technique he was more likely leaning on was the ancient craft of “millefiori” or “thousand flowers,” which is still used today in art and small objects like paperweights (1). Look up some examples, they are quite beautiful!

But as math usually ruins everyone’s day, it was determined that a mere 18-or-so intercepted images would be enough to crack the scrambling code, and the CIA eventually killed the cipher project to refocus (in honesty this time) on the gastroscopes they had said they were working on. The medical application evolved down its own path, but optical fiber bundles arguably had their first “real job” back in the 1960s reading the punchcards used in computers at the time. Early card readers used an array of 12 small and short-lived bulbs, with one to illuminate each row in a moving card. A fellow named Bob Potter at American Optical replaced this setup with a fiber bundle that collected light from one long-lived bulb and split the light into 12 smaller bundles, one for each hole. American Optical also developed 9–15-in bundles for NASA testing of the Saturn rocket boosters during the Apollo program, so these were finding some real-world usage (1).

We’ve gotten to the point in time where fiber bundles are a real thing, and people are figuring out they can be used to solve a wide range of problems. But let’s take a break from the history lesson, at least for the moment, and let’s talk about some of the key benefits fiber bundles offer us today in the world of spectroscopy and photonics:

“Mapped” bundles can be aligned to carry a coherent image.

Patterned bundles can sync with some type of complex receiver.

Randomized bundles do more than scramble images for the CIA, and can be useful to uniformly distribute multiple sources or perform some other type of photonic averaging or amplification.

Bundles allow you to mix and match various fiber types to create a diverse “box of chocolates.”

Assemblies may include single- or multi-mode fiber, different numerical apertures, diameters, claddings, tapering, and any other aspect you would want to optimize. For example, a bundle may be created with a 50:50 randomized mix of UV-visible (UV-vis) and visible-near-infrared (vis-NIR) grade fiber, creating an extended broadband range with the durability and flexibility benefits described below.

This is arguably the greatest benefit, as you can deploy bundle assemblies in very harsh and rugged environments with minimal risk of breakage or optical distortion. We will talk about this in greater detail since it can be such a blessing.

We’ve briefly touched on the concept of coherent bundles when looking at the history of Lamm’s work, and there is more to say about that shortly. But patterned bundles, which are meant more for data-based connectivity rather than image transmission, are a fulcrum in the world of communications. Like candy buttons where the right color is placed in the proper spot with its others, patterned bundles ensure signals are grouped as they should be. While Dorset, England is probably best known for its Jurassic Coast, its police force in 1975 was the first non-experimental entity to use a fiber-optic telecommunication link, and by the mid-1980s, Sprint had established the world’s first 100% fiber optic network (2). These fiber links require assemblies with consistent patterns and layouts, and the first telecommunications ribbon cables used 240 μm fiber in linear arrays of 6–12 held by mylar tape. A typical telephone line would have 12 ribbons of 12 fibers each, stacked on top of each other and surrounding a sheath embedded with strength supports in the center; a submarine version of this would essentially be the same thing with two layers of outer armor added (3). Today’s data servers use bespoke fiber-bundle patterns and connectors to relay the right information to the right place.

We all love and live for our communication technologies, but most readers are here for the real treat—that is, the theory behind it all. Thinking about the most basic multifiber assembly presents us with just two fibers side by side, and such multicore fibers can be split into two groups: Assemblies where fibers are designed for a controlled coupling of the guided modes between the cores, and assemblies where modes in the different cores are practically decoupled (4). Coupled assemblies may be similar to a KitKat bar, where the wafers are independent, yet connected, versus a Twix bar where each side is entirely decoupled; we will leave the mathematical proof behind the superiority of the left-side Twix for another article.

You can do some pretty cool things by arranging cores the right way. In a common twin-core fiber with two identical single-mode cores within the cladding, illuminating one core will equally excite two transversal modes in the fiber, one symmetrical and one asymmetrical. Energy is periodically transferred back and forth between cores, where the first core will be totally depleted of energy after some length, but this is now guided by the second core and will start to transfer back to the original core; this continues ad infinitum as light travels down the fiber. You can imagine how this would be useful in amplifiers and lasers, and the beat length, which is the fiber length necessary to complete one cycle of this, can be adjusted with core separation, geometry, refractive index, and wavelength (4).

Perhaps the greatest benefit of fiber bundles comes from their inherent resistance to harsh environments and rough handling. There are actually two areas to touch on here: mechanical durability and optical stability. Starting with the former, have you ever had a chocolate bar in a backpack or other tight space, only to take it out and find it broken into several pieces? Note this does not happen with a pack of M&M’s due to the small size of the individual components. Similarly, fiber assemblies using single, large fibers are prone to breakage from overbending or mechanical shock, whereas fiber bundles using many small-diameter fibers can undergo extreme bends and hits without compromising the internal fibers...or at least saving the vast majority of the signal. Theory states that fiber bending will cause some high-order modes to hit the core-cladding interface at less than the critical angle and become lossy, while low-order modes should be unaffected (5). More complete predictions, however, state all modes will lose power at a bend—the wavefront propagation in a fiber will be compelled to move faster at the outer periphery of the bend than at the center of the fiber, and since it would be necessary to exceed the speed of light in the material to accomplish this feat, some loss must occur. Because all modes are represented in the wavefront, all lose power at a bend (5). Therefore, it is difficult to achieve requirements of high bandwidth and low sensitivity to bending with single large fiber, but bundles using small core diameters solve this because the bend ratio to the fiber radius is larger for the same absolute bending radius (4). Keep this in mind because it will come up shortly when we loop back to gastroscopes. Figure 2 provides a sketch of this metaphor between the form factors of chocolate and optical fiber.

Techniques such as two-step cladding can further improve bandwidth and bend immunity, and graded-index profiles, where fibers have a continually diminishing refractive index from the fiber axis to the cladding, can increase bandwidth by two to three orders of magnitude, when done right (4). The benefit to that technique is how it keeps the beams in the fiber center, but that’s another topic for another day.

A popular configuration for fiber used in common optical oxygen sensor systems is randomized multimode bundles, and an old demonstration we often showed in prepandemic trade show days was literally tying the fiber between the probe and electronics into a knot and then hitting the knot with a large rubber mallet. The demonstration showed how stable the optical signal remained despite dramatic bending and physical impact. Such resilience makes bundles instantly attractive for handheld probes, vehicle systems, industrial sensors, and other applications with high vibration and movement. If you’ve ever had a fiber-based absorbance setup such as using a standard 1-cm cuvette, you may have noticed your broadband spectrum became “wobbly” or “wavy” any time you touched the connecting patch fiber or even bumped the table. This is due to the mode-mixing effect seen in typical single-core, multimode fiber. By moving to a bundle, you will see this effect becomes muted or even disappears entirely because all of the mode-mixing “noise” from the individual fibers is now averaged across potentially hundreds of fibers in the bundle. Handheld dip probes, for example, see great improvement in optical stability and repeatability when using bundles to transmit light, at least in the UV-vis range. Going back to the chocolate bar versus M&M analogy, fiber bundles let the chocolate melt in your mouth and not in your hand...or rather, they let the optical signal make it to its intended destination, rather than degrade along the way. So, in terms of both external durability and internal photonic quality, bundles are essentially “the M&Ms of fiber optics.”

Let’s loop back to Lamm’s work on coherent image transfer back in the 1930s, but this time with a modern twist. The most important use of coherent bundles is the endoscope, which saves thousands of lives each year by seeing problems that would otherwise remain undetected. A standard endoscope can have over 10,000 fibers, which creates a 1-mm spot capable of 70 μm resolution (6). This is great for seeing and transmitting 2D images, but a group at the ARC Centre of Excellence for Nanoscale BioPhotonics at RMIT University in Melbourne has developed a clever method to extract 3D information from such images.

As you can imagine, the pixelation of an image coming out of a bundle is a function of how many fiber cores make up that image, and in normal operation this image is somewhat homogenized or “downsampled” to remove that pixelation effect, basically adding a slight blur to make it less choppy (7). This ends up eliminating any modal information that may be contained in the cores. What Orth and associates (7) have shown is that the intracore intensity structure of the output light carries information about the angular distribution of the input light, which can then be post-processed to map the original depth of the object; this is demonstrated in Figure 3 with a single fluorescent point source as the input, and also showing a mockup depth map of multiple point sources at various distances from the input facet.

Several paragraphs earlier, we were talking about high- and low-order modes experiencing loss at a bend, but how the small diameters used in bundles helps to minimize this. The technique proposed by Orth and colleagues is possible because the higher-order modes, which are preferentially excited at larger angles of incidence, carry more energy near the core-cladding interface than the low-order modes, such that light is effectively pushed to the edge of each core with increasing ray angle. They then use a process called digital aperture filtering to remove peripheral light for each core, creating a synthetically constricted numerical aperture. Using light field moment imaging (LMI), they can extract missing azimuthal information to calculate the average ray direction at a given point in the image. Orth’s paper obviously goes into much greater detail, and is a recommended read, but the key takeaway is that this now adds depth mapping to a traditional 2D image. This group used a bundle of 30,000 fibers with 2 μm diameter, and was able to achieve better than 10 μm depth accuracy with samples up to 80 μm away from the fiberface.

What would Heinrich Lamm think of this modern evolution of his inspiring first steps? What an amazing technological leap in less than a century to be roughly translating a large, distorted, 2D “V” shape into then generating 3D depth maps with 10 μm accuracy! We’ve tracked these human-led developments in the world of fiber bundles, and it’s time to bring our discussion to a close, but it’s also worth noting a fiber-bundle design developed by the world’s most prevalent inventor: nature. Remember the millefiori we talked about earlier, with the fused glass rods cut like a Swiss roll cake? There is a mineral called ulexite found mainly in Boron, California, that’s created through evaporation of mineral-rich lakes, and this naturally creates the same type of structure seen in the millefiori. These are sold in toy and rock shops as “TV stones,” because they translate the image from one face coherently to the other. Try to find one next time you’re in a novelty or rock shop, and maybe it will inspire your kids to have the next brilliant idea in fiber optic design and write the next chapter of the story.

(2) “History of Fiber Optics,” Timbercon. 

https://www.timbercon.com/resources/blog/history-of-fiber-optics/ 

(3) Christian, N.; Passauer, L. “Fiber Optic Component Design, Fabrication, Testing, Operation, Reliability, and Maintainability,” Noyes Data Corporation, 1989.

An Introduction to Fiber Optics System Design

(7) Orth, A.; Ploschner, M.; Wilson, E. R.; Maksymov, I. S.; Gibson, B. C. Optical Fiber Bundles: Ultra-Slim Light Field Imaging Probes. 

Columns: Lasers and Optics Interface

A brief history of fiber-optic bundles helps us understand how this technology how it has evolved from use in telecommunications to enabling extraction of three-dimensional (3D) information from two-dimensional (2D) images.

Fiber-Optic Bundles Create a World of Pure Imagination

Although tracing product origin by the analysis of their multi-element composition has been successfully applied in several commodities, the potential for doing the same with timber has not yet been determined. Laura Boeschoten and her team studied the reactive soil element pools and the multi-elemental composition in sapwood and heartwood for 37 Azobé (

) trees at two forest sites in Cameroon, measuring 46 elements using inductively coupled plasma mass spectrometry (ICP-MS), and conducted the first study relating wood multi-elemental composition of forest trees to soil chemical and physical properties in an effort to combat the illegal trade of timber. Laura spoke to 

 about this study and the technique’s potential.

Your paper (1) deals with using forensic methods, specifically ICP-MS, to independently trace timber origin in order to combat the illegal timber trade. Briefly state why you chose this technique over others.

ICP-MS has shown a lot of potential for origin identification of commodities—for example, for wine and cabbage. One of the pros of the method is that you can measure a large number of elements simultaneously, thus providing a lot of information easily. That is why we thought it would be interesting to try this for timber as well.

How can multi-elemental composition provide a chemical fingerprint for timber tracing?

Trees have a different multi-elemental composition in different regions. We are looking for the differences between locations, to find something like the fingerprint or barcode of a specific site. Based on that, you can compare the chemical composition wood of unknown origin to a reference dataset. This tells you which location your sample is most similar to, which may help trace location when the origin is unknown or when it needs to be verified.

Please describe details on how you applied random forest analysis for determination of timber origin.

We used a classification model to test whether we could correctly assign our trees to their respective origin. We chose a random forest model for this, as it is a robust machine learning method that compares many different decision trees and then gives you the best classification model for your dataset. The random forest model can thus predict the most likely origin of wood based on the chemical composition.

Please describe your discovery and development process, and how this differed from what was previously done by yourself or others.

We were the first to show the potential of multivariate analysis for tracing timber origin. The idea was developed some years ago when our team was researching other methods for the same purpose, and they saw that the method was applied in other commodities. We then tested for variation in timber, and tried to understand what drives differences in wood chemistry between different origins based on soil composition to be able to apply the method in timber.

We found that wood of two locations could be allocated to the correct origin with great success, and we found that wood chemistry was associated mostly with soil clay content and soil organic matter.

What are the implications of your research, and how do you see them combatting the illegal timber trade, both short term and long term?

We hope to test this method on a larger scale and in other species, to further understand how it works, and where it has the most potential. Then it will be a great addition to the timber tracing toolbox; together with other methods, we will be able to independently verify timber origin. This is an essential step in combatting illegal timber trade, as, currently, the origin is only verified by documents and external tags, both of which are prone to falsification.

Were there any particular challenges you encountered in your work?

Most of the challenges for this project were related to fieldwork. Of course, we needed a lot of wood and soil to test this method, so we had all the challenges related to setting up a field campaign. You need all the right people and material in the right place at the right time, which was challenging sometimes. And then Covid started, so we outsourced most of our fieldwork to local partners. This strengthened our collaborations, but also made it harder to understand what was going on in the field.

What sort of feedback did you receive regarding this work?

We received a lot of positive feedback on our project: forestry companies, competent authorities, non-profit organizations, and the Dutch customs laboratory were even involved from the start, and are very enthusiastic about the progress.

What are the next steps in this research?

As explained earlier, the next step will be to test this method on a larger scale with more species. This will improve our understanding of the full potential of the method.

(1) Boeschoten, L. E.; Sass-Klaassen, U.; Vlam, M.; Comans, R. N. J.; Koopmans, G. F.; Rocha Venâncio Meyer-Sand, B. et al. Clay and Soil Organic Matter Drive Wood Multi-Elemental Composition of a Tropical Tree Species: Implications for Timber Tracing. 

Columns

ICP-MS was used to measure 46 different elements in 37 Azobé (Lophira alata) trees at two forest sites in Cameroon to determine the multi-elemental composition of trees in an effort to combat the illegal trade of timber.

Clay and Soil Organic Matter Drive Wood Multi-Elemental Composition of a Tropical Tree Species: Implications for Timber Tracing

Raman spectroscopy can characterize size-related properties of semi- conductor nanomaterials according to the change of Raman shift. When limited to physical mechanisms, it is often difficult to predict the size-dependent Raman shift of semiconductor nanomaterials. To predict the size-dependent Raman shift more accurately and efficiently, a simple and effective method was created, demonstrated, and achieved via the deep learning model. The deep learning model is implemented by multi-layer perceptron. For size-dependent Raman shifts of three common semiconductor nanomaterials (InP, Si, CeO

), the prediction error was 1.47%, 1.18%, and 0.58%, respectively. The research has practical value in material characterization and related engineering applications, where physical mechanisms are not the focus and building predictive models quickly is key.

Raman spectroscopy can characterize size-related properties of semiconductor nanomaterials according to the change of Raman shift. When limited to physical mechanisms, it is often difficult to predict the size-dependent Raman shift of semiconductor nanomaterials. To predict the size-dependent Raman shift more accurately and efficiently, a simple and effective method was created, demonstrated, and achieved via the deep learning model. The deep learning model is implemented by multi-layer perceptron. For size-dependent Raman shifts of three common semiconductor nanomaterials (InP, Si, CeO

Semiconductor nanomaterials play a major role in many research areas, including energy conversion (1), sensing (2), electronics (3), photonics (4), and biomedicine (5). Size effect is widely accepted as an inevitable consideration when controlling special properties of semiconductor nanomaterials for different applications of interest. Because Raman shifts show obvious changes with the decrease of the material size, Raman spectroscopy, as a powerful tool, can effectively characterize the size of semiconductor nanomaterials. Predicting the size-dependent Raman shift is in favor of the research on size-related properties of semiconductor nanomaterials. A great deal of theoretical research has attempted to determine the size dependence of the Raman shift in semiconductor nanomaterials, such as phonon modes (6–8), thermodynamic methods (9,10), and atomic coordination models (11). Because of the limitation of complex physical mechanisms, establishing a theoretical model is time-consuming work. It is easier and simpler to obtain bond number model (12,13), but coordination numbers of surface atoms in this model need to be revised frequently for different semiconductor nanomaterials, which is a trivial task. However, convenience and speed are the most urgent needs for engineering. As a result, from the application point of view, a simpler and more effective method is needed for predicting the size-dependent Raman shift of semiconductor nanomaterials.

Deep learning is one of the most focused research fields in recent years, aiming to learn features and directly build predictive models from large-scale raw data sets. Deep learning has been widely researched on qualitative and quantitative analysis of Raman spectroscopy (14–16). Deep learning is a data-driven method and thus can perform direct feature extraction of experimental data without considering complex physical mechanisms. Moreover, deep learning possesses the powerful ability of the non-linear feature extraction. As a result, deep learning is very suitable for the regression analysis of complex physical problems. Once a well-trained deep learning model is achieved, the size-dependent Raman shift of semiconductor nanomaterials can be easily predicted.

In this article, to predict the size-dependent Raman shift of semiconductor nanomaterials, a deep learning model was achieved by multi-layer perceptron. It is discussed that the influence of the training data set fluctuation and the type of activation function on the performance of the deep learning model. Finally, according to the test on three common semiconductor nanomaterials (InP, Si, and CeO

), the results show that the size-dependent Raman shift can be predicted via the well-trained deep learning model.

Data sets of three common semiconductor nanomaterials (InP, Si, and CeO

) were generated by the bond number model that is used to simulate the experimental data (that is, the size-dependent Raman shift of semiconductor nanomaterials). To be in agreement with the experiment, the random noise was added into the bond number model to simulate the fluctuation of Raman shift in the experiment (this is called the training data set fluctuation in this article). The fluctuation amplitude indicates how far the measured value deviates from the true value because of noise. The random noise (

) that generates random values from 0 to 1 with uniform distribution. The random function is implemented by Python library function. The fluctuation amplitude (

. The truth value refers to the situation where the fluctuation amplitude is 0 cm

. The bond number model (13), random noise, and data set are as follows:

 is the size of semiconductor nanomaterials with the shape of cuboctahedron; 

 is the size-dependent Raman shift of semiconductor nanomaterials with random noise that is training data; 

 is the atom number on one edge of the cuboctahedron; and 

Finally, the training data set is generated by varying 

 from 2 to 100, and the size of the data set is 99. As a result, the size of the semiconductor nanomaterial is discrete. Material parameters of three semiconductor nanomaterials are given in Table I.

Size-dependent Raman spectra (InP, Si, and CeO

) have been reported in the literature (17–19). However, to visually observe the data set generated by the bond number model, Raman spectra simulated by the Lorentzian function are given in Figure 1. Because the relationship between the Raman shift and size is reflected by the Raman peak position of the longitudinal optical (LO) mode, the broadening of spectral bands and other spectral bands are not considered in the simulated spectrum. Figure 1 shows the change of Raman shift when n (equation 2) varies from 2 to 11. The definition of Lorentzian function and the relevant parameters are given in the “Supporting Information” section.

The deep learning model is achieved by multi-layer perceptron. The schematic of multi-layer perceptron is shown in Figure 2, including one input layer, two hidden layers (H1 and H2), and one output layer. Each hidden layer is followed by one batch normalization (BN) layer. The sizes of input layer and output layer are both one. The multi-layer perceptron can be written out as (20):

The tanh is selected as activation function (

) to extract nonlinear feature and conduct nonlinear regression analysis. The impact of different activation functions on the performance of the deep learning model is compared in the next section. The definition of tanh is as followed (20):

Experimental Configuration and Hyper Parameters

The deep learning model is trained on dual CPUs (Intel Xeon 6226R) with a memory size of 128 Gb. The multi-layer perceptron is achieved via the PyTorch that is an open source framework. Stochastic gradient descent is selected as the loss function from the PyTorch library. Hyper parameters including the number of neurons in hidden layers (H1 or H2), the learning rate, the batch size and the epoch are determined by grid searching method. The final values are given in Table II.

To evaluate the effect of training and the performance of the trained model, the training error and the prediction error are calculated correspondingly. The 

 is the deviation of the predicted value from the truth value for the training data set, and the 

 is the deviation of the predicted value from the truth value for the testing data set. The training error indicates the training effect of the model, and the prediction error indicates the generalization ability of the trained model.

First, the influence of the training data set fluctuation is discussed. In the experiment, the experimental value of the Raman shift generally deviates from the truth value because of the resolution of the Raman spectrometer. Generally, the resolution of a high-precision Raman spectrometer can be less than 0.2 cm

, such as Horiba T64000, whereas the resolution of a common portable Raman spectrometer can be less than 4 cm

. Therefore, the fluctuation amplitude of Raman shift from 0 cm

 was chosen to study the influence of the training data set fluctuation on the training of deep learning model. Here, InP data sets with different fluctuation am- plitudes are used, and the training data set is generated by equation 4. According to Figure 3, in the training process, the loss value is markedly decreasing with the training epoch. To show it more clearly, loss values of the last five epochs are zoomed into view and shown in the inset of Figure 3. As shown in Figure 3, when the fluctuation amplitude of the training data set varies from 0 cm

, the loss value at the 2600th training epoch increases from 0.07 to 7.25 correspondingly. Although the fitting of the multi-layer perceptron to the training data is degraded under the larger fluctuation amplitude of the training data set, the loss value can converge steadily.

To observe the influence of the training data set fluctuation on the training result further, the training error given in equation 9 is calculated under different fluctuation amplitudes of training data sets. The training errors are shown in Figure 4. Compared with the fluctuation amplitude of 0 cm

, the training error increases with the increase of the fluctuation amplitude of the training data set. When there is no the training data set fluctuation (that is, the training data is equal to the truth value), the training error is 0.63%. For the fluctuation amplitude of 10 cm

, the training error is 2.01%. Although the training data set fluctuation will cause the degraded performance of the deep learning model, the training accuracy is still satisfying. To illustrate extreme case, all errors in the whole article are expressed with the maximum of error. In addition to the training error, the prediction error given in equation 10 is also calculated. For the ideal training data (the fluctuation amplitude of the training data set is 0 cm

), the training error and the prediction error are actually the same, so the prediction error is also 0.63%. For the fluctuation amplitude of 10 cm

, the prediction error is 3.46%, which further indicates that the precise training data is a key factor to determine the accuracy of the deep learning model. The relevant results are shown in Table III.

Next, the role of activation function in the model training is compared. In the above section, the activation function of tanh is used. For comparison, another non-linear activation function sigmoid is used in this article. Optimal hyper parameters for the sigmoid are re-optimized. Optimal hyper parameters for sigmoid and corresponding results are given in Table S1, Table S2, Figure S1, and Figure S2 ("Supporting Information" section). For the activation function of sigmoid, the loss value at convergence, the training error and the prediction error are worse than those of tanh. The loss value at convergence varies from 0.42 to 7.92 when the fluctuation amplitude of the training data set varies from 0 cm

. The training error and the prediction error are both 1.61% for the ideal training data. When the fluctuation amplitude of the training data set is 10 cm

, the training error is 2.35% and the prediction error is 4.95%. Therefore, for the nonlinear regression analysis, the activation function is also important in determining the accuracy of the deep learning model. However, for semiconductor nanomaterials meeting the bond number model, the activation function of tanh is better.

To further demonstrate the feasibility, the fitting to the size-dependent Raman shifts of Si and CeO

 is studied as well. All hyper parameters and network structure of the deep learning model are the same as those of InP. Nevertheless, the deep learning model is retrained by the training data set of Si and CeO

. Results in Figure 5 are obtained under the fluctuation amplitude of the training data set of 2 cm

 are 1.18% and 0.58%, respectively. Because the resolution of a high-precision Raman spectrometer can be less than 0.2 cm

, the prediction error has the potential to be reduced further.

A deep learning model for predicting the size-dependent Raman shift of semiconductor nanomaterials was demonstrated and achieved via multi- layer perceptron. The prediction error of the size-dependent Raman shifts of three semiconductor nanomaterials (InP, Si, CeO

) was 1.47%, 1.18%, and 0.58%, respectively. Through the research of this article, predicting the size-dependent Raman shift of semiconductor nanomaterials can be greatly simplified, which is very conducive to engineering applications.

This study was supported by National Natural Science Foundation of China (61905047), and Fundamental Research Funds for the Central Universities of China (3072021CF2510).

(1) Nehra, M.; Dilbaghi, N.; Marrazza, G.; Kaushik, A.; Abolhassani, R.; Mishra, Y. K.; Kim, K. H.; Kumar, S. 1D Semiconductor Nanowires for Energy Conversion, Harvesting and Storage Applications. 

A deep learning model for predicting the size-dependent Raman shift of semiconductor nanomaterials was demonstrated and achieved via multi-layer perceptron.

April 2023

CSA: Much Ado About Nothing?

Fiber-Optic Bundles Create a World of Pure Imagination

Clay and Soil Organic Matter Drive Wood Multi-Elemental Composition of a Tropical Tree Species: Implications for Timber Tracing

Prediction of the Size-Dependent Raman Shift of Semiconductor Nanomaterials via Deep Learning

Study on Estimating Total Nitrogen Content in Sugar Beet Leaves Under Drip Irrigation Based on Vis-NIR Hyperspectral Data and Chlorophyll Content