K.H. Esbensen and V.

Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 115

The sampling interface—A critical Theory of Sampling success factor in process

sampling and PAT

K.H. Esbensena and V. Sivalingamb

a
KHE Consulting, Copenhagen, Denmark. E-mail: [email protected]
b
Dept. of Process, Technology, University of South-Eastern Norway, Porsgrunn, Norway

In the process industry arena, analytical endeavors are today well served by the Process Analytical
Technology (PAT) framework, offering a plentitude of on-line analytics, mostly spectroscopic: UV-VIS,
NIR, RAMAN, NMR, ‘acoustic chemometrics’ a.o. This context is generally considered satisfactory, but
there is a major catch. The present paper focuses on a fundamental feature in process sampling, the
link between an on-line analytical technique and a moving stream of matter, which is to be character-
ised, monitored, and analysed. A process sampling operation can serve many objectives; process
regulation, product or interim product characterisation, optimisation of raw material consumption,
maximizing energy efficiency, pollution, and environmental impact management – or process, plant,
company, corporation decision making at supervisor and management levels concerning economics
and risk management. All these objectives can only be meaningfully undertaken if based on relevant
information, that is, if based on reliable analytical data, which in turn is 100% dependent upon repre-
sentative samples, or representative sensor signals. The singular common element in all of this is the
process sampling interface. This paper develops a critique of most current interfaces, with a scathing
verdict: the dual role of the process sampling interface as both delineating an appropriate aliquot vol-
ume while facilitating the specific analytical technique, has largely been overlooked, with a significant
negative effect that is spelled out in full detail across implemented solutions across widely different
application sectors. Most current process sampling interfaces do not comply with TOS’ demands, put-
ting representativity severely at risk for heterogenous materials.

Introduction – setting the scene

The Theory of Sampling (TOS) has always been front and center concerning how to extract representative samples from
moving lots (process sampling), e.g. from a conveyor belt or from ducted material streams. TOS’ preeminence is that iden-
tical Governing Principles (GP), Sampling Unit Operations (SUO) and Sampling Error Management (SEM) can be applied
identically to both stationary and to moving lots. The 70-year application history in TOS’ traditional realms is highly successful
regarding physical extraction of increments and samples, especially in the Mining, Minerals and Metals processing and
refining industry (M3), in cement production and in many other process industry sectors dealing with broken ores, particulate
matter, minerals and slurries. A comprehensive, up-to-date introduction to TOS in these sectors can be found in three recent
TOS textbooks, and in the wealth of references to the extensive literature found herein1-3. These are conveniently presented
in a comparative overview, easily available4.

Recently in these industry sectors, a pronounced interest can be observed for applying modern sensor technologies
for on-line monitoring because of the obvious advantages for fast, real-time information used for process monitoring and
control allowing optimized decision making, all with significantly added economic results. These approaches take a broader
view on available analytical techniques that can be applied on-line, always closely related to the characteristics of the ma-
terials and lots involved (physical, chemical, other). A key example outside TOS’ traditional realm is the use of Near InfraRed
(NIR) spectroscopy in the broad pharmaceutical manufacturing context (‘pharma’ for short). In terms of societal importance
and total turnaround economy, one can meaningfully equate M3 and pharma.

In the process industry arena today, analytical endeavors are well served by the Process Analytical Technology (PAT)
framework, offering a plentitude of on-line mostly spectroscopic analytics: UV-VIS, NIR, RAMAN, NMR, ‘acoustic chemo-
metrics’ a.o. A comprehensive introduction to this field can be found in the “PAT bible (2010)”5.

The present paper focuses on the fundamental feature in common in process sampling, the link between an on-line
analytical technique and the moving stream of matter, which is to be characterized, monitored, and analysed. A process
sampling operation can serve many objectives, process regulation, product or interim product characterization, optimization
of raw material consumption, maximizing energy efficiency, pollution, and environmental impact management – or process,
plant, company, corporation decision making at supervisor and management levels concerning economics and risk man-
agement. All these objectives can only be meaningfully undertaken if based on relevant information, that is if based on
reliable analytical data, which in turn is 100% dependent upon representative samples, or representative sensor signals
from the start.

doi: https://doi.org/10.1255/tosf.141 Published under a Creative Commons BY-NC-ND Licence

116 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

The singular common element in all of this is – the process sampling interface.

Below it is shown that this commonality is the critical success factor w.r.t. representative sampling sensu TOS. To reveal
the critical role of the process sampling interface, examples shall be presented from TOS’ traditional realm (not many are
needed taking into account the WCSB10 context), but deliberately also from what at first sight may appear as very different
industrial sectors, pharma and similar industries (food/ feed processing, production and manufacturing for example), but
which are never-the-less dealing with the same matter, i.e. heterogeneous aggregate mixtures, only often from a finer grain-
size realm (powders, slurries). From these industrial sectors focus shall be on NIR spectroscopic applications, which are all
in the realm of PAT.

While at first sight the M3 and pharma/food/feed (PFF) industry sectors are likely to be found disparate in the extreme in
their process technology contexts and w.r.t. the wildly different materials involved, it turns out they are closely related in their
identical need for representative process sampling to conduct reliable process monitoring and process control, and QC/QA
of interim/finished product. In fact, for PFF the need for appropriate sampling Quality Management (QM) is identical to those
for M3. This means that M3 and PFF industries are fully aligned and joined precisely at the juncture which is the focus for
the present paper – the process sampling interface. The common role of the process sampling interface is to bring a specific
on-line analytical technique in appropriate contact with the material/product/streaming material that is to be characterized.
The role of the process sampling interface is to facilitate the analytic interaction with the moving material – without giving
rise to sampling errors to any degree more than absolute necessary. This turns out to be a formidable challenge.

In this paper, the analytical specifics in themselves are not of interest (see the abundant PAT literature). Below it is all,
and only about what comes before analysis. Also, the analytical competence per se is very rarely an issue; it is almost
universally well taken care of in the form of thoroughly validated analytical procedures.

So, are the M3, PFF, and all similar particulate matter realms fully competent regarding the specific process sampling
issues? After all, solutions to this task can look at the unparallel success for historical TOS applications over ~70 years in
science, technology, industry, trade, society and in the environmental sciences. For more than five decades the Theory of
Sampling (TOS) is rightly famous as the framework that covers all necessary elements, i.e, design of sampling procedures
and equipment, correct implementation, correct practical operation, sufficient maintenance, a necessary and sufficient the-
oretical competence framework.

Well, it turns out there are still aspects of TOS application that can benefit from a general perspective - and none more
so than the process sampling interface.

TOS vs. PAT – an exemplar contrast from which to learn

Within TOS’ traditional realms it is well known how to sample particulate matter and mixtures, both from stationary lots and
from moving streams of matter. When it comes to how to extract physical increments and samples in a representative
fashion, the foundation of TOS reigns supreme, and all new applications always take their point of departure from this
framework and the from the wealth of accumulated practical experience, as witnessed by the historical library of Proceedings
from the World Conferences on Sampling and Blending WCSB1-WCSB10. Sampling by physical extraction is very well
served indeed, see6-9 for example.

By contrast, for ‘sensor sampling’ i.e., Process Analytical Technologies (PAT), which is claimed to be a ‘no sampling
needed’ approach, there is no similar theoretical framework for the sampling step, which is rather often treated in a somewhat
stepmotherly fashion. Instead, a pronounced practical approach rules in this realm5, in which the question of how to achieve
representative sensor signals is the first item on any agenda, here intimately related to the design and implementation of
the specific analytical instrumentation, which is to be put in operation by way of a sensor interface connecting to, and inter-
acting with, the streaming flux of matter.

This endeavor is closely related to the specific analytical technique involved (the use of NIR is dominant, but other ana-
lytical techniques find their use as well in specific cases). This is always based on specific calibration approaches; calibration
is always needed as PAT instrumentation is multi-channel, multi-wavelength – producing multivariate spectral data. There
can be observed a drive: “get the multivariate spectral signals – and all will be well”, which is based on the overwhelming
success of the data analytical discipline chemometrics in the form of multivariate calibration of process sensor signals from
multi-channel analytical instruments10. One often gets the distinct impression that the specifics of a PAT sensor interface is
more driven by the needs and requirements of the sensor analytics than of considerations for the analytical support (vol-
ume/mass) that can be realized. This is where the issue of the effective increment volume crops up.
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 117
Preluding what is more fully illustrated in systematic graphics below, PAT solutions make use of three principal types of
interfaces, three sampling modi, which are very well known in TOS. Pat interfaces are almost always of the type (1,2)
instead of (3):

1. Taking part of the stream, all of the time; or

2. Taking part of the stream, part of the time; instead of
3. Taking full cross-steam increments part of the time (the only TOS-compliant sampling approach)

These sampling modi have been illustrated with particularly clear graphics in a companion contribution in these proceed-
ings, see Figs. 2-4 in Novaes & Hidding (2022)11.

With modus (3), focus rightly is on sampling the entire stream, full slices of the stream, while sampling in process time.
While with (1,2), the full width/breadth/depth of the stream is never covered in full, leading to TOS-incorrect sampling,
massive IDE, thereby breaking one the two necessary-and-sufficient requirements for representative process sampling
(unbiasedness and sufficient precision).

Based on such a failing PAT focus, the issue of representativity here is mostly, or sometimes wholly, only related to
spectral and reference sample measurement uncertainty (MU) – but data modeling errors and uncertainties are also
acknowledged if the chemometrics is not up to the job (data model under – or over fitting). With this PAT focus, the realm
‘before sensor signal acquisition’ is often unwittingly ignored, even though this is the very domain where sampling errors
abound. This issue constitutes the central theme of this presentation: In the PAT realm sampling errors are incurred in the
exact same fashion as when extracting physical samples from moving streams. In the PAT realm this is largely a surprising
insight – indeed this parallel is also able to shed light on how process sampling sometimes is subpar, even in the TOS
arena.

For these reasons, sampling error effects impacting on the quality of sensor and probe signals largely constitute a terra
incognito in the PAT realm. This situation invites all parties to take a serious look at the need to comply with TOS’ stipulations
for representativity, no matter with what interface configuration sampling is performed. It turns out that there is a complete
duality between TOS and PAT, which must be acknowledged and followed where- and whenever the goal is representative
sensor signals, see chapter three in5, and also10,11. This duality is the central theme for this paper.

Be aware that a dear friend goes under many names. Thus, what in chemometrics is termed ‘multivariate’ (multi-variate,
meaning many variables), in many other scientific and technological fields has acquired the name ‘hyper-spectral’ a.o. A
good example of the latter, and a good example of the emerging awareness of the spectral (multivariate, hyper-spectral)
advantage by frontline players, would be Mittrup et al (2017)12, which is also among the pioneers applying NIR spectroscopy
in M3.

The contradiction between TOS and PAT has been highly frustrating, indeed a situation that has lasted for the better
part of the last 10-15 years. The scope can best be illustrated by the fact that the first edition of the famed source book:
“Process Analytical Technologies (PAT)” (2007) did not contain a specific chapter on sampling (sic). While the second
edition, out just three years later (2010)5, features a comprehensive chapter, aptly named: “Process sampling: Theory of
Sampling (TOS) – the missing link in Process Analytical Technologies (PAT)” – the title says it all!

Focusing
In many industry sectors, there is today an emerging rush to take advantage of ‘modern sensor technology’ (including
‘advanced’ statistics, machine learning and other artificial intelligence approaches), to provide real-time process information
to support and optimise business decision making. Although this trend manifests itself slightly differently in different process
industry sectors, typically because the sector viewpoint is intimately related to one dominating specific analytical approach,
e.g., XRF, PGNAA, NIR, it is rare to find awareness of the overarching hidden commonality involved, the role of process
sampling interface. Focusing on this issue makes it possible to review and critique a multitude of current applications to
avoid making the same mistakes over and over in many disparate realms in science, technology, industry, trade, society,
and the environment.

In medias res - PAT

Process monitoring and control in technology and industry is incomplete without full understanding of all sources of variation,
causing bias and impression. It is particularly important to be able to decouple sampling and analytical variations caused
by bias-generating errors because these can, and shall be reduced maximally, indeed preferentially eliminated. This is the
first, well-known mandate of TOS: elimination of sampling bias. By contrast, the analytical bias is fundamentally different,
not always clearly acknowledged as is a major point in1. The analytical bias is treated very well within the specific analytical
realm and literature.

A case in point: It is not enough to make use of Multivariate Statistical Process Control (MSPC) or similar QCQAQM
monitoring approaches based on charted analytical results, because these may well be compromised by unknown signifi-
cant sampling errors, if not adequately identified, quantified, and reduced to below a relevant a priori acceptance threshold:
This situation results in unnecessarily large bias and impression.
118 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

More precisely: Process data are affected both by analytical measurement errors as well as extraction sampling errors
or sensor signal acquisition errors. This understanding is of critical importance, because the latter two categories of sam-
pling error effects are all too often unrecognized, or unknown. But worse, they typically dominate over analytical errors by
factors 5-10-20+ the more heterogeneous materials targeted and/or if proper sampling competence is not brought to bear
in the design, implementation, maintenance, and operation of the total process measurement system. Alarmingly, almost
always the Total Analytical Errors (TAE) constitute only a very small, sometimes vanishingly small proportion of the total
Measurement Uncertainty (MU) – What makes “all the difference” quantitatively and qualitatively, is the Total Sampling
Error (TSE). Below we are only interested in how to minimize TSE in process sampling.

In the PAT and in the process technology context in general, it is not sufficiently known that sensor signal acquisition
gives rise to identical error types as does physical sample extraction.

Below we bring forth the critical analogy between Process Analytical Technologies (PAT) and conventional physical
sample extraction and show how three QCQAQM tools, the Replication Experiment (RE), interleaved sampling (IS), and
variographic process characterisation (VC) can be involved. RE, IS, VC constitute a necessary and sufficient on-line ap-
proach for total variability decomposition. Without proper sampling error management (error identification, reduction and
elimination), on-line analytical instrument calibration and validation will unavoidably incorporate significantly inflated data
uncertainties, the bias complement of which is inconstant and without correction possibilities, all of which assuredly com-
promises the ultimate process monitoring and control objectives significantly.

The presentation below reflects the close interaction between TOS, PAT and chemometrics, a competence overlapping
that is not much in focus in the IPGSA realm. Here, space consideration does not allow more than referral to a select suite
of key, comprehensive literature sources for interested parties who wish to go deeper with respect to fundamental theory
with all three scientific fields11,5-10.

A practical starting platform for the present objective shall be a fundamental sampling duality.

The fundamental duality between physical and sensor sampling

Below all efforts have been made to make illustrations completely anonymous; no identification of persons, locations, in-
strument brand or companies is intended. Only the didactic value of correct vs. incorrect, faulty, or fatal non-compliance
with TOS is of interest.

Fig. 1. The fundamental sampling duality between physical increment/sample extraction (left) and sensor signal
acquisition (right). Installing PAT sensors has often been claimed to be equivalent to ‘no sampling needed’ –
alas! Illustration copyright KHE Consulting teaching collection, used with permission.

Installing PAT sensors has often been claimed to be equivalent to ‘no sampling needed’ – alas, this is a complete fallacy.
The issues treated below are not novelties for the sampling community – rather they should be trivial but are not entirely.
But the main thrust of this paper is directed at related scientific communities, such as chemometrics, process technology,
process engineering, processing, and manufacturing plant design – all facing the same issues. For these communities, it
has been a rude awakening how the “Eureka moment” in the right panel of Fig. 1. dissolved into a whimper as soon as
proper Total Measurement Uncertainty and full system performance validation was invoked. Alas, TSE error effects also
abound in the PAT sensor world. “Why - there is literally no physical sampling involved?”
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 119
First lesson: Both physical extraction sampling errors and sensor technology spectral acquisition errors are in play – in
fact they are identical, as even rudimentary TOS-analysis will show.

The sampling interface

Figure 2 shows the key insight presented in this paper: TOS’ mandatory rules that guarantee representative increment
extraction from a moving lot (physical process sampling) are identical when applied to secure a representative sensor signal
from a ducted material stream (sensor or PAT process sampling). There is an often-overlooked subtlety involved here,
which is the distinction between a sample container, holder, cup, vial – and a sampling interface, see further below.

Fig. 2. Key commonality between physical increment/sample volume and sensor signal support in process sam-
pling: Identical rules apply w.r.t. correct increment delineation and extraction (complete and undisturbed volumet-
ric coverage) for both physical extraction as well as for delineation of the sensor signal support volume as facili-
tated by a sampling interface. This illustration can either be viewed as looking down on a heterogeneous 1-D
moving material lot f.ex. on a conveyor belt - or it can be understood as a vertical section through a horisontal
conduit transporting a heterogeneous material (indicated by irregular texture). From both points of view, TOS-
correct process sampling necessitates recovery of, or spectral coverage of, the full delineated cross-stream incre-
ment volume, shown by the three renditions in the upper panel Only fully covering increment volumes with planar-
parallel boundaries will result in representative increments or samples. Illustration copyright KHE Consulting
teaching collection, used with permission.

Note the non-representative increment geometries indicated in Fig. 2. The upper panel shows fatal ‘grab sampling’,
either as partial 3-D volumes or as superficial grabbed increments from the top of conveyed material only. The central panel
illustrates increment delineation/extraction f.ex. resulting from a valve opening capturing a “partial fraction of the moving
stream of matter all the time” only (as well exemplified by the infamous Shark Fin valve), while the bottom panel shows
flawed cross-stream increment delineations, flawed as they furnish unbalanced increments w.r.t. transverse stream heter-
ogeneity. None of these configurations corresponds to the imperative demand for an uncompromised full slice of the moving
stream, none of these deliver TOS-compliant increments.

Thus, Fig.2 shows the non-negotiable demands for correct vs. incorrect increment delineation, which determines whether
a sampling process will give rise to Increment Delineation Error (IDE) effects, or not. By implication, the illustration also
shows the demands for elimination of Increment Extraction Errors (IEE). These demands on a practical physical sampling
system, for example from a conveyor belt, applies with equal criticality for sensor sampling, but contrary to what is often
assumed, there is not a dichotomous divide between physical – and sensor sampling. Fig. 3 shows the most often occurring
principal process sampling options of either modality; it is noteworthy that all the shown examples from current process
technology do not deliver representativity1-5.
120 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

Fig.3. Overview of typical, often occurring sampling system configurations and locations encountered in process
technology across diverse industry sectors. “Sampling valves” are designated A through I, while typical locations
of sensor technology solutions (PAT) are marked as red arrows. None of the process system solutions shown here
are representative however, see5. Illustration copyright KHE Consulting teaching collection, used with permission.

Whether designed for physical increment/sample extraction by use of “sample valves” [A through I] from vessels and
containers, horisontal or vertical pipelines – or designed to take advantage of the much claimed “no sampling necessary”
sensor technology (PAT), none of the sampling systems depicted in Fig. 3 will give rise to representative samples or signals,
which is easily verified by reference to the ‘full slice imperative’ shown in Fig. 2.

Figures 2 & 3 and the remarks above, lend full justification to the duality argument illustrated in Fig 1.

There is another duality presented below. Illustrative examples of the main message in this paper shall come, appar-
ently willy-nilly, from both the pharma PAT/NIR world and from the traditional TOS process realm addressing M3 for ex-
ample. This is deliberate, as it is meant to bring forth the universality of the discussion of the process sampling interface
that follows.

But first a few reflections on sensor spectroscopic penetration depth, which are of key interest.

Spectral penetration depth also matters

NIR analysis is influenced by how lots, samples, and smaller sub-samples are handled before the aliquot is presented to
the analyser. An iconic author in the NIR field (P. Williams, 2001) listed ~30 sample-associated factors that affect the
accuracy and precision of NIR analysis, but only gave indicative actions for how to avoid some of these and emphasized
that: “If [these] are not efficiently carried out, consistent and accurate NIR analysis is not possible, no matter how sophisti-
cated the software.”, (Williams quoted in13). We might here add: “… and regardless of the analytical competence”. This view
is in complete agreement with the principles presented by TOS.

Thus, with respect to analytical errors and their effect (NIR and for almost any other analytical modality), in a specific
sense it is all about what comes before analysis – because this is where the quality, relevance and validity of the subsequent
analytical results originate! Disregard for this realm is like the original sin …

One specific analytical issue can also be seen as impacting the sampling realm, which is of paramount importance for
sensor process sampling. This concerns the effective spectroscopic penetration depth. We shall use NIR as an exemplar
for many other spectroscopic techniques, although their individual penetration depths will vary, and can be widely different.
However, the principal argument below is universal: penetration depth is a bona fide sampling issue, penetration depth
uncertainty matters!
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 121

Fig. 4. Schematics for NIR spectroscopy/spectrometry, illustrating the thorny issue of varying penetration depth.
Illustration copyright, Rodolfo Romanach teaching collection, used with permission. See also discussion in13

The analytical setup in Fig. 4 illustrates the business end of sensor sampling of a powder mixture, based on the interac-
tion of the incident radiation with the material, either in an extracted sample contained in a sample holder (full cylinder
height) or in an assumed, correctly delineated volume serving the same purpose as the sample container. This volume is
delineated by the sensor sampling interface.

The key issues is that this interaction is simultaneously a sampling operation (sensor sampling), as well as the physical
basis for quantitative analysis through a multivariate calibration model, allowing prediction of a specific material property
(often a concentration) (Esbensen and Swarbrick 2018)10, Romañach and Esbensen 2016)13. However, this type of chemo-
metric prediction of analytical concentration will be fraught with unnecessarily inflated prediction errors (prediction vari-
ances), with varying material-dependent impacts, due an obvious volume mismatch between the assumed aliquot volume
(full cylinder height) and the physically realized support volume for the spectral signals, which is only a fraction hereof.

In this case, sensor sampling is intricately related to the physics of diffuse reflection in NIR spectroscopy. The volume
mismatch issue is in general material dependent. As the top layers of the sample will receive more radiation than the bottom
layers, less radiation reaches the bottom part of a material on a volume-to-volume comparison. For a given material with a
specific physical constitution and different grain-size distribution a.o., the depth penetration also depends on the wavelength
used. The lower wavelength radiation penetrates to greater depths at where molecules absorb less NIR radiations. Thus,
the effective sample mass interacting with NIR radiation also varies according with wavelength. A more comprehensive
introductory description of these issues can be found in13 and in “Sampling” column in Spectroscopy Europe/World SE-2
(2021)14, entitled “What are Sampling Errors – and WHAT can be done about them”, which has this ingress:

“The objective of this column is to provide easy-to-understand examples of sampling errors. Prompted by
recent participations and presentations at on-line conferences and meetings, we believe there is a need
for a more fulfilling introduction and exemplification of the concept and real-world consequences of com-
mitting “sampling errors”. WHAT is a sampling error? WHAT is the result of sampling errors? WHAT can
we do about sampling errors? […] The point of departure will be in the Theory of Sampling (TOS) and in
the near infrared spectroscopy analysis and pharma application sectors, but the focus will be developed to
be more general, so that readers can carry-over to other scientific and application areas of interest.”
The vast professional NIR literature is readily available (a plethora of references can be found in the literature referred to
above).

The key issue is that the effective volumes are not identical – which according to TOS’ framework will unavoidably lead
to a bias.

It is sometimes argued that since this mismatch error is ‘identical’ for all samples (calibration, validation, prediction sam-
ples), it will in effect create a ‘constant bias’, which will influence calibration, validation, and prediction samples identically,
that can therefore be corrected for (as can an analytical bias). However, this is a fatal misunderstanding because this
understanding tacitly presupposed that all materials, mixtures, and powders are identical when involved in NIR analysis, all
differences in sample preparation, handling, pouring, compactification, and presentation a.o. notwithstanding. But this is
clearly never the case for the multitude of very different materials subjected to NIR analysis in this world.

These relationships cannot be resolved universally by the traditional route of demanding a “bigger sample”. The solution
is hoped to be NIR radiation interacting with an increased analytical mass in a well-designed fashion to accomplish practical
composite sampling and at the same time guarantee sampling correctness (elimination of sampling bias). This objective
has been the holy grail for decades of development of suitable ‘sample cells’ within the NIR realm, far from all of which
having reached this goal, however. While the specifical analytical problems of this kind are well recognised, it takes more
122 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

than a minimum of TOS competence to look through the immediate manifestations of many ‘sample interfaces’, salient
examples appear below.

This kind of sensor sampling error cannot be corrected for by any physical pre- or post-treatment, any mathematical
algorithm or by any software approach. These issues will always affect adversely on the possibility of developing fully
optimised calibration models. All NIR spectroscopic methods are similar to a multiple path-length sample cell where particles
may interact with radiation more times than once. The mass analysed by the radiation may be estimated in some cases but
is generally not known in detail. Most other analytical chemists know exactly the material that is analysed, including its
mass. The authoritative “HANDBOOK of NIR Analysis (2020)”, chapter 19, characterises this fundamental sensor sampling
enigma in the following way (quote): “NIR spectroscopists are spectacularly the most successful analytical chemists that do
not know the exact sample mass being analysed!” Factor in the appropriate density of the analysed material, then mass
=~volume in this scathing dictum.

For multivariate calibration truly to be the chemometric be-all, end-all solution to process sampling, it is imperative that
‘an increment – is an increment’ - in the sense that ‘a process sampling interface increment volume – is simultaneously also
the physical reference sample volume of matching size’. This is obviously not something that can be taken for granted
across the gamut of analytical techniques in the realm of PAT.

The process sampling interface is a harsh mistress at this most fundamental scale. It pays to pay attention, also because
other issues occur at the scale of the process sampling interface itself.

Spectral averaging is not composite sampling

Development of PAT systems to meet TOS’ Fundamental Sampling Principle (FSP) is an ongoing challenge. Full compli-
ance with FSP requires that an entire cross section of a process stream shall be obtainable for it to be a bona fide repre-
sentative increment, Fig. 2. This requirement is probably the most often occurring sin-by-omission, or sin-by-neglect, in
many current PAT and in similar process technological applications.
As but one example, the Field-of-View (FOV) of a beam of NIR irradiation only interacts with, say, the top 1-2 mm or so
of a moving bed of matter. For a process stream the depth of which may be many, many times greater; there is clearly a
massive Increment Delimitation Error (IDE) at work here. This situation must rightly be considered beam grab sampling,
incurring a significant support volume mismatch. The parallel with physical grab sampling is striking – the duality is com-
plete.
Today’s dominating modus operandum is pledging allegiance to a massive averaging of spectral scans acquired during
the process flow. While this in some, restricted way appears to constitute an improved procedure, the penetrative question
is, how much better is a number of grab samples, scanned or physically obtained grabs, if none of them complies with the
necessary principles for increment representativity? In the informed optics of TOS, this ‘way out’ for PAT analysts will
unfortunately not necessarily lead to representative averaged signals even though apparently using spectral composite
sampling - because it is still only the uppermost part of the streaming matter which is ‘averaged’.

A representative composite sample (physical or optical) requires that all constituent parts of the lot moving through the
sampling interface, have the opportunity to become part of the composite sample. This can manifestly not be obtained by
any variation of increments which only represent the top of a moving stream, or any similar series of partial increment
volumes only, Fig. 2. This is a structural impossibility. Spectral averaging is not by itself TOS-compliant composite sampling,
thereby affecting TSE – but spectral averaging does help to get the analytical uncertainty better constrained, thereby af-
fecting TAE (only). This all takes place under the complete boundary condition of the Total Measurement System Error
TMSE := TSE + TAE.

From PAT/NIR sensor sampling to physical sample extraction

What is the relationship between the pharma/NIR/PAT powder and powder mixture worlds briefly described above – and
the traditional TOS realm of sampling of aggregated particulate materials of geological, biological a.o. origin over a very
wide range of grain-sizes, say from decimeters to micrometers? Despite all the drastic intrinsic differences, from the point
of view of process sampling it is all the same - only the scale changes.

Key distinction
Perusing the relevant literature, one often comes across a misunderstanding concerning the distinction between a “sample
cell” and a “sampling cell”. The difference is very much not just a semantic difference – it constitutes the essential insight
for full understanding of the role of the process sampling interface.

A “sample cell” (noun) (think of a sample holder, a sample cup, a sample vial, a ‘petri dish’ … …) presupposes that
‘the sample’ has been sampled before being put into this container. All the thorny issues regarding sampling have been left
in the analytical ante-room, the realm ‘before analysis’. For good order it is noted that this realm is typically comprised by
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 123

an integral series of sampling stages: primary -, secondary, tertiary sampling ... the last of which produces the analytical
sample. The primary characteristic of a sample cell is that it is passive; it ‘merely’ is filled by the already-sampled analytical
aliquot. “Someone else is responsible for the sampling” – heard this before?

A “sampling cell” (noun) is an active devise that serves a dual purpose (think of a ‘flow through’ cell). A sampling cell
is designed to delineate the operative volume support (sensor increment) from which spectroscopic signals is obtained,
thus effectively determining the analytical sample mass - while simultaneously facilitating the spectroscopic interaction with
the moving material (sensor field-of-view, analytical duct window opening a.o.). The former function has everything to do
with TOS’ concepts of IDE and IEE. A broader view is needed than the basic truism: all sampling cells enclose a specific
volume, determined by the design of the sampling interface. Here is the most important distinction:

The dynamic analytical support volume delineated by the process sampling interface – must be equal to the volume of
a ‘full slice’ increment, in the sense of Fig. 2. Indeed, physically it must be the same volume. No exceptions are acceptable,
if representativity is the objective.

Graphics – a picture tell a thousand words

The following figures are presented as principal sketches and generic renditions; they only serve the purpose of illustrat-
ing the characteristics of an installed process sampling interface. No identification of instrument brand, OEM, plant, com-
pany, corporation, or persons is intended. On the contrary every effort has been made only to present schematic didactic
information pertaining to how to perform proper process sampling.

Figures 5-13 below all focus on the role of the sampling interface. The examples shown cover a wide and dominating
range of contemporary process technological implementations and PAT installations.

It is highly significant that nearly all the illustrations are telling the story of how not to perform process sampling! The
reason is sobering – there are not many practical process sampling equipment and system solutions found on the market
today, or in contemporary industrial installations, that factually comply with TOS’ requirements for representative process
sampling. Unless the gamut of all the world’s materials correspond to infinitely diluted chemical solutions or to perfectly
mixed composite materials. Not many do (see also the compositional caveat below). For all of these, the following will be
a dramatic eye-opener,

Fig. 5. Principal sketch of process sampling interfaces for analytical techniques purporting to deliver sensor sig-
nals that are representative of the streaming matter in conveyed flows, ducted pipes or from reactors. The flow
intersecting volumes, depicted in red in cross-section views of the flow, do not comply with the imperative demand
for a full cross-sectional slice, cfr. Fig. 2. Illustration copyright KHE Consulting teaching collection, used with
permission.
124 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

Fig. 6. Principal sketch of process sampling interfaces for analytical techniques purporting to deliver sensor sig-
nals that are representative of the streaming matter in conveyed flows, ducted pipes or from reactors. The flow
intersecting volumes, depicted in red in cross-section views of the flow, do not comply with the imperative de-
mand for a full cross-sectional slice, cfr. Fig. 2. Illustration copyright KHE Consulting teaching collection, used
with permission.

Fig. 7. Principal sketch of process sampling interfaces for analytical techniques purporting to deliver sensor sig-
nals that are representative of the streaming matter in conveyed flows, ducted pipes - or from reactors. The
flow-intersecting volumes, depicted in red in cross-section views, do not correspond to the imperative demand
for a full cross-sectional slice, cfr. Fig. 2. Illustration copyright KHE Consulting teaching collection, used with
permission.
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 125

Fig. 8. Principal sketch of process sampling interfaces for analytical techniques purporting to deliver sensor sig-
nals that are representative of the lot material residing in storage containers and similar. The realized analytical
support volumes, depicted in red in cross-section views, do not correspond to the imperative demand for a full
cross-sectional slice, cfr. Fig. 2. A series of containers or drums, for example from a production line or forming
an incoming raw product line, also constitute a process sampling situation, whether sampled by an installed sen-
sor (left) or a handheld device (right). Illustration copyright KHE Consulting teaching collection, used with permis-
sion

Fig. 9. Principal sketch of process sampling interfaces purporting to deliver physical increments that are repre-
sentative of the streaming matter in conveyed flows, ducted pipes or from reactors. The realized analytical sup-
port volumes, depicted in red in cross-section views of the flow, do not correspond to the imperative demand for
a full cross-sectional slice, cfr. Fig. 2. Inserting probes and screw samplers makes no difference, despite many
claims to the contrary in the marketplace. The deliverable is a far cry away from a correct, full slice of the stream,
cfr. Fig. 2. Illustration copyright KHE Consulting teaching collection, used with permission.
126 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

Fig. 10. Principal sketch of process sampling interface purporting to deliver physical increments that are repre-
sentative of the streaming matter in conveyed flows, ducted pipes or from reactors. The realized analytical sup-
port volume, depicted in red in cross-section views of the flow, do not correspond to the imperative demand for a
full cross-sectional slice, cfr. Fig. 2. Process control samplers, including pressure pipe and poppet samplers,
make no difference, despite many claims to the contrary in the marketplace. Illustration copyright KHE Consult-
ing teaching collection, used with permission.

Fig. 11. Principal sketch of process sampling interfaces purporting to deliver physical increments that are repre-
sentative of the streaming matter in conveyed flows, ducted pipes or from reactors. The realized analytical sup-
port volume (“sample inlet”) does not correspond to the imperative demand for a full cross-sectional slice, cfr.
Fig. 2. Poppet samplers makes no difference, despite many claims to the contrary in the marketplace. Illustration
copyright KHE Consulting teaching collection, used with permission.
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 127

Fig. 12. Principal sketch of process sampling interface purporting to deliver physical increments that are repre-
sentative of the streaming matter in conveyed flows and ducted pipes. The realized analytical support volume, here
depicted in green in cross-section view of the flow, do not correspond to the imperative demand for a full cross-
sectional slice, cfr. Fig. 2. This illustration shows the famous ‘shark fin’ sampler. It fails to extract a complete slice
of the moving matter but delivers only a narrow central ‘part of the whole stream all of the time’. The cross-sectional
areas indicated in red are structurally never available for sampling and will therefore never be able to comply with
TOS’ Fundamental Sampling Principle (FSP) resulting in a highly significant sampling bias. Despite many claims in
the marketplace, it is obvious that the shark fin sampler is not representative. Illustration copyright KHE Consulting
teaching collection, used with permission.

Fig. 13. Principal sketch of a process sampling interface purporting to deliver sensor signals that are representative
of the streaming matter in conveyed flows. The realized analytical support volume (red) does not in any way corre-
spond to the imperative demand for a full cross-sectional slice, cfr. Fig. 2. Many hopeful suggestions involve sen-
sors, or cameras, capturing views (narrow-angle, or broad-angle) of moving steams of matters, which fall under the
category process sampling. It is obvious however, that such solutions will always fail w.r.t. the imperative slice
rule, at best delivering sensor grab samples. The duality with traditionally extracted physical grab samples from
the top of the conveyed matter only is striking. Grab sampling is never a solution, neither in TOS’ domain, nor in
the PAT sensor domain. Illustration copyright KHE Consulting teaching collection, used with permission.
128 The Sampling Interface—a Critical Theory Of Sampling Success Factor In Process Sampling And Pat

Underlying assumption of cross-sectional homogeneity

Upon reflection, all the non-representative exemplifications presented above, would to some degree be able to function – if
only all instantaneous flow cross-sections were homogenous. This underlying assumption is widespread in very many pro-
cess technology contexts where TOS literacy has not been obtained.

A warning, also part of the process sampling interface discussion, concerns the by-pass valve, often used to try to mass-
reduce (sub-sample) a streaming flow of matter on the fly. While this objective is highly desirable, it is also very difficult to
realise in practice, because of a fundamental misunderstanding. Not any by-pass valve follows TOS’ demands for correct
increment cutting. In fact, of the five alternatives shown in Fig. 14, only one is TOS compliant! Check out your next blueprint
showing ‘by-pass streams’. Exactly how is this by-pass diversion effectuated? There are many examples in which use is
made of one of the non-compliant stream diversion options shown in Fig. 14, the effect of which can be of varying importance,
but never neglectable. For full description, see2,5,13.

Fig. 14. The by-pass valve enigma. How to divert a moving stream in a representative fashion? Of the five alterna-
tive designs for diverting a mass- and flow-reducing by-pass stream, only one is TOS-compliant (top rightmost
example). Illustration from KHE Consulting teaching collection, used with permission.

A (very small) compositional caveat

To be fair, many of the process sampling system solutions depicted above actually do work according to their fundamental
objectives of securing representative process samples – but this only applies to one-phase material systems, e.g. infinitely
diluted solutions (‘fluids’) or mono-disperse particulate materials (uniform materials) a.o. in which there by definition are no
compositional heterogeneity, no untoward grain-size distributions variability, and no grain size-grade heterogeneity issues.

Upon reflections, these latter constitute the overwhelmingly dominating proportion of the kind of materials treated with
process sampling. Thus, the sampling difficulties illustrated in this paper concerns what in TOS parlance is known as ‘sig-
nificantly heterogeneous materials’, which comprise the by far overwhelming proportion of materials targeted for process
sampling and analysis in science, technology, industry, trade, commerce, environment … and society.

An ongoing concern is that solutions devised for ‘easy’ chemical or physical systems have led to a various process sam-
pling interfaces that do a fair job within their quite specific restricting limits, but only here. Great problems ensue when such
system solutions unwittingly are being carried over to other types of materials of the significantly heterogeneous type –
because here they are bound to fail, as the intrinsic heterogeneity of such systems (compositional heterogeneity (CH),
distributional heterogeneity (DH), grain-size distribution heterogeneity and grain-size-grade heterogeneity obviously devi-
ates from easy-to-sample ‘uniform’ material systems.

So, unfortunately, as a caveat, this one is very, very small and it cannot be used as justification for carrying-over of any
process sampling solution from the ‘easy’ realm.
K.H. Esbensen and V. Sivalingam, Proceedings of WCSB10: TOS Forum Issue 11, 115–129 (2022) 129

How to gauge the performance of a process sampling solution?

All process system solutions, in which the sampling interface interacting with a specific moving material has not been previ-
ously characterized, must be subjected to a complete system’s performance validation. It is necessary to validate all sam-
pling systems for its status w.r.t. an a priori decided TSE variance threshold, which would qualify the sampling system as
‘fit-for-purpose’ representative, or not. There is no space here to go into detailed descriptions of the well-known quality
assessment approaches just listed below; but full details can be found in very many references within the TOS realm. Suffice
here to mention1-5.

Full system sampling performance facilities:

 Interleaved sampling (IS)

 Replication Experiment (RE)
 Variographic characterization (VC)

So, what to do, then?

TOS calls for compliance validation of all existing designs not previously characterised and for innovative compliance in all
future designs of process sampling interfaces. Validation in the view of the above presentations is of critical importance
regarding representativity, especially concerning already installed systems that are “too expensive to fail”. For all such cases,
the key issue is simple: does the existing, or the newly designed sampling solution, capture the entire cross-section of
flowing streams of matter in an TOS-appropriate manner, or not. It cannot be stated simpler …

The hopeful quest for a universal process sampling interface will probably go on – regarding both sensor sampling and
physical sample extraction. But is there a universal interface design? Whatever the answer to this question, TOS must serve
as the guiding framework for all future developments with Fig. 2 outlining the principal, extremely simple, yet apparently so
difficult demand: “Obey the slice rule – or representativity is lost”. However, there is no other way!

References
1. K.H. Esbensen, Introduction to the Theory and Practice of Sampling, IMPOpen, Chichester, UK. (2020).
2. F. F. Pitard, Theory of Sampling and Sampling Practice 3rd ed., CRC Press, Taylor & Francis Group, Boca Raton, FL.
(2019).
3. G. J. Lyman, Theory and Practice of Particulate Sampling an Engineering Approach, Materials Sampling & Consultancy
PTY LTD., Southport, Queensland, AUS. (2019).
4. TOS Forum. 10, https://www.impopen.com/tosf-toc/20_10,IMPOpen, 15-19, IMPOpen, Chichester, UK. (2020).
5. K.A. Bakeev (Ed.) Process Analytical Technology: Spectroscopic Tools and Implementation Strategies for the Chemi-
cal and Pharmaceutical Industries 2nd ed., John Wiley & Sons, LTD., West Sussex, UK (2010).
6. K.H. Esbensen and P. Minkkinen (Eds.) Proceedings of the First World Conference and Sampling and Blending, Spe-
cial Issue: 50 Years of Pierre Gy’s Theory of Sampling, Chemometrics and Intelligent Laboratory Systems, 74:1, Else-
vier B.V., Oxford, UK. (2004)
7. K.H. Esbensen and C. Wagner (Eds.) Proceedings for the 7th World Conference on Sampling and Blending, in TOS
forum, vol 5, IMPOpen, Chichester, UK (2015).
8. S.C. Dominy and K.H. Esbensen (Eds). Proceedings of the 8th World Conference on Sampling and Blending, AusIMM,
Carlton Victoria, AUS. (2017).
9. BGRIMM Group (Eds.) Proceedings for the 9th World Conference on Sampling and Blending, BGRIMM Group, Beijing,
China (2019).
10. K.H. Esbensen and B. Swarbrick, Multivariate Data Analysis: An introduction to Multivariate Analysis, Process Analyt-
ical Technology and Quality by Design, 6th edition, Camo Software A/S, Oslo, Norway (2018).
11. A.R. Novaes and B.M. Hidding, ‘Choosing metallurgical samplers or static cutters for process control in slurry: when
or why to avoid the increment delimitation error, in Proceedings of the 10th World Conference on Sampling and Blend-
ing. IMP Open, Chichester, UK, pp. 171–175 (2022).
12. D. Mittrup, O. Dominguez and M. Haest, “A new approach to implement quality assurance and quality control to tech-
nological innovations—quality of spectral data capture and processes in the minerals industry” in Proceedings of the
8th World Conference on Sampling and Blending, Ed. by S.C. Dominy and K.H. Esbensen, AusIMM, Carlton Victoria,
AUS. (2017).
13. A.P. Ferreira, J.C. Menezes and M. Tobyn (Eds.) Multivariate Analysis in The Pharmaceutical Industry, Elsevier, Aca-
demic Press, London, UK (2018).
14. R.J. Romañach, A. Joubert Castro and K.H. Esbensen, “WHAT are sampling errors—and WHAT can we do about
them? Part 1”, Spectrosc. Europe 32(2), 36–42 (2021). https://doi.org/10.1255/sew.2021.a11