Resolving the inverse problem in pulse response analysis of TAP reactors
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Resolving the inverse problem in pulse response analysis of TAP reactors".
Mira: Here's the gist of the paper. Pulse experiments in the temporal analysis of products (TAP) reactor are one of the most important methods for studying transient kinetics of gas-solid catalytic reactions.
Kai: Here, they explicitly recognize the inverse problem inherent in the Y-procedure and treat it using well-developed tools from the field of inverse problems.
Mira: The work develops a new method for dealing with this ill-posed step of the Y-procedure by formulating it as a discrete inverse problem and solving it via the truncated singular value decomposition (TSVD) method.
Kai: The core challenge addressed The main difficulty in applying the Y-procedure is the solution of Eq.
Paper summary: Kai: So, we're talking about this paper today called "Resolving the inverse problem in pulse response analysis of TAP reactors". It looks like they’re tackling a big problem in studying how gases react inside these kinds of reactors. Mira Exactly. The main thing is that the Y procedure, which is this model-free way to figure out the reaction rate R and concentration C from just measuring the outlet flux, has a real issue when you try to use it. Lev I’m thinking about that noise amplification thing; if you try to go backward from what you measured at the end of zone three, you get these tiny fluctuations amplified into huge errors in what was supposed to be happening earlier in the reactor Kai. Mira That's right. And this paper is saying they need to treat it as a proper inverse problem and use tools from that field instead of just using some standard noise filter.
Lev: They’re proposing framing the whole thing as a discrete inverse problem and solving it using something called truncated singular value decomposition, or TSVD Kai. This sounds like they're trying to get around the ill-posed step in the Y procedure where you have to find that unknown flux entering zone three given only what you measured at the outlet Mira.
Kai: Right. And they formulate this by representing the unknown flux entering zone three with a basis of localized square pulses, which lets them set up an equation like A w equals b Lev. So, instead of just guessing a smooth curve, they break it down into these specific pulse shapes and solve for the coefficients w Kai.
Mira: The problem is that the matrix A they get ends up having a very large condition number. That means the problem is really ill-posed, which necessitates using regularization techniques like TSVD Lev. And to pick how many of those modes to keep, they introduce this metric called Psi, which measures how far the reconstructed reaction rate function R departs from being a single-valued function of concentration C Kai.
Lev: So they’re not just throwing in a random number for the cutoff mode number; they are finding an optimal value by minimizing that Psi, and for state defining experiments, where R is one-to-one with C, this minimum is at twenty-eight Mira. That's a concrete piece of information about how to choose their regularization parameter.
Kai: And it sounds like they show this strategy works even when you’re using Fourier filtration too, because they find a minimum for Psi at lambda equal to thirteen Lev. It seems like this method can automate the choice of that filter parameter without needing someone to manually look at the results Mira.
Lev: They also demonstrate that this TSVD reconstruction produces better results than the old Fourier-filtration method, especially when dealing with nonlinear kinetics Kai. That’s a key claim because nonlinearity is where those standard smoothing techniques usually fall apart and start oversmoothing things Mira.
Kai: So, the implication here is that you can automate this Y procedure using these inverse problem tools instead of relying on human judgment for setting regularization or filtering parameters Lev. It moves it from being a manual process to something that can be automated.
Mira: And what this means for people working on reaction kinetics is that they might be able to reconstruct the R vs C relationship more accurately, especially when the reaction isn't linear Kai. But we have to remember the authors specifically state this method is not applicable to state altering experiments where R(C) is multivalued Lev.
Lev: And they do give a hint on how to guide that selection process by looking at a Picard plot, which shows you where the information-rich modes transition into noise-dominated ones Kai. It’s a visual way to see what’s working.
Kai: So, for those of us in the lab, it means we can use these discrete inverse problem techniques like Tikhonov regularization or even machine learning methods on top of this work Mira. It paves the way for other ways to handle this kind of problem Lev.
Mira: Ultimately, they show that you don't need to rely on ad hoc noise filters when dealing with the ill-posed nature of the Y procedure in pulse response analysis of TAP reactors Kai. This paper is a strong argument for using formal inverse problem methods here.
Lev: We saw how this method converges as the number of modes increases, unlike direct solutions which just blow up Mira. And they show that the reconstructed profiles for R and C converge on increasing M, which means you get better results by sampling more data Kai.
Kai: So, to wrap up on this paper, "Resolving the inverse problem in pulse response analysis of TAP reactors", it’s about taking a known difficult step in the Y procedure and solving it as a discrete inverse problem using TSVD Lev. It gives us an objective way to pick parameters instead of just guessing Mira.
Mira: It means we can automate pulse response analyses, potentially enabling real-time reconstruction of R versus C curves Kai. The paper suggests this could be applied to other reaction systems as well, provided the relationship R is single-valued with respect to C Lev.
Conclusion: Kai: So we've been looking at how they solved that tricky inverse problem in pulse response analysis of TAP reactors today, and now we need to talk about what that actually means for us as experimentalists.
Mira: Right, so basically they took this really hard step in the Y procedure—that part where you have to go backward in time from the outlet measurement—and they framed it as a discrete inverse problem using something called truncated singular value decomposition, or TSVD.
Lev: Yeah, and the main takeaway is that instead of just applying some ad hoc noise filter, which is what people usually do, they use this more rigorous mathematical approach to handle those amplified measurement noises.
Kai: It seems like the whole point of this work is making the Y procedure more automated. They show a strategy for picking a regularization parameter based on minimizing a measure called Psi, instead of just picking some arbitrary number.
Mira: That objective selection strategy is pretty neat because they found that for state-defining experiments, where the reaction rate R is just one value for every concentration C, the best cutoff mode number is twenty eight.
Lev: For real hardware running this stuff, that objective parameter selection helps a lot because you don't have to spend hours tuning a filter parameter by hand. It lets you automate it.
Kai: So what does this mean for the world of reaction engineering? It means we can reconstruct the relationship between the rate and concentration curve much more accurately than we could before, especially for nonlinear reactions.
Mira: Especially when the kinetics are nonlinear, where those standard smoothing techniques tend to oversmooth things and hide important details in the data. This TSVD reconstruction actually gives a better picture of that R versus C relationship.
Lev: But we have to keep an eye on what they say about limitations. They explicitly state this method doesn't work for state altering experiments where the reaction rate R is multivalued with respect to concentration C. That’s a big restriction.
Kai: So, it’s a powerful tool for single-valued systems, but we still have to be careful about when we apply it in our experiments. But this opens the door for using other inverse problem techniques on pulse response analysis moving forward.
Anjali Aleriaa, Evgeniy Redekopb, A. K. Sureshc, Jason R. Picardo
Department of Chemical Engineering, Indian Institute of Technology Bombay · Department of Chemistry, University of Oslo · Indian Institute of Technology Madras Zanzibar
physics.chem-ph, physics.data-an, physics.flu-dyn
Submitted: 2026-07-06
Updated: 2026-07-06
Comments: 15 pages, 11 figures
Journal ref: Chemical Engineering Science (2026)
DOI: 10.1016/j.ces.2026.125567
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 82/100
The gist: Pulse experiments in the temporal analysis of products (TAP) reactor are one of the most important methods for studying transient kinetics of gas-solid catalytic reactions.
Key concepts
- Y-procedure
- A model-free analysis framework used to infer the relationship between a reaction rate (R) and reactant concentration (C) from measurements of gas outlet flux in a thin-zone catalytic reactor. It involves three steps: reconstructing the flux entering the reaction zone, determining the input flux to that zone, and calculating the difference to find R(t).
- Ill-posed Problem
- The mathematical difficulty where solving for an unknown quantity (like flux entering Zone 3) from known measurements (outlet flux) is unstable. Small measurement noise gets severely amplified when trying to backtrack in time, making naive solutions unreliable.
- Truncated Singular Value Decomposition (TSVD)
- A regularization technique used to solve ill-posed problems. It involves calculating the SVD of a matrix and then discarding modes corresponding to very small singular values. This effectively filters out the noise present in high-frequency modes, leading to a stable and accurate solution.
- Objective Regularization Strategy
- A method developed for selecting the optimal regularization parameter (cutoff mode number) without human judgment. For state-defining experiments, this strategy minimizes a measure of 'multivaluedness' ($\Psi$), providing an automated way to choose the best smoothing level.
Terminology
Summary
Pulse experiments in the temporal analysis of products (TAP) reactor are one of the most important methods for studying transient kinetics of gas-solid catalytic reactions. p. 0 Here, they explicitly recognize the inverse problem inherent in the Y-procedure and treat it using well-developed tools from the field of inverse problems. p. 0 The work develops a new method for dealing with this ill-posed step of the Y-procedure by formulating it as a discrete inverse problem and solving it via the truncated singular value decomposition (TSVD) method. p. 1
The core challenge addressed
The main difficulty in applying the Y-procedure is the solution of Eq. (1) in zone 3, given only outlet boundary conditions. p. 3 When backtracking from the measured F∗out to obtain F3(L, t), naive solutions amplify small fluctuations in F∗out so that the reconstructed flux is dominated by noise. p. 3 Previous implementations used Fourier-filtering, but this requires human intervention to decide the degree of filtration and can suffer from oversmoothing. p. 1 The proposed method avoids ad hoc filtering by representing the unknown flux entering zone 3 with a basis of localized square pulses to formulate a discrete inverse problem Aw = b. p. 1
Formulation of the discrete inverse problem
The formulation represents the flux entering zone 3, F3(L, t), as a sum of basis functions using Eq. (10). The next step involves determining the outlet flux that emerges from zone 3 when the input flux to zone 3 takes the form of any one basis function. p. 5 This leads to a system of M linear algebraic equations for the unknown coefficients w, resulting in Aw = b. p. 6
Regularized solution via TSVD
The ill-posed nature of the problem manifests in A having a very large condition number, necessitating regularization methods like TSVD. The direct solution is obtained as wfull = VΣ−1UTb = XM i=1 uT i b σi vi. p. 6 The noise in b is present in the high-index modes with small singular values, and truncating the SVD expansion to retain only small index components yields wTSVD = Xmcut i=1 uT i b σi vi. p. 6
Objective selection of regularization parameter
For state-defining experiments where R is a single-valued function of C, an objective strategy is developed by defining a measure of multivaluedness Ψ = Z Cmax ∫[0] [R+(c) − R−(c)]2 dc. Minimizing this degree of multivaluedness Ψ yields the optimal value for the cutoff mode number mcut, which is shown to be mcut = 28 for state-defining experiments. p. 9 This strategy can also be applied to Fourier-filtration, where a minimum in Ψ is found at λ = 13. p. 1
Performance and conclusion
The TSVD reconstruction is shown to produce significantly better reconstructions than the Fourier-filtration method, especially for nonlinear kinetics. The optimal TSVD reconstruction, selected by minimizing Ψ, is found to be quite good for all tested values of mcut. p. 8 The ability to select the regularization parameter objectively enables the automation of the Y-procedure. p. 0 This method demonstrates that human judgement is no longer required to choose the regularization or filtration parameter. p. 1 The formulation paves the way for further development of the Y-procedure through other well-developed techniques from the field of inverse problems. p. 0 The method is not applicable to state altering experiments where R(C) is multivalued. p. 0 The Picard plot reveals the transition from information rich modes to noise-dominated modes, guiding the choice of mcut. p. 2 The TSVD method converges as M increases, unlike direct solutions which diverge. p. 8 The reconstructed profiles of R(t) and C(t) converge on increasing M for the TSVD method. p. 8 The TSVD reconstructions are more accurate than those obtained from Fourier-filtration, particularly near the maximum value of concentration where the Fourier-filtration produces oversmoothing. p. 8 The paper concludes that this method can be automated, enabling real-time reconstruction of R−C curves. p. 8 The formulation paves the way for further development of the Y-procedure through the use of other well-developed techniques from the field of inverse problems. p. 0 The paper is 2607. p. 0047763.
--- Page 1 ---
Resolving the inverse problem in pulse response analysis of TAP reactors Anjali Aleriaa, Evgeniy Redekopb, A. K. Sureshc, Jason R. p. 0 Picardoa,∗ aDepartment of Chemical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai, 400076, India bDepartment of Chemistry, University of Oslo,, Oslo,, Norway cAbstract Pulse experiments in the temporal analysis of products (TAP) reactor are one of the most important methods for studying transient kinetics of gas-solid catalytic reactions. p. 0 The Y-procedure (Yablonsky et al. p. 0, Chem. Eng. Sci. 62, 6754, 2007) is a model-free analysis framework for inferring the relationship between the reaction-rate R and the reactant concentration C from measurements of the outlet flux of gas. p. 0 While elegant in conception, its application is hindered by the amplification of measurement noise that results from having to backtrack diffusive transport from the outlet to the reaction zone. p. 0 Here, we explicitly recognize the inverse problem inherent in the Y-procedure and treat it using well-developed tools from the field of inverse problems. p. 0 While previous implementations of the Y-procedure used Fourier-based filtering, we do not pre-process the measurements with an ad hoc noise-filter. p. 0 Instead, we use a basis of localized square pulses to formulate a discrete inverse problem, whose regularized solution is obtained via the truncated singular value decomposition (TSVD) method. p. 0 This method requires one to select a cutoff mode number; while we show how the choice of this regularization parameter can be guided by a Picard plot, we also develop an objective selection strategy for state defining experiments, for which R(C) is a single-valued function, i. p. 0e., for every value of C, there is a single value of R. p. 0 We apply our proposed inverse-problem approach to synthetic measurement data corresponding to linear and nonlinear reactions and compare the results with the Fourier-filtration method. p. 0 The former produces better reconstructions of the R vs C relationship, especially for nonlinear reactions. p. 0 Our work facilitates the automation of pulse response analyses and enables the application of other discrete inverse-problem techniques, such as Tikhonov regularization or machine-learning methods. p. 0 Keywords: TAP reactor, Y-procedure, Inverse problem, SVD
--- Page 2 ---
Zone 2 Input Pulse Valve QMS Fin (t) Fout(t) Zone 1 Zone 3 Figure 1: Schematic of a thin-zone TAP reactor. A pulse of gas molecules diffuses through the bed of inert particles in zone 1, reacts in the thin zone of catalystloaded particles in zone 2, and diffuses again through the inert zone 3, to finally emerge as the outlet pulse. p. 1 p. 1 Fin(t) and Fout(t) are the inlet and outlet flux profiles;the former is known while the latter is measured by the Quadrupole Mass Spectrometer (QMS). p. 1 p. 1 (TZTR) [20], wherein an excellent approximation to a uniform concentration is realized in the middle reaction zone by making it sufficiently thin. p. 1 p. 1 The Y-procedure takes advantage of the separation between reaction and transport in the thin-zone TAP reactor, to reconstruct the temporal variation of reaction rates and concentrations in the thin reaction zone, thereby yielding a reaction rate vs. p. 1 p. 1 concentration curve without having to assume a kinetic model. p. 1 p. 1 Since its introduction, the Y-procedure has been extended to include irreversible adsorption and to compute the surface concentration; it has also been implemented within a data-science workflow. p. 1 p. 1 The Y-procedure is based on three key steps: (i) The measured outlet flux profile (leaving zone 3) is used to reconstruct the flux leaving the thin reaction zone (entering zone 3), as well as the concentration C(t) in the reaction zone (see Fig. p. 1 p. 1 1); (ii) The known temporal variation of the input flux to zone 1, i. p. 1 p. 1e, the input pulse profile, is used to obtain the flux entering the thin reaction zone (zone 2); (iii) The difference between the fluxes entering and leaving the reaction zone is used to compute the reaction rate R(t). p. 1 p. 1 Now, while the second step involves the straightforward simulation of diffusion in zone 1, the first step requires one to go backward in time and infer the flux that must have entered zone-3 so as to yield the measured outlet flux at the exit of zone 3. p. 1 p. 1 Solving the diffusion equation backwards in time is Hadamard-illposed. p. 1 p. 1 As a consequence, even tiny measurement noise in the outlet flux will get amplified by a naive backward solution so that the reconstructed flux entering zone-3 is entirely corrupted by noise. p. 1 p. 1 Recognizing this issue, Yablonsky et al. p. 1 p. 1 [19] used Fourier-filtering to smooth the measured outlet flux. p. 1 p. 1 This pre-processing step has henceforth been the typical approach to dealing with the illposedness of the first step. p. 1 p. 1
--- Page 3 ---
Zone 1 Zone 3 Zone 2 Input Pulse Valve QMS Fin (t) Fout(t) Figure 1: Schematic of a thin-zone TAP reactor. A pulse of gas molecules diffuses through the bed of inert particles in zone 1, reacts in the thin zone of catalystloaded particles in zone 2, and diffuses again through the inert zone 3, to finally emerge as the outlet pulse. Fin(t) and Fout(t) are the inlet and outlet flux profiles;the former is known while the latter is measured by the Quadrupole Mass Spectrometer (QMS). (TZTR) [20], wherein an excellent approximation to a uniform concentration is realized in the middle reaction zone by making it sufficiently thin. The Y-procedure takes advantage of the separation between reaction and transport in the thin-zone TAP reactor, to reconstruct the temporal variation of reaction rates and concentrations in the thin reaction zone, thereby yielding a reaction rate vs. concentration curve without having to assume a kinetic model. Since its introduction, the Y-procedure has been extended to include irreversible adsorption and to compute the surface concentration; it has also been implemented within a data-science workflow. The Y-procedure is based on three key steps: (i) The measured outlet flux profile (leaving zone 3) is used to reconstruct the flux leaving the thin reaction zone (entering zone 3), as well as the concentration C(t) in the reaction zone (see Fig. 1); (ii) The known temporal variation of the input flux to zone 1, i.e, the input pulse profile, is used to obtain the flux entering the thin reaction zone (zone 2); (iii) The difference between the fluxes entering and leaving the reaction zone is used to compute the reaction rate R(t). Now, while the second step involves the straightforward simulation of diffusion in zone 1, the first step requires one to go backward in time and infer the flux that must have entered zone-3 so as to yield the measured outlet flux at the exit of zone 3. Solving the diffusion equation backwards in time is Hadamard-illposed. As a consequence, even tiny measurement noise in the outlet flux will get amplified by a naive backward solution so that the reconstructed flux entering zone-3 is entirely corrupted by noise. Recognizing this issue, Yablonsky et al. [19] used Fourier-filtering to smooth the measured outlet flux. This pre-processing step has henceforth been the typical approach to dealing with the illposedness of the first step.
--- Page 4 ---
In Yablonsky et al. [19], the issue of noise-amplification was dealt with by Fourier-filtering the measured signal F∗out(t). p. 3 Since this operation requires the use of the Fourier transform, it is convenient to also solve Eq. p. 3 (1) using the Fourier transform method, as done in [19]. p. 3 Applying the Fourier transform in time to Eq. p. 3 (1) in zone 3 yields a boundary value problem for C˜3(x, ω) which is the Fourier transform of C3(x, t) (here ∼ denotes a Fourier transform and ω is the frequency). p. 3 A similar treatment of zone 1 yields a relation between F˜1(L, ω) and F˜in(ω). p. 3 In this manner, the Fourier transforms of R and C can be expressed in terms of the Fourier transform of F∗out(t). p. 3 The Discrete Fourier Transform (DFT) is used to numerically compute the Fourier transforms. p. 4 If the input and output fluxes and sampled at Nt time instants at regular intervals ∆t, over a total time window T, then the DFT yields the Fourier transform at discrete frequencies ω = 2πn/T where n = −Nt/2 + 1, …, Nt/2. p. 4 The Inverse DFT, applied to R˜ and C˜, then yields R(t) and C(t) on the uniform temporal grid [19]. p. 4
--- Page 5 ---
The transfer functions W and Z in Eq. (8c) may be shown to increase rapidly with ω and hence, as anticipated, small amplitude high-frequency noise in F∗out(t) is greatly amplified in the reconstructed R(t) and C(t). p. 4 In order to reconstruct the underlying, true relationship between R and C, Yablonsky et al. p. 4 [19] applied a Fourier filter to F∗out(t) by multiplying F˜ ∗out(ω) with a smoothing filter S (ω) given by S (ω) = exp(−ω2/2∆t2/λ2 / 2), where λ is a smoothing parameter. p. 4 The value of λ is, of course, unknown a priori. p. 4 It is typically chosen based on a qualitative examination of the results obtained using different values of λ. p. 4
--- Page 6 ---
Figures 3(a-c) presents reconstructions of the flux entering zone 3, F3(L, t), obtained by using the Fourier-filtration method with λ = 5, 10, and 20. The reconstructed concentration in the thin reaction zone, C(t), is presented in the insets of Figs. p. 4 3(a-c). The exact functions for F3(L, t) and C(t) from the forward simulation that generated Fout(t) is also shown for comparison (red-dashed lines). p. 4 The reconstructions of the reaction rate function R(C) are shown in Figs. p. 4 3(d-f), along with the exact kinetic curve (R = kC, with k = 0. p. 41 m s−1). We see that a small value of λ allows spurious fluctuations to persist in the reconstructed results (Figs. p. 4 3(a,d)), while a large value of λ leads to oversmoothing—the peak of F3(L, t) is not fully captured (Fig. p. 4 3(c)) and so the curve of R(C) does not extend over the entire range of C values realized by the pulse (Fig. p. 4 3(f)).
--- Page 7 ---
Indeed, the ad hoc nature of the filter in Eq. (9) and the subjectivity associated with choosing λ are both drawbacks of the Fourier-filtration procedure. p. 4 Both these difficulties are alleviated by the inverse problem methodology presented in the next section. p. 4
--- Page 8 ---
The TSVD method requires one to first compute the singular value decomposition (SVD) of the M × M matrix A. The two sets of M vectors ui and vi form orthogonal bases for the range and the domain of the operator A [32]. p. 6 The measurement noise in b will be present in the high-index (large i) highfrequency SVD modes that have small values of σi. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Implement an automated Y-procedure using a discrete inverse problem formulation to reconstruct reaction rates without assuming a kinetic model. This allows for the application of other discrete inverse-problem techniques such as Tikhonov regularization or machine learning methods.
-
Develop an objective selection strategy for the regularization parameter by minimizing a positive semi-definite measure of multi-valuedness, Psi, in state defining experiments. This strategy can be used to select the cutoff mode number mcut in the truncated singular value decomposition method for automated pulse response analysis.
-
Integrate a machine learning approach for inverse problems that is agnostic to the specific basis set used by methods like TSVD. This allows the system to learn particularities of inlet and outlet pulses, potentially offering better reconstructions, especially in the presence of non-Gaussian noise.
-
Extend the inverse problem formulation to handle state altering pulses where the reaction rate R depends on both gas phase concentration C and adsorbed species concentration. Develop a strategy for selecting regularization parameters when the true R vs C curve is multivalued in these experiments.
Sources
Related papers
- A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions
- A Machine Learning Model for the Chemistry of a Solvated Electron
- Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set
- Charge Transfer with a Spin. II: A Framework for Diabatization which Localizes Charge and Spin
- Diabatic Seam Space Sampling for Hydrogen Tunneling Systems with Nuclear-Electronic Orbital Theory
- Augmented Roothaan-Hall Hessian Applied to Spin-Restricted Open-Shell Density-Functional Theory