Photo-induced currents and short-term memory for reservoir computing in a ferroelectric semiconductor

Atreya Majumdar; Dennis Meier; Ingvild Hansen; Karin Everschor-Sitte; Manuel Zahn; Yan Meng Chong

REVIEW 3 major objections 5 minor 47 references

Photo-induced currents and short-term memory for reservoir computing in a ferroelectric semiconductor

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The time-evolving photocurrent of the ferroelectric semiconductor ErMnO3 is a physical reservoir whose fading memory lets a simple linear readout identify the preceding light pulse with ~93% accuracy.

desk verdict Solid experimental demonstration of photocurrent memory in ErMnO3, but the missing n=1 control leaves the reservoir-computing claim overreaching. read the letter →

arxiv 2607.20810 v1 pith:EEF5TTSP submitted 2026-07-23 cond-mat.mtrl-sci

Yan Meng Chong , Atreya Majumdar , Manuel Zahn , Ingvild Hansen , Karin Everschor-Sitte , Dennis Meier This is my paper · ORCID

classification cond-mat.mtrl-sci PACS 72.40.+w

keywords reservoircomputingphysicalphoto-inducedcurrentshort-termmemoryferroelectricsemiconductorErMnO3hexagonalmanganitevirtualnodes

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the photo-induced current of the ferroelectric semiconductor ErMnO3 is, on its own, enough to perform the memory-dependent step of reservoir computing: recognizing which input pulse came immediately before the current one. Under white-light illumination the photocurrent rises nonlinearly and, after the light is turned off, relaxes over timescales from tens of milliseconds to seconds; sampling that time trace at ten points turns one slow physical signal into a high-dimensional reservoir state. Because the relaxation outlasts the inter-pulse interval, the state carries a fading trace of the previous pulse — directly evidenced by a paired-pulse-facilitation ratio above 100%. A linear readout trained in a single least-squares step then identifies the present pulse at ~99% accuracy and the immediately preceding pulse at ~93%, whereas the same readout on the raw input intensities scores only ~33%, the chance level. The authors further show the relaxation timescale is tunable through contact engineering and reducing-atmosphere annealing, which is what would let such a reservoir be matched to the timescale of a given task.

What carries the argument

The key mechanism is time-multiplexed virtual-node sampling applied to a single physical variable. The photocurrent response to each light pulse is sampled at n=10 points spaced by a time interval τ across the 5 s response window, forming the reservoir state vector x(k) = [1, r(tk), r(tk+τ), …, r(tk+(n−1)τ)]^T, the leading 1 acting as readout bias. Two least-squares-trained weight vectors map this same state onto predictions of the present and previous pulse intensities, with the continuous outputs thresholded at the midpoints between the three intensity levels. The physical substrate — nonlinear current rise plus slow trap-governed relaxation — supplies the nonlinearity and the fading memor

What would settle it

A decisive test: train the readout on the raw photocurrent without square-root background subtraction, and separately on the baseline-corrected signal, and check that past-pulse accuracy stays near 93% under both treatments. A complementary control is to shuffle the past-pulse labels against the measured traces and confirm accuracy collapses to ~33%; and to fit the baseline on a reversed or reshuffled pulse sequence — if the fitted a+b√(t−t0) changes with pulse order, the drift is not history-independent and the subtraction is not neutral.

Watch

Extended reading notes

Core claim

The central discovery is that the dynamic photocurrent of ErMnO3 is not a nuisance signal to be read at steady state but a usable physical reservoir. The shape of the current during a 5 s light pulse depends on both the present intensity and the intensity of the preceding pulse: averaging over all two-pulse configurations yields three distinct current profiles for each present-pulse level, one per possible past pulse. Sampling each response window at ten uniformly spaced times produces the reservoir state that feeds a linear readout. Trained by ordinary least squares, the readout labels the present pulse with ~99% accuracy and the immediately preceding pulse with ~93% accuracy, where the sam

Load-bearing premise

The load-bearing premise is that the slow, monotonic baseline drift of the photocurrent over the 33-minute run carries no information about the pulse history and can be cleanly removed by fitting a square-root function; if that drift is actually a long-term accumulation of the photo-response, the subtraction could either erase part of the memory the reservoir is supposed to exploit or imprint an artificial trend that inflates past-pulse accuracy.

Editorial extensions

If this is right

A single ferroelectric semiconductor device, with no recurrent network in software, can carry the memory-dependent part of a temporal classification task: a linear readout on sampled photocurrent lifts past-pulse recognition from chance level (33%) to about 93%.
The reservoir's memory timescale is an engineering parameter — Schottky contacts give the fastest response, as-grown Ohmic contacts an intermediate one, and annealing in a reducing atmosphere prolongs relaxation to seconds — so the same material can be matched to tasks with different characteristic timescales.
Because the present and past predictions are read from the same reservoir state through different trained weights, switching to a new task requires retraining only the linear readout, not re-running the physical reservoir.
The noiseless circuit model reaches ~100% past-pulse accuracy through the same training pipeline, which the paper treats as evidence that the photocurrent waveform encodes full information and that the experimental gap is noise and baseline drift.
Several relaxation timescales could in principle be combined in one device to raise reservoir dimensionality, and the ferroelectric domain and domain-wall structure offers further, locally distinct responses for downscaling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the same protocol could be run as a quantitative short-term-memory benchmark (for example, memory capacity of a linear readout on the reservoir state), which would allow ErMnO3 to be compared directly, on a task-independent scale, with memristive, spintronic, and photonic reservoirs.
Editorial inference: the task only probes one step of memory (the immediately preceding pulse). Since the relaxation is double-exponential with decay constants near 66 ms and 468 ms, the state likely carries information about pulses further back; testing recognition of u(k−2) and u(k−3) would measure the effective memory depth and could clarify whether the experimental shortfall is a memory-depth
Editorial inference: the background-subtraction step deserves a stress test. A control that trains the readout on the raw photocurrent without the square-root baseline fit, or fits the drift with a history-dependent model and checks whether past-pulse accuracy moves, would show whether the 93% figure is robust to how the slow drift is treated.
Editorial inference: a single-node control — training the same readout on the photocurrent sampled at one optimized time instead of ten — would isolate whether the benefit comes from the temporal trajectory shape or merely from the current level carrying the previous pulse's tail.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This paper does something real and useful: it shows that the photocurrent response of ErMnO3 carries memory of the previous light pulse and that this memory is tunable via contact engineering. The measurement methodology is careful — 1000 random train/test splits, a fair 33% baseline, clean figures. The ~93% past-pulse accuracy is a legitimate experimental result; the material does have short-term memory.

The soft spot is that the paper attributes that accuracy to the reservoir transformation — the 10 time-multiplexed virtual nodes — but never runs the control that would justify that attribution. The baseline it compares against is the raw input intensities u(k), which trivially contain no information about u(k−1). The right baseline is a single photocurrent sample, r(t_k), at the onset of the pulse. Given the reported relaxation times (20 ms to several seconds) and the 5 s pulse window, r(t_k) is essentially the tail of the previous pulse's response, so a one-dimensional state may already separate the three intensity classes. The accuracy-versus-n curve in Appendix C starts at n=2; n=1 is missing. If n=1 already gives ~93%, then the time-multiplexed reservoir is not doing the work — the memory is in the raw current, and the 'reservoir computing' framing is an overstatement. If n=1 were poor and the accuracy jumped at n=2 or 3, the claim would hold. Right now we don't know.

The baseline-drift subtraction (Appendix A) is a lesser concern: a smooth square-root drift is nearly orthogonal to a random pulse sequence, so it's unlikely to fabricate 93%; at worst it may remove some genuine long-term memory. The equivalent-circuit simulation (Appendix D) is fitted to the same transient it is then used to explain, which is circular but only used to attribute the 7% gap to noise, so it doesn't affect the central result. Data and code are stated to be on Zenodo but no link or hash appears in the text; minor workflow issue.

Bottom line: this is an incremental but sound experimental contribution, and the authors are honest with their methods. But the central claim needs the n=1 control before the reservoir-computing interpretation is justified. If I were the editor, I'd send it to review, and the referee should ask for that single experiment.

Referee Report

3 major / 5 minor

Summary. The paper reports an experimental demonstration of photo-induced currents in the ferroelectric semiconductor ErMnO3 as a physical reservoir for a temporal classification task. A random sequence of light pulses with three intensity levels is applied, and the resulting photocurrent is sampled at 10 virtual nodes within each 5 s pulse window. A linear readout, trained by least squares, classifies the present and immediately preceding pulse intensities. The authors report ~99% accuracy for the present pulse and ~93% for the past pulse, compared with ~33% for a readout trained on the raw input intensities. They also present a circuit-model simulation that reaches ~100% accuracy, and they show that the photocurrent relaxation timescale can be tuned via contact type and annealing.

Significance. If the central claim holds, the work is a useful experimental demonstration of short-term memory in a light-driven ferroelectric-semiconductor device, with strengths including an open data/code repository, 1000 random train/test splits, and a fair comparison to a trivial no-memory baseline. However, the specific attribution of the improvement to time-multiplexed reservoir computing (the high-dimensional projection) is not fully established because the natural single-node control is missing. The empirical baseline-drift model also rests on an unvalidated assumption. The paper is of interest to the physical-reservoir-computing community, but the experimental evidence needs additional controls to support the reservoir-computing interpretation.

major comments (3)

[Section II, Fig. 4, Appendix C] The 'without reservoir' baseline in Fig. 4 is trained on the raw input intensities u(k), which trivially cannot contain information about u(k−1) in a random sequence. The relevant control is a single photocurrent sample (n=1), such as r(t_k) at pulse onset. Figure 2(c,d) shows that the current at the onset of the present pulse already differs substantially for different past pulses, so n=1 may already achieve high accuracy. The accuracy-versus-n curve in Fig. 6 starts at n=2 and the text does not report n=1. Please provide the n=1 accuracy (and ideally a baseline using a single measured current readout) to show that the time-multiplexed, high-dimensional reservoir state is actually responsible for the ~93% accuracy. Without this, the conclusion in Section II that 'the reservoir transformation, not the raw input, supplies the information about u(k−1)' is under-supported.
[Appendix A, Eq. (A1)] The baseline drift is modeled as a fitted square-root function I_bg(t)=a+b√(t−t0) and is assumed to be 'uninformative about the immediate pulse history.' This is an empirical assumption. If the drift contains history-dependent components (e.g., persistent photocurrent accumulation), the background subtraction could remove part of the memory the reservoir is supposed to exploit, or impose an artificial trend. Although the random pulse sequence makes it unlikely that the drift alone fabricates 93% accuracy, a robustness check (e.g., repeating the pipeline without background subtraction or with an alternative detrending) would support the assumption and strengthen the central claim. Please add such an analysis or explicitly discuss the sensitivity.
[Section II, Appendix D] The circuit-model simulation is fitted to a single experimental transient (Fig. 7(c)) and then applied to the pulse sequence to obtain ~100% past-pulse accuracy. Since the model parameters are derived from the same material's response, this simulation is not independent evidence that the photocurrent waveform 'already encodes enough' for perfect recognition. The statement in Section II attributing the experimental ~7% shortfall to 'measurement non-idealities' is therefore partly speculative. Please soften this claim or provide a more controlled simulation study (e.g., varying the memory timescale and showing the accuracy degrades when τ1, τ2 are shortened).

minor comments (5)

[Appendix A, Eq. (A1)] The equation is written as 'I_bg(t) = a+b√t−t0', which is ambiguous; the square root should clearly cover (t−t0), i.e., a+b√(t−t0).
[Appendix C, Fig. 6] The accuracy-versus-n curve starts at n=2. Please state explicitly why n=1 is not included, and if possible add the n=1 point (even as an open symbol) to the figure.
[Section II, paragraph after Fig. 2(a)] The comparison of current values at t=25–30 s and t=150–155 s is described as showing dependence on preceding light pulses, but this example chiefly illustrates the slow baseline drift. Rephrase to avoid confusing drift with memory from u(k−1).
[Appendix D] The phrase 'developed by intention pictorially' is unclear and should be rewritten. Also, the choice R4 = R3 A2/A1 is not explained; please provide the rationale or a reference.
[Appendix B] When using scikit-learn's LinearRegression with a constant column already in X, the default fit_intercept=True makes the intercept redundant. Please specify fit_intercept=False (or equivalently mention that the constant column supplies the intercept).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central result is an empirical classification using measured photocurrent reservoir states with a least-squares readout; auxiliary simulation and missing n=1 control are support/interpretation concerns, not reductions of the claim to its inputs.

full rationale

The derivation chain is experimental rather than definitional. The paper measures photocurrent responses to a random sequence of light pulses, samples each response into 10 virtual nodes, and trains a linear least-squares readout to classify the present and past pulse intensities. The reported accuracies (present ~99%, past ~93%) are measured outcomes of this pipeline, not quantities that are fitted into the pipeline and then read back out. The 'without reservoir' baseline uses the raw input intensities alone, which is a legitimate control for whether the photocurrent transformation adds information about u(k-1), even though it is not a control for whether one photocurrent sample would suffice. The absence of an n=1 node count in Appendix C (the curve starts at n=2) is a missing control for the specific claim that the time-multiplexed, high-dimensional reservoir projection is necessary; however, this is an under-support/interpretational limitation, not a circular reduction. Appendix A's background subtraction assumes the slow drift carries no immediate-pulse-history information; if that assumption is wrong, the subtraction could alter the measured memory, but this is an untested empirical assumption, not an equation-level tautology. The Appendix D equivalent-circuit simulation is fitted to a photocurrent transient and then used to illustrate near-100% simulated accuracy; because the classification accuracy is not used to set the circuit parameters, this is a consistency/simulation result rather than an independent prediction, but it does not reduce the central experimental claim to its own inputs. The only overlapping-author citation (Ref. 43, used for the circuit model) is prior dielectric-spectroscopy work and is not load-bearing for the main experimental demonstration. Overall, the core result is self-contained empirical evidence: a measured dynamical system, a linear readout trained by least squares, and a test-set accuracy that is not guaranteed by the training procedure.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central experimental claim is empirical and does not depend on a derived model; it uses standard linear regression. The main auxiliary assumptions are the fitted baseline-drift model and the fitted circuit simulation, both clearly described but not independently validated.

free parameters (3)

Baseline drift coefficients a, b = fitted per dataset (Eq. A1)
The square-root baseline I_bg(t)=a+b√(t−t0) is fitted to the full active window and subtracted before reservoir feature extraction. If misspecified, this could bias the 93% result.
Equivalent-circuit parameters C1, C2, R4 = C1=τ1/R1, C2=τ2/R2, R4=R3·A2/A1; R1=R2=3×10^5 Ω, R3=3×10^7 Ω chosen arbitrarily
Used only for the auxiliary noiseless simulation (Appendix D). These parameters are fit to or chosen for one experimental decay curve, not for the central experimental reservoir result.
Number of reservoir nodes n = n=10
Selected because accuracy saturates by n≈8 (Appendix C). This is a hyperparameter choice, not fitted to the target, and does not affect the qualitative conclusion.

assumptions (4)

standard math Ordinary least-squares regression gives the optimal readout (Appendix B)
Standard closed-form linear regression is used to train the readout weights; this is a conventional and unproblematic assumption.
ad hoc to paper The slowly drifting baseline is uninformative about immediate pulse history and can be modeled as a+b√(t−t0)
Appendix A. This is load-bearing for isolating short-term memory from the measured photocurrent; if false, the 93% result could be biased by the drift correction.
domain assumption Photocurrent relaxation dynamics are reproducible and provide fading memory
Section II and Fig. 1(e). The paired-pulse facilitation is interpreted as short-term memory; the reservoir interpretation relies on this reproducibility.
ad hoc to paper The equivalent circuit model with two parallel RC branches adequately emulates the photocurrent transients
Appendix D. The circuit is modified from Ref. 43 and its parameters are fit to a single decay curve; this assumption only affects the auxiliary simulation, not the central experimental claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Photo-induced currents and short-term memory for reservoir computing in a ferroelectric semiconductor." pith.science (2026). https://pith.science/paper/EEF5TTSP

@misc{pith2026260720810,
  author       = {Pith},
  title        = {Pith review of: Photo-induced currents and short-term memory for reservoir computing in a ferroelectric semiconductor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EEF5TTSP}},
  note         = {Machine review of arXiv:2607.20810}
}

abstract

Physical reservoir computing represents an energy efficient approach for processing temporal signals by exploiting the intrinsic nonlinear dynamics and fading memory of a physical system. Recently, ferroelectric semiconductors moved into focus as reservoir materials motivated by their versatile electronic responses to external stimuli. Here, we explore the fundamental possibility to recognize time-varying light pulses via photo-induced currents, using the small-band-gap p-type semiconductor ErMnO$_3$ as a model system. Under white light illumination, ErMnO$_3$ exhibits non-linearly evolving photo-induced currents and controllable relaxation dynamics that naturally realize the high-dimensional projection and fading memory capabilities required for reservoir computing. The reservoir capability of ErMnO$_3$ is reflected by the improved recognition accuracy of "Past" input pulses, which increases from ~33% to ~93% after applying reservoir transformation to the input signal. The results present ferroelectric hexagonal manganites as a promising platform for photo-induced current-based reservoir computing and highlight the potential of light-driven oxide semiconductors for temporal information processing.

Figures

Figures reproduced from arXiv: 2607.20810 by the authors.

**Figure 2.** FIG. 2. Short-term memory in the photo-induced current of ErMnO [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** FIG. 3. Reservoir readout for pulse recognition using virtual nodes. (a) Photo-induced current in response to a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: FIG. 4. Recognition accuracy for identifying the present [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: FIG. 5. Removal of the photocurrent baseline drift by back [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: FIG. 6. Recognition accuracy as a function of the number [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: FIG. 7. Equivalent circuit modeling of the dynamic electric response of ErMnO [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references

[1]

the bulk material. While the intrinsic dielectric constantε bulk creates a capacitive contribution, the real part resistance consists of a frequency indepen- dent partσ dc and a so-called universal dielectric response (UDR) 44 with power-law scaling with re- spect to frequencyω,σ UDR ∝ω ν,0< ν <1
[2]

the contacts between electrodes and bulk material, generating via a Maxwell-Wagner relaxation a RC- circuit contribution characterized byσ contact and εcontact
[3]

Based on these considerations, we derive a modified circuit representing the photocurrent scenario

the ferroelectric domain walls, acting as additional capacitive barrier layers, characterized byσDW and εDW. Based on these considerations, we derive a modified circuit representing the photocurrent scenario. It is de- veloped by intention pictorially and can be abstracted and in principle treated more rigorously by signal pro- cessing methods: •In the fi...
[4]

Lukoševičius and H

M. Lukoševičius and H. Jaeger, Reservoir computing ap- proaches to recurrent neural network training, Computer Science Review3, 127 (2009)

2009
[5]

Lukoševičius, H

M. Lukoševičius, H. Jaeger, and B. Schrauwen, Reservoir computing trends, KI - Künstliche Intelligenz26, 365 (2012)

2012
[6]

Verstraeten, B

D. Verstraeten, B. Schrauwen, D. Stroobandt, and J. Van Campenhout, Isolated word recognition with the liquid state machine: a case study, Information Processing Let- ters95, 521 (2005)

2005
[7]

Jaeger and H

H. Jaeger and H. Haas, Harnessing nonlinearity: Predict- ing chaotic systems and saving energy in wireless com- munication, Science304, 78 (2004)

2004
[8]

Pathak, B

J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott, Model- free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach, Physical Re- view Letters120, 024102 (2018)

2018

Show all 47 references

[9]

Arcomano, I

T. Arcomano, I. Szunyogh, J. Pathak, A. Wikner, B. R. 10 Hunt, and E. Ott, A machine learning-based global at- mospheric forecast model, Geophysical Research Letters 47, e2020GL087776 (2020)

2020
[10]

Verstraeten, B

D. Verstraeten, B. Schrauwen, M. D’Haene, and D. Stroobandt, An experimental unification of reservoir computing methods, Neural Networks20, 391 (2007)

2007
[11]

Tanaka, T

G. Tanaka, T. Yamane, J. B. Heroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Recent advances in physical reservoir comput- ing: A review, Neural Networks115, 100 (2019)

2019
[12]

Jaeger,Short term memory in echo state networks, GMD Report 152 (German National Research Center for Information Technology, 2001)

H. Jaeger,Short term memory in echo state networks, GMD Report 152 (German National Research Center for Information Technology, 2001)

2001
[13]

Fernando and S

C. Fernando and S. Sojakka, Pattern recognition in a bucket, inAdvances in Artificial Life (ECAL 2003), Lec- ture Notes in Computer Science, Vol. 2801 (Springer Berlin Heidelberg, 2003) p. 588

2003
[14]

C. Du, F. Cai, M. A. Zidan, W. Ma, S. H. Lee, and W. D. Lu, Reservoir computing using dynamic memris- tors for temporal information processing, Nature Com- munications8, 2204 (2017)

2017
[15]

J. Moon, W. Ma, J. H. Shin, F. Cai, C. Du, S. H. Lee, and W. D. Lu, Temporal data classification and forecast- ing using a memristor-based reservoir computing system, Nature Electronics2, 480 (2019)

2019
[16]

Torrejon, M

J. Torrejon, M. Riou, F. A. Araujo, S. Tsunegi, G. Khalsa, D. Querlioz, P. Bortolotti, V. Cros, K. Yakushiji, A. Fukushima, H. Kubota, S. Yuasa, M. D. Stiles, and J. Grollier, Neuromorphic computing with nanoscale spintronic oscillators, Nature547, 428 (2017)

2017
[17]

Prychynenko, M

D. Prychynenko, M. Sitte, K. Litzius, B. Krüger, G. Bourianoff, M. Kläui, J. Sinova, and K. Everschor- Sitte,Magneticskyrmionasanonlinearresistiveelement: A potential building block for reservoir computing, Phys- ical Review Applied9, 014034 (2018)

2018
[18]

Yokouchi, S

T. Yokouchi, S. Sugimoto, B. Rana, S. Seki, N. Ogawa, Y. Shiomi, S. Kasai, and Y. Otani, Pattern recog- nition with neuromorphic computing using magnetic field–induced dynamics of skyrmions, Science Advances 8, eabq5652 (2022)

2022
[19]

Msiska, J

R. Msiska, J. Love, J. Mulkers, J. Leliaert, and K. Everschor-Sitte, Audio classification with skyrmion reservoirs, Advanced Intelligent Systems5, 2200388 (2023)

2023
[20]

K. Raab, M. A. Brems, G. Beneke, T. Dohi, J. Rothörl, F. Kammerbauer, J. H. Mentink, and M. Kläui, Brown- ian reservoir computing realized using geometrically con- fined skyrmion dynamics, Nature Communications13, 6982 (2022)

2022
[21]

Appeltant, M

L. Appeltant, M. C. Soriano, G. Van Der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso, and I. Fischer, Information processing using a single dynamical node as complex system, Nature Com- munications2, 468 (2011)

2011
[22]

M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, and J. Sun, Emerging opportunities and challenges for the future of reservoir computing, Nature Communications 15, 2056 (2024)
[23]

Liang, J

X. Liang, J. Tang, Y. Zhong, B. Gao, H. Qian, and H. Wu, Physical reservoir computing with emerging elec- tronics, Nature Electronics7, 193 (2024)

2024
[24]

O. Lee, R. Msiska, M. A. Brems, M. Kläui, H. Kure- bayashi, and K. Everschor-Sitte, Perspective on uncon- ventional computing using magnetic skyrmions, Applied Physics Letters122, 260501 (2023)

2023
[25]

Pinna, G

D. Pinna, G. Bourianoff, and K. Everschor-Sitte, Reser- voir computing with random skyrmion textures, Physical Review Applied14, 054020 (2020)

2020
[26]

Everschor-Sitte, A

K. Everschor-Sitte, A. Majumdar, K. Wolk, and D. Meier, Topological magnetic and ferroelectric systems for reservoir computing, Nature Reviews Physics6, 455 (2024)

2024
[27]

B. B. Van Aken, T. T. Palstra, A. Filippetti, and N. A. Spaldin, The origin of ferroelectricity in magnetoelectric YMnO3, Nature Materials3, 164 (2004)

2004
[28]

T. Choi, Y. Horibe, H. T. Yi, Y. J. Choi, W. Wu, and S.- W.Cheong,Insulatinginterlockedferroelectricandstruc- tural antiphase domain walls in multiferroic YMnO3, Na- ture Materials9, 253 (2010)

2010
[29]

V. Wood, A. Austin, E. Collings, and K. Brog, Magnetic properties of heavy-rare-earth orthomanganites, Journal of Physics and Chemistry of Solids34, 859 (1973)

1973
[30]

Fiebig, T

M. Fiebig, T. Lottermoser, D. Fröhlich, A. V. Goltsev, and R. V. Pisarev, Observation of coupled magnetic and electric domains, Nature419, 818 (2002)

2002
[31]

S. H. Skjærvø, E. T. Wefring, S. K. Nesdal, N. H. Gaukås, G. H. Olsen, J. Glaum, T. Tybell, and S. M. Selbach, Interstitial oxygen as a source of p-type conductivity in hexagonal manganites, Nature Communications7, 13745 (2016)

2016
[32]

T. S. Holstad, D. M. Evans, A. Ruff, D. R. Småbråten, J. Schaab, C. Tzschaschel, Z. Yan, E. Bourret, S. M. Sel- bach, S. Krohns, and D. Meier, Electronic bulk and do- main wall properties inB-site doped hexagonal ErMnO3, Physical Review B97, 085143 (2018)

2018
[33]

J. He, U. Ludacka, K. A. Hunnestad, D. R. Småbråten, K. Shapovalov, P. E. Vullum, C. Hatzoglou, D. M. Evans, E. D. Roede, Z. Yan, E. Bourret, S. M. Selbach, D. Gao, J. Akola, and D. Meier, Local p- and n-type doping of an oxide semiconductor via electric-field-driven defect mi...

2025
[34]

D. M. Evans, T. S. Holstad, A. B. Mosberg, D. R. Småbråten, P. E. Vullum, A. L. Dadlani, K. Shapovalov, Z. Yan, E. Bourret, D. Gao, J. Akola, J. Torgersen, A. T. J. Van Helvoort, S. M. Selbach, and D. Meier, Conductivity control via minimally invasive anti-frenkel defects in a...

2020
[35]

Schaab, S

J. Schaab, S. H. Skjærvø, S. Krohns, X. Dai, M. E. Holtz, A. Cano, M. Lilienblum, Z. Yan, E. Bourret, D. A. Muller, M. Fiebig, S. M. Selbach, and D. Meier, Electri- cal half-wave rectification at ferroelectric domain walls, Nature Nanotechnology13, 1028 (2018)

2018
[36]

Y. Geng, N. Lee, Y. J. Choi, S.-W. Cheong, and W. Wu, Collective magnetism at multiferroic vortex do- main walls, Nano Letters12, 6055 (2012)

2012
[37]

J. A. Mundy, J. Schaab, Y. Kumagai, A. Cano, M. Sten- gel, I. P. Krug, D. M. Gottlob, H. Doğanay, M. E. Holtz, R. Held, Z. Yan, E. Bourret, C. M. Schneider, D. G. Schlom, D. A. Muller, R. Ramesh, N. A. Spaldin, and D. Meier, Functional electronic inversion layers at ferro- ele...

2017
[38]

Sheng, I

Y. Sheng, I. Fina, M. Gospodinov, and J. Fontcuberta, Switchable photovoltaic response in hexagonal LuMnO3 single crystals, Applied Physics Letters118, 232902 (2021)

2021
[39]

W. S. Choi, D. G. Kim, S. S. A. Seo, S. J. Moon, D. Lee, J. H. Lee, H. S. Lee, D.-Y. Cho, Y. S. Lee, P. Murugavel, 11 J. Yu, and T. W. Noh, Electronic structures of hexago- nalRMnO 3 (R= Gd, Tb, Dy, and Ho) thin films: Opti- cal spectroscopy and first-principles calculations, ...

2008
[40]

Huang, T

X. Huang, T. R. Paudel, S. Dong, and E. Y. Tsym- bal, Hexagonal rare-earth manganites as promising pho- tovoltaics and light polarizers, Physical Review B92, 125201 (2015)

2015
[41]

Bourret, Growth of high-quality hexagonal ErMnO3 single crystals by the pressurized floating-zone method, Journal of Crystal Growth409, 75 (2015)

Z.Yan, D.Meier, J.Schaab, R.Ramesh, E.Samulon,and E. Bourret, Growth of high-quality hexagonal ErMnO3 single crystals by the pressurized floating-zone method, Journal of Crystal Growth409, 75 (2015)

2015
[42]

L. Chen, G. Zheng, G. Yao, P. Zhang, S. Dai, Y. Jiang, H. Li, B. Yu, H. Ni, and S. Wei, Lead-free perovskite narrow-bandgap oxide semiconductors of rare-earth man- ganates, ACS Omega5, 8766 (2020)

2020
[43]

D. R. Småbråten, F. H. Danmo, N. H. Gaukås, S. P. Singh, N. Kanas, D. Meier, K. Wiik, M.-A. Einarsrud, and S. M. Selbach, Controlling electronic properties of hexagonal manganites through aliovalent doping and thermoatmospheric history, Physical Review Materials9, 024408 (2025)

2025
[44]

D. Y. Guo, Z. P. Wu, Y. H. An, X. C. Guo, X. L. Chu, C. L. Sun, L. H. Li, P. G. Li, and W. H. Tang, Oxygen vacancy tuned ohmic–schottky conversion for enhanced performance inβ-Ga 2O3 solar-blind ultraviolet photode- tectors, Applied Physics Letters105, 023507 (2014)

2014
[45]

Y. M. Chong, A. Majumdar, M. Zahn, I. Hansen, K. Everschor-Sitte, and D. Meier, Data and code for: Photo-inducedcurrentsandshort-termmemoryforreser- voir computing in a ferroelectric semiconductor, Zenodo (2026), version 1

2026
[46]

Puntigam, J

L. Puntigam, J. Schultheiß, A. Strinic, Z. Yan, E. Bour- ret, M. Altthaler, I. Kézsmárki, D. M. Evans, D. Meier, and S. Krohns, Insulating improper ferroelectric domain walls as robust barrier layer capacitors, Journal of Ap- plied Physics129, 074101 (2021)

2021
[47]

A. K. Jonscher, The universal dielectric response, Nature 267, 673 (1977)

1977

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

the bulk material. While the intrinsic dielectric constantε bulk creates a capacitive contribution, the real part resistance consists of a frequency indepen- dent partσ dc and a so-called universal dielectric response (UDR) 44 with power-law scaling with re- spect to frequencyω,σ UDR ∝ω ν,0< ν <1

[2] [2]

the contacts between electrodes and bulk material, generating via a Maxwell-Wagner relaxation a RC- circuit contribution characterized byσ contact and εcontact

[3] [3]

Based on these considerations, we derive a modified circuit representing the photocurrent scenario

the ferroelectric domain walls, acting as additional capacitive barrier layers, characterized byσDW and εDW. Based on these considerations, we derive a modified circuit representing the photocurrent scenario. It is de- veloped by intention pictorially and can be abstracted and in principle treated more rigorously by signal pro- cessing methods: •In the fi...

[4] [4]

Lukoševičius and H

M. Lukoševičius and H. Jaeger, Reservoir computing ap- proaches to recurrent neural network training, Computer Science Review3, 127 (2009)

2009

[5] [5]

Lukoševičius, H

M. Lukoševičius, H. Jaeger, and B. Schrauwen, Reservoir computing trends, KI - Künstliche Intelligenz26, 365 (2012)

2012

[6] [6]

Verstraeten, B

D. Verstraeten, B. Schrauwen, D. Stroobandt, and J. Van Campenhout, Isolated word recognition with the liquid state machine: a case study, Information Processing Let- ters95, 521 (2005)

2005

[7] [7]

Jaeger and H

H. Jaeger and H. Haas, Harnessing nonlinearity: Predict- ing chaotic systems and saving energy in wireless com- munication, Science304, 78 (2004)

2004

[8] [8]

Pathak, B

J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott, Model- free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach, Physical Re- view Letters120, 024102 (2018)

2018