Bayesian Optimization of Multi-Bit Pulse Encoding in In2O3/Al2O3 Thin-film Transistors for Temporal Data Processing

Benius Dunn; Javier Meza-Arroyo; Jen-Sue Chen; Julia W.P. Hsu; Weijie Xu; Yu-Chieh Chen

arxiv: 2510.07421 · v1 · submitted 2025-10-08 · ❄️ cond-mat.dis-nn · cond-mat.mtrl-sci· cs.LG

Bayesian Optimization of Multi-Bit Pulse Encoding in In2O3/Al2O3 Thin-film Transistors for Temporal Data Processing

Javier Meza-Arroyo , Benius Dunn , Weijie Xu , Yu-Chieh Chen , Jen-Sue Chen , Julia W.P. Hsu This is my paper

Pith reviewed 2026-05-18 09:16 UTC · model grok-4.3

classification ❄️ cond-mat.dis-nn cond-mat.mtrl-scics.LG

keywords Bayesian optimizationthin-film transistorsreservoir computingtemporal encodingpulse parametersstate separabilityneuromorphic hardwareIn2O3/Al2O3

0 comments

The pith

Bayesian optimization over pulse parameters in In2O3/Al2O3 thin-film transistors enables high-fidelity 6-bit temporal encoding.

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

The authors apply Bayesian optimization to tune five pulse parameters in solution-processed Al2O3/In2O3 thin-film transistors for physical reservoir computing. They use the normalized degree of separation as the metric to improve distinguishability of multi-state outputs for time-series data encoding. A model trained on 4-bit encoding tasks successfully guides optimization for the more complex 6-bit case, which reduces the number of experiments needed. When applied to encoding and reconstructing a sequence of binary-patterned moving car images over six frames, the optimized conditions yield higher accuracy than default settings. Interpretability analysis shows that gate pulse amplitude and drain voltage have the strongest influence on state separation.

Core claim

Bayesian optimization identifies pulse parameter combinations that maximize the normalized degree of separation among six output states in the thin-film transistor. This produces more faithful encoding of temporal data, demonstrated by improved reconstruction accuracy for a moving car image sequence. The optimization model built from simpler 4-bit data transfers effectively to the 6-bit problem.

What carries the argument

Bayesian optimization of five key pulse parameters using the normalized degree of separation (nDoS) metric to enhance output state separability in the TFT-based reservoir.

If this is right

Optimized operating conditions improve the accuracy of encoding and reconstructing binary-patterned images of a moving car across six sequential frames.
A 4-bit trained model can direct the optimization of 6-bit encoding tasks while requiring fewer experimental runs.
Gate pulse amplitude and drain voltage are identified as the most influential parameters for higher state separation.
The approach offers a systematic way to find optimal conditions for reservoir devices instead of empirical selection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Applying this transfer learning from lower to higher bit depths may allow efficient scaling to 8-bit or higher encodings in similar hardware.
Validating the nDoS metric against reconstruction performance in additional temporal tasks, such as sensor data streams, would test its broader utility.
Similar Bayesian optimization strategies could be tested on other physical reservoir materials to identify platform-specific optimal parameters.

Load-bearing premise

The normalized degree of separation metric reliably predicts which pulse parameters will produce output states that stay distinguishable during actual temporal data reconstruction.

What would settle it

Reconstructing the moving car image sequence using the optimized pulse parameters shows no reduction in error compared to using default or randomly chosen parameters.

read the original abstract

Utilizing the intrinsic history-dependence and nonlinearity of hardware, physical reservoir computing is a promising neuromorphic approach to encode time-series data for in-sensor computing. The accuracy of this encoding critically depends on the distinguishability of multi-state outputs, which is often limited by suboptimal and empirically chosen reservoir operation conditions. In this work, we demonstrate a machine learning approach, Bayesian optimization, to improve the encoding fidelity of solution-processed Al2O3/In2O3 thin-film transistors (TFTs). We show high-fidelity 6-bit temporal encoding by exploring five key pulse parameters and using the normalized degree of separation (nDoS) as the metric of output state separability. Additionally, we show that a model trained on simpler 4-bit data can effectively guide optimization of more complex 6-bit encoding tasks, reducing experimental cost. Specifically, for the encoding and reconstruction of binary-patterned images of a moving car across 6 sequential frames, we demonstrate that the encoding is more accurate when operating the TFT using optimized pulse parameters and the 4-bit optimized operating condition performs almost as well as the 6-bit optimized condition. Finally, interpretability analysis via Shapley Additive Explanations (SHAP) reveals that gate pulse amplitude and drain voltage are the most influential parameters in achieving higher state separation. This work presents the first systematic method to identify optimal operating conditions for reservoir devices, and the approach can be extended to other physical reservoir implementations across different material platforms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Bayesian optimization tunes five pulse parameters in these In2O3/Al2O3 TFTs for better 6-bit temporal encoding, and a 4-bit trained model transfers nearly as well to the 6-bit case.

read the letter

This paper's main point is that Bayesian optimization finds better operating conditions for multi-bit pulse encoding in solution-processed In2O3/Al2O3 TFTs used as physical reservoirs, and that a model fit on simpler 4-bit data works almost as well for the 6-bit version. The transfer step is the practical part that could actually save lab time. The work does a clean job of framing a repeatable workflow: search over five pulse parameters with nDoS as the objective, then apply SHAP to show gate amplitude and drain voltage matter most. The moving-car binary image sequence across six frames is used as the concrete test, and the optimized conditions are reported to give more accurate reconstruction than default settings. That demonstration plus the transfer result gives the paper a usable takeaway for experimental groups working on similar oxide devices. The soft spot is the missing numbers. The abstract claims higher fidelity and more accurate encoding but does not supply reconstruction errors, frame-wise accuracies, or direct comparisons with error bars against non-optimized baselines. The stress-test concern about nDoS maximization not automatically guaranteeing reliable temporal distinguishability is reasonable to raise; if the full results section only optimizes on pairwise separation and then asserts task-level gains without showing the link, that part needs tightening. Otherwise the gap looks minor rather than load-bearing. This is aimed at people building physical reservoirs with thin-film transistors or looking for data-driven tuning methods instead of trial-and-error. A reader who wants a concrete procedure they can adapt to their own hardware will get value from it. The paper deserves a serious referee because the core method is straightforward to reproduce and targets a real experimental bottleneck in the subfield. I would send it to peer review with the expectation that the authors add the quantitative task metrics and clarify how well nDoS predicts the actual reconstruction performance.

Referee Report

2 major / 3 minor

Summary. The manuscript applies Bayesian optimization over five pulse parameters in In2O3/Al2O3 thin-film transistors to maximize the normalized degree of separation (nDoS) metric, claiming high-fidelity 6-bit temporal encoding for physical reservoir computing. It further reports that a model trained on 4-bit data can guide 6-bit optimization with comparable performance, demonstrates improved accuracy in reconstructing a moving-car binary image sequence across six frames, and uses SHAP analysis to identify gate pulse amplitude and drain voltage as the most influential parameters.

Significance. If the quantitative links between nDoS maximization and task-level reconstruction accuracy hold, the work supplies a systematic, reproducible route to operating-condition selection that reduces reliance on empirical tuning. The transfer-learning demonstration and SHAP-based interpretability are concrete strengths that could lower experimental overhead and aid generalization to other reservoir materials.

major comments (2)

[Moving-car image sequence reconstruction] Moving-car image sequence reconstruction: the central claim that optimized parameters produce 'more accurate' encoding requires explicit quantitative metrics (pixel error, frame-wise accuracy, or nDoS values with error bars) comparing optimized versus baseline conditions; without these, it is unclear whether nDoS gains translate to reliable distinguishability under sequential readout, drift, and noise.
[Bayesian optimization and transfer learning] 4-bit to 6-bit transfer learning: the statement that the 4-bit optimum 'performs almost as well' as the 6-bit optimum is load-bearing for the cost-reduction claim, yet the manuscript supplies no tabulated performance numbers or details on how the 4-bit surrogate model is applied to the 6-bit search space.

minor comments (3)

[Abstract] The abstract uses 'high-fidelity' without a numerical threshold or achieved nDoS value; a single representative number would improve clarity.
[Methods] The exact definitions of the five pulse parameters and the normalization procedure for nDoS should be stated explicitly in the methods for reproducibility.
[Results figures] Figure captions for the reconstruction examples should include the quantitative error metric used to declare 'more accurate' encoding.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review of our manuscript. We appreciate the positive assessment of the significance of our Bayesian optimization framework and the transfer-learning demonstration. We address the two major comments below and will incorporate the requested quantitative details and clarifications in the revised manuscript.

read point-by-point responses

Referee: Moving-car image sequence reconstruction: the central claim that optimized parameters produce 'more accurate' encoding requires explicit quantitative metrics (pixel error, frame-wise accuracy, or nDoS values with error bars) comparing optimized versus baseline conditions; without these, it is unclear whether nDoS gains translate to reliable distinguishability under sequential readout, drift, and noise.

Authors: We agree that explicit quantitative metrics are necessary to rigorously support the claim of improved reconstruction accuracy. The original manuscript primarily presented visual comparisons of reconstructed frames; we acknowledge this leaves the translation from nDoS to task-level performance open to interpretation. In the revised version we will add a dedicated results subsection and accompanying table that reports (i) mean pixel error (with standard deviation across repeated trials), (ii) frame-wise binary accuracy, and (iii) nDoS values with error bars for the baseline, 4-bit-optimized, and 6-bit-optimized operating conditions. These metrics will be obtained under the same sequential readout protocol used for the moving-car sequence, thereby directly addressing concerns about drift and noise. revision: yes
Referee: 4-bit to 6-bit transfer learning: the statement that the 4-bit optimum 'performs almost as well' as the 6-bit optimum is load-bearing for the cost-reduction claim, yet the manuscript supplies no tabulated performance numbers or details on how the 4-bit surrogate model is applied to the 6-bit search space.

Authors: We thank the referee for identifying this gap. The transfer-learning result is central to our claim of reduced experimental overhead, yet the manuscript indeed lacks tabulated side-by-side metrics and a precise description of the surrogate transfer procedure. In the revision we will (i) expand the Methods section to specify how the 4-bit Gaussian-process surrogate is reused (posterior mean and uncertainty are used to seed the 6-bit acquisition function), (ii) add a table that lists achieved nDoS, number of experimental trials required, and reconstruction accuracy for both the direct 6-bit optimization and the 4-bit-guided optimization, and (iii) include a brief discussion of the observed performance gap (or lack thereof) to substantiate the “almost as well” statement. revision: yes

Circularity Check

0 steps flagged

No circularity: experimental optimization grounded in measurements

full rationale

The paper describes an experimental workflow in which Bayesian optimization is applied to measured nDoS values obtained from physical TFT devices; the resulting parameter sets are then directly tested on a moving-car image-sequence reconstruction task. No equations, first-principles derivations, or statistical predictions are claimed that reduce by construction to fitted constants or self-referential definitions. The 4-bit-to-6-bit transfer is presented as an empirical observation rather than a derived result. The work is therefore self-contained against external benchmarks (device measurements and task accuracy) and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard domain assumptions about the suitability of TFT nonlinearity for reservoir computing and on the validity of nDoS as a proxy for encoding fidelity; no new physical entities are postulated and the pulse parameters are treated as tunable inputs rather than free parameters fitted to the final result.

free parameters (2)

gate pulse amplitude
One of the five key pulse parameters explored by Bayesian optimization; identified by SHAP as highly influential.
drain voltage
One of the five key pulse parameters; SHAP analysis shows strong influence on state separation.

axioms (2)

domain assumption The In2O3/Al2O3 TFT exhibits intrinsic history dependence and nonlinearity that can be harnessed for temporal data encoding in reservoir computing.
Invoked in the opening paragraph as the foundation for physical reservoir computing.
domain assumption Maximizing normalized degree of separation (nDoS) across output states produces encodings that remain distinguishable during downstream reconstruction tasks.
Central to the choice of optimization objective and the claim of high-fidelity 6-bit encoding.

pith-pipeline@v0.9.0 · 5835 in / 1724 out tokens · 34958 ms · 2026-05-18T09:16:54.927355+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

We show high-fidelity 6-bit temporal encoding by exploring five key pulse parameters and using the normalized degree of separation (nDoS) as the metric of output state separability.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_injective unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the encoding is more accurate when operating the TFT using optimized pulse parameters

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.