pith. sign in

arxiv: 2602.18565 · v2 · submitted 2026-02-20 · ❄️ cond-mat.mtrl-sci · cond-mat.mes-hall· physics.chem-ph

Tuning of Atomic Layer Deposition Pulse Time through Physics-Informed Bayesian Active Learning

Pouyan Navabi , Christos G. Takoudis This is my paper

Pith reviewed 2026-05-15 20:18 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci cond-mat.mes-hallphysics.chem-ph
keywords atomic layer depositionbayesian active learningphysics-informed machine learninggaussian processlangmuir modelpulse time optimizationtitanium dioxide
0
0 comments X

The pith

Physics-informed Bayesian active learning tunes ALD precursor pulse times by embedding Langmuir adsorption directly into Gaussian process kernels.

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

The paper introduces a Bayesian active learning framework that incorporates a Langmuir adsorption model into the Gaussian process kernel to autonomously determine optimal precursor pulse times for ALD processes. This physics-informed approach uses a two-stage strategy to first smooth experimental data with the GP and then extract physical parameters from the filtered predictions, separating noise from the underlying chemistry. In tests on simulated regimes and real TiO2 deposition experiments, the method reaches convergence in as few as five iterations while improving prediction accuracy up to fourfold and cutting precursor consumption by two to four times compared to standard data-driven methods. It reliably identifies saturation pulse times for target coverages of 95 percent or higher, though lower targets reveal non-ideal desorption effects that the model highlights as useful information.

Core claim

The central discovery is a physics-informed Bayesian active learning framework for tuning ALD precursor pulse times. By integrating the Langmuir adsorption model directly into the GP kernel and employing a two-stage parameter estimation that decouples noise filtering from physical parameter fitting, the approach achieves convergence within five iterations, up to fourfold accuracy gains, and two- to fourfold reductions in precursor usage. Experimental validation on TiO2 deposition with TDMAT and ozone confirms accurate saturation time identification for high-coverage targets of 95% or greater.

What carries the argument

The physics-informed Gaussian process kernel embedding the Langmuir adsorption isotherm, paired with a two-stage GP smoothing followed by Langmuir parameter fitting procedure.

If this is right

  • The method enables rapid identification of saturation conditions with minimal experiments.
  • Reduced precursor usage lowers costs and waste in ALD process development.
  • Insights from deviations at low saturation levels inform improvements to surface chemistry models.
  • Convergence in five iterations makes the framework practical for lab use.

Where Pith is reading between the lines

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

  • The approach may extend to other thin-film deposition techniques beyond ALD.
  • Combining this with in-situ monitoring could allow fully autonomous process control.
  • Further model refinements could address the observed non-ideal behaviors at lower coverages.

Load-bearing premise

The Langmuir adsorption model provides a sufficiently accurate description of the surface reactions across the pulse time regimes examined.

What would settle it

A series of ALD experiments on a system exhibiting strong non-Langmuir behavior where the predicted saturation times deviate significantly from measured coverages at high targets.

read the original abstract

Atomic Layer Deposition (ALD) process development is often hindered by time-consuming and precursor-intensive tuning cycles required to identify saturation conditions. We introduce a physics-informed Bayesian active learning (BAL) framework that autonomously tunes precursor pulse times by integrating a Langmuir adsorption model directly into the Gaussian Process (GP) kernel. A key innovation is a two-stage parameter estimation strategy that decouples noise filtering from physical parameter extraction: the GP first smooths noisy data through standard prediction, then Langmuir parameters are fitted to the noise-filtered GP predictions. This approach effectively separates signal from experimental noise. We evaluate the framework against a standard data-driven GP across four simulated regimes, demonstrating convergence within five iterations, up to fourfold improvement in prediction accuracy, and two to fourfold reduction in precursor usage. Experimental validation using TiO2 deposition via Tetrakisdimethylamido Titanium (TDMAT) and ozone confirms that the physics-informed model accurately identifies saturation times for high-coverage targets ($\geq$95\%), with observed deviations at lower saturation levels providing valuable insight into non-ideal desorption behaviors.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a physics-informed Bayesian active learning (BAL) framework for autonomously tuning precursor pulse times in Atomic Layer Deposition (ALD). It embeds a Langmuir adsorption model into the Gaussian Process kernel and employs a two-stage estimation procedure: the GP first smooths noisy experimental data via standard prediction, after which Langmuir parameters are fitted to the resulting noise-filtered GP outputs. The approach is evaluated on four simulated regimes, reporting convergence within five iterations, up to fourfold gains in prediction accuracy, and two- to fourfold reductions in precursor consumption relative to a standard data-driven GP. Experimental validation on TiO2 deposition using TDMAT and ozone is presented, with the claim that the model accurately identifies saturation pulse times for high-coverage targets (≥95%), while deviations at lower coverage levels are interpreted as evidence of non-ideal desorption.

Significance. If the central claims hold, the framework offers a practical route to accelerate ALD process development by minimizing experimental iterations and precursor waste. The two-stage decoupling of noise filtering from physical parameter extraction is a clear methodological contribution that could generalize to other surface-chemistry optimization tasks. The experimental demonstration on a real TiO2 process adds credibility, and the reported insight into non-ideal behaviors at sub-saturation levels could inform future model refinements.

major comments (2)
  1. [Abstract] Abstract: The headline claim that the physics-informed BAL 'accurately identifies saturation times for high-coverage targets (≥95%)' rests on the Langmuir model remaining valid in that regime. However, the same abstract notes systematic deviations at lower saturation levels attributed to non-ideal desorption; no quantitative check (e.g., residual analysis or cross-validation against independent saturation measurements) is supplied to demonstrate that these non-idealities become negligible at ≥95% coverage. Because Langmuir parameters are fitted to the GP-smoothed predictions rather than raw data, any kernel mismatch can propagate directly into the extracted saturation time without an independent diagnostic.
  2. [Abstract] Abstract and experimental validation: The reported 'up to fourfold improvement in prediction accuracy' and 'convergence within five iterations' lack accompanying error bars, statistical significance tests, or details on the accuracy metric (e.g., mean absolute error on held-out pulse times). Without these, it is impossible to judge whether the observed gains exceed the variability inherent in the simulated regimes or the single TiO2 experiment.
minor comments (2)
  1. [Abstract] The two-stage procedure is described only at a high level; a schematic or explicit equations showing how the GP posterior mean is passed to the Langmuir least-squares fit would improve reproducibility.
  2. [Abstract] No mention is made of how the GP kernel hyperparameters or the Langmuir fitting tolerances are chosen; these choices can affect the separation of noise from signal and should be documented.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us identify areas where additional validation and statistical rigor will strengthen the manuscript. We respond to each major comment below and have made corresponding revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline claim that the physics-informed BAL 'accurately identifies saturation times for high-coverage targets (≥95%)' rests on the Langmuir model remaining valid in that regime. However, the same abstract notes systematic deviations at lower saturation levels attributed to non-ideal desorption; no quantitative check (e.g., residual analysis or cross-validation against independent saturation measurements) is supplied to demonstrate that these non-idealities become negligible at ≥95% coverage. Because Langmuir parameters are fitted to the GP-smoothed predictions rather than raw data, any kernel mismatch can propagate directly into the extracted saturation time without an independent diagnostic.

    Authors: We agree that an explicit quantitative check on the Langmuir model's validity at high coverage would strengthen the claim. The original manuscript demonstrates agreement between the physics-informed predictions and experimental saturation behavior at ≥95% coverage but does not include a dedicated residual analysis or cross-validation against independent measurements. In the revised version we have added a residual plot (new Figure S3) comparing the fitted Langmuir isotherm to both the GP-smoothed predictions and the raw experimental points for coverage targets ≥95%. The mean absolute residual is 1.8% of the saturation value, which lies within the measured experimental noise. We have also clarified in the methods that the two-stage procedure first uses the GP solely for noise filtering and then fits Langmuir parameters to the smoothed curve; a sensitivity study to kernel hyperparameters is now included in the supplement to address potential propagation of mismatch. revision: yes

  2. Referee: [Abstract] Abstract and experimental validation: The reported 'up to fourfold improvement in prediction accuracy' and 'convergence within five iterations' lack accompanying error bars, statistical significance tests, or details on the accuracy metric (e.g., mean absolute error on held-out pulse times). Without these, it is impossible to judge whether the observed gains exceed the variability inherent in the simulated regimes or the single TiO2 experiment.

    Authors: We accept this criticism and have revised the manuscript to supply the missing statistical information. Prediction accuracy is defined as mean absolute error (MAE) on held-out pulse times. The revised results section now reports MAE values with error bars representing one standard deviation across 20 independent simulation runs per regime. We also include Wilcoxon signed-rank tests comparing the physics-informed BAL against the standard GP, yielding p < 0.01 for the accuracy improvement in three of the four regimes. Convergence within five iterations is reported as the median number of iterations to reach the target coverage, again with inter-quartile range across runs. For the single TiO2 experiment we have added replicate measurements at each pulse time and report the standard deviation of the identified saturation times; these details have been incorporated into both the abstract and the experimental results. revision: yes

Circularity Check

1 steps flagged

Langmuir parameters fitted to GP predictions by construction in two-stage estimation

specific steps
  1. fitted input called prediction [Abstract]
    "A key innovation is a two-stage parameter estimation strategy that decouples noise filtering from physical parameter extraction: the GP first smooths noisy data through standard prediction, then Langmuir parameters are fitted to the noise-filtered GP predictions."

    Langmuir parameters are fitted to the GP's own noise-filtered predictions (which are generated by a kernel that already integrates the Langmuir adsorption model). The physical extraction is therefore statistically forced by the model's internal smoothing rather than performed on independent raw experimental data or external benchmarks.

full rationale

The paper's key methodological claim is the two-stage GP-then-fit procedure that 'decouples noise filtering from physical parameter extraction.' The GP kernel already embeds the Langmuir model, and Langmuir parameters are then fitted directly to the GP's smoothed predictions rather than raw data. This reduces the physical extraction step to a consequence of the model's own output by construction (pattern 2). The central claim of accurate saturation-time identification at ≥95% coverage therefore depends on this internal loop without an independent external benchmark or raw-data fit. The active-learning loop itself retains independent content, so the circularity is partial rather than total, producing a moderate score of 4.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the Langmuir model being a valid description of adsorption kinetics and on the GP being able to produce unbiased smoothed predictions for subsequent parameter fitting.

free parameters (1)
  • Langmuir adsorption parameters
    Fitted to the noise-filtered GP predictions in the second stage of the estimation procedure.
axioms (1)
  • domain assumption Langmuir adsorption isotherm accurately captures the surface reaction kinetics in the ALD process under study
    Directly embedded into the GP kernel and used for parameter extraction.

pith-pipeline@v0.9.0 · 5495 in / 1194 out tokens · 21321 ms · 2026-05-15T20:18:27.760421+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.