acceptodds
Under review as a conference paper at ICLR 2027

ALDTwin – A Hybrid Physics-Based and Data-Driven Digital Twin of Atomic Layer Deposition Process for Semiconductor Innovation and Manufacturing

Abstract

Atomic layer deposition (ALD) is essential for conformal thin-film growth in advanced semiconductor fabrication, where equipment monitoring and maintenance are critical for atomic-scale precision and process consistency. However, in shared cleanroom environment, monitoring equipment health and predicting process outcomes are challenging because the same tools support diverse recipes, users, materials and operating conditions, while reliable fault and maintenance labels are scarce. We present ALDTwin, a hybrid physics-based and data-driven digital twin with an LLM agent-coordinated workflow for monitoring ALD processes under weakly labeled conditions. The methodology distinguishes intentional recipe-induced chamber conditioning effects from unintended equipment-state changes, including drift, contamination and malfunctions. We demonstrate ALDTwin using sensor data from a shared lab-scale cleanroom ALD system. The LLM agent coordinates data preparation and qualification before invoking simulation and anomaly models. Validated sensor logs are transformed into process-relevant features, after which an authoritative monitoring layer evaluates tool state, run completeness and data quality. Forty-four process metrics are compared with adaptive recipe-specific baselines to suppress expected recipe-dependent variation and identify equipment-state deviations. Qualified runs are passed to a physics-based feature scale simulation for estimating deposition outcomes, while incomplete or abnormal runs bypass simulation but remain part of longitudinal tool-state assessment, with abnormal runs escalated for engineer inspection. In parallel, one LSTM autoencoder detects abrupt deviations within multivariate sensor traces, while a second identifies gradual equipment drift across successive runs using process-relevant metrics. Finalized metric-based anomaly scores are incorporated into a Kalman forecasting layer to project near-term tool behavior. By separating probabilistic LLM coordination from deterministic qualification and physics-based evaluation, ALDTwin integrates safeguards, unsupervised learning, simulation and temporal forecasting within an interpretable, grounded, self-verifying framework for trustworthy process monitoring and decision support.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Atomic Layer Deposition, Semiconductor Manufacturing, Digital Twin, Process Monitoring & Control, Physics-Based Simulations, Agentic AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GNJQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ALDTwin integrates recipe-specific LSTM autoencoders (raw-sensor and metric-based), Kalman drift forecasting, a calibrated physics-based ALD kinetic simulator, and an LLM orchestration layer for post-run monitoring of shared lab-scale ALD tools. Evaluated on 2021–2025 data across two recipes (HfO₂, Al₂O₃) on one tool.

Critical strengths

Recipe-aware design cleanly separates recipe-induced variation from equipment drift; deterministic qualification layer correctly retains authority over learned components. Physics simulation (surface-coverage ODE + Hertz–Knudsen flux, per-chemistry calibration) is rigorous and well-described. Limitations are honestly acknowledged; integrated case studies (Appendix I) make operational behavior concrete.

Critical weaknesses

No quantitative baselines or ablations: impossible to assess whether LSTM/Kalman complexity beats simpler detectors (one-class SVM, control charts) or how much each component contributes. Evaluation scale too small: Kalman assessed on 5 held-out runs; anomaly validation rests on one informal retrospective correlation with no fault labels. LLM role is underjustified: the described function (formatting rule-based decisions into a report) could be a template; no evaluation of whether LLM reports improve experimenter decisions. No uncertainty quantification on physics-simulation outputs despite their use in downstream fabrication decisions.

Questions for authors

What specific capability does the LLM add beyond deterministic formatting, and how is that evaluated? How sensitive are simulated film thickness estimates to calibration uncertainty in the kinetic parameters? The metric-based drift score finalizes after nine subsequent same-recipe runs—what is the practical latency for infrequently used recipes?

7 comments
AuthorsOpenReview ↗

Response to the Reviewer

We thank the reviewer for recognizing the strengths of the recipe-aware design, authoritative deterministic qualification, physics-based ALD simulation and integrated operational case studies. We appreciate the specific concerns regarding component-level validation, the role of the LLM, simulation uncertainty and practical monitoring latency and address them below as follows:

AuthorsOpenReview ↗

Quantitative Baselines and Ablations

We agree that quantitative comparisons against established baseline methods are important for demonstrating the effectiveness of the individual computational components and justifying their integration within ALDTwin.

To address this concern, we have expanded the evaluation to include systematic comparisons across all three data-driven modeling components: (i) raw sensor-based anomaly detection, (ii) metric-based tool drift assessment and (iii) time series forecasting. For the raw sensor and metric-based models, the proposed LSTM autoencoders are compared against appropriate conventional unsupervised anomaly detection and statistical monitoring approaches. For time series prediction, the implemented local level Kalman forecaster is compared against simpler persistence-based forecasting and alternative Kalman formulations.

These comparisons are designed around consistent recipe-specific datasets, chronological evaluation protocols and the corresponding objectives of each modeling task. The resulting analyses are now incorporated into the revised manuscript [sections 5.1, 5.2 in the main text and sections E.4 and F.1 in the Appendix], thus providing a more comprehensive assessment of the contribution of each component.

AuthorsOpenReview ↗

Role of LLM-based Reporting and Future Agentic Integration

In the present implementation, the LLM agent primarily serves as the user facing coordination and reporting layer, integrating outputs from deterministic qualification, unsupervised anomaly detection, Kalman forecasting and physics-based simulation into interpretable reports for experimentalists. Importantly, all numerical computations and operational decisions remain governed by their respective computational modules, ensuring that language generation cannot override authoritative equipment state decisions or safety constraints.

However, the present reporting functionality represents an initial stage of a broader research direction involving interconnected scientific agents for autonomous semiconductor fabrication process development. As part of this ongoing work, we are developing dedicated LLM-coordinated characterization agents for X-ray photoelectron spectroscopy (XPS), Raman spectroscopy, photoluminescence (PL) and ellipsometry. These agents use domain-specific scientific workflows and validated numerical algorithms to automate measurement data processing, calibration, spectral interpretation and extraction of physically meaningful material properties.

A natural extension of ALDTwin is to integrate its process monitoring and physics-based simulation capabilities with these autonomous characterization agents. Such integration would enable process conditions and equipment-state information to be connected directly with experimentally measured film thickness, chemical composition, oxidation states, interfacial properties and other relevant material characteristics. An overarching orchestration agent could combine these complementary sources of information to recommend subsequent process adjustments based on measured outcomes, thereby progressing toward an autonomous, characterization-guided closed-loop fabrication process recipe-development framework.

Reviewer_egDgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a digital twin framework for atomic layer deposition process. A framework employing model-based filtering and LSTM-based monitoring stages with an LLM orchestrator assessing the process condition is presented. The method is showcased on two datasets identifying trends and process conditions. The paper discusses outlook on generalizing to different implementation cases and multi system conditions.

Critical strengths

The paper proposes a comprehensive strategy for handling the monitoring and prediction of machine health for an ALD system. From the presented results, the dataset seems to span a long period and is comprehensive enough to show interesting phenomena, which is not easy to obtain for complex manufacturing systems. The paper presents an interesting framework for machine monitoring and would fit better in a different venue, but the fit to the ICLR topics and contributions is not clear.

Critical weaknesses

The statement about "Current equipment monitoring models identify abnormal tool behavior, while digital twins optimize processes & predict deposition outcomes." seems like a simplified view of existing works. There are a number of works looking at the combination of these topics for industrial digital twins. If the authors claim this gap for ALD related works, this should be clarified, but this still points to a very narrow literature gap and is not suitable for the ICLR scope.

The authors should compare their work against the vast literature on prognostic health monitoring and physics-based modeling for anomaly scores and drift detection. This is common in industrial systems, and the novelty of the paper beyond employing an LLM in the framework is not clear. The authors should review recent papers that employ LLMs with such process monitoring and anomaly detection methods.

The generality of the method is not clear. The presented framework has lots of small design decisions specific to the dataset considered. A general method to be applied to a new process or a similar process is not clear, which greatly limits the applicability and reproducibility of the work.

The results are not compared against simpler baselines, and the use of complex tools such as LLMs in the architecture is not well justified.

Questions for authors
  • The authors mention process monitoring and control but the control aspect of the framework is not clear as no control action is taken. Please clarify and correct as necessary.
    • Are the runs following the same recipe? Are the steps of the run hardcoded or learned from the incoming data? Please clarify.
    • Can you argue about the generality of the method for new materials and recipes on the machine? Do you employ any method to model the relationship of operations to model the process?
    • How did you design the parameters and the nine runs for the tool drift score? Do you employ simple linear models or track the baselines to capture the drifts in the process?
    • Please provide details of the simulator used in the Kalman filter. Is this a linear model similar to the one in the appendix, if so which states are estimated from which outputs and how was the filter validated? Did you validate this model or tune its parameters based on measured signals? How do you calibrate the noise covariances for the estimator?
    • The filter and the dynamics used for simulation are valid only for the baseline model you assume. Can the framework deal with parametric and multiplicative drifts or would you need to employ different filtering methods in this case?
    • Please comment on the validity of the noise model assumption over the actual process parameters.
    • Can you comment on the time scale of the 100 tool drift scores for reporting to the engineer? Did you decide this based on historical data? Is it possible that this is too early or late for certain types of anomalies and drifts?
    • Have you compared the framework components to any baselines, including linear estimators to understand the baseline performance and expected improvement with the proposed architecture?
15 comments
AuthorsOpenReview ↗

Response to the Reviewer

We thank the reviewer for their detailed evaluation of our work and for highlighting important concerns regarding methodological novelty, generalizability, model assumptions, baseline comparisons and the practical interpretation of process monitoring and control. We appreciate these observations and address the critical weaknesses and individual technical questions below.

AuthorsOpenReview ↗

Research Gap, Methodological Novelty and Relevance

A fundamental distinction between much of the existing literature on semiconductor equipment monitoring and our work lies in the operating environment being considered. Many existing studies focus on industrial high volume manufacturing systems, where production processes are relatively standardized, established recipes are repeatedly executed and equipment monitoring benefits from comparatively consistent operating conditions and more structured process and maintenance records. In contrast, ALDTwin specifically addresses shared research lab fabrication environments, where the same equipment serves multiple users conducting experiments with heterogeneous recipes, materials and operating conditions. Furthermore, equipment utilization is often irregular, while reliable fault, failure and maintenance labels are generally unavailable in historical datasets. These characteristics make it particularly challenging to distinguish intentional recipe-induced variations from unintended changes in equipment condition.

We agree that semiconductor equipment health monitoring, anomaly detection, process modeling and digital twins have individually been studied extensively. Our intention is not to suggest that these capabilities are absent from existing industrial manufacturing systems or that ALDTwin introduces a fundamentally new machine learning architecture. Rather, our work addresses the practical challenge of integrating these capabilities into a unified monitoring methodology suitable for heterogeneous, weakly labeled research environments. The primary contribution of ALDTwin is a systematic, recipe-aware and auditable framework that integrates three interconnected aspects of semiconductor equipment monitoring: (i) unsupervised learning from historical equipment data without requiring labeled failure events, (ii) near-real-time post-run monitoring and operational decision support, and (iii) time series forecasting of evolving equipment behavior for preventive maintenance decision support.

Unlike approaches that address individual monitoring or prediction tasks separately, ALDTwin connects complementary evidence from multivariate sensor measurements, physically interpretable process metrics, deterministic run qualification, learned anomaly indicators, temporal forecasting and physics-based deposition simulation within a cohesive computational workflow. Importantly, this integration preserves explicit operational decision authority, model version consistency and human oversight, thus ensuring that learned models and LLM-generated interpretations provide supporting evidence without overriding critical safety decisions.

Our contribution therefore lies primarily in developing and demonstrating an integrated, physically grounded and operationally relevant methodology for equipment monitoring under heterogeneous and weakly labeled conditions, rather than introducing new theoretical results concerning the individual computational algorithms. We have already clarified this distinction in our manuscript and have also now include some more details in the Related Work section to strengthen the positioning of ALDTwin relative to existing semiconductor equipment monitoring and digital twin approaches.

AuthorsOpenReview ↗

Comparison with Existing Prognostic Health Monitoring, Physics-Based Modeling and LLM-Enabled Frameworks

We appreciate the recommendation to position our framework more comprehensively within the established literature on prognostic health monitoring and physics-based process modeling.

As already highlighted in the response for the previous comment, ALDTwin differs principally in its operational integration of recipe-conditioned unsupervised monitoring, authoritative equipment qualification, physics-based deposition estimation and finalized score forecasting within a single auditable workflow designed for weakly labeled research-lab fabrication type of environments.

Regarding LLM integration, we acknowledge that language model-based monitoring, manufacturing orchestration and fault interpretation are increasingly active research directions. In the current implementation, the LLM agent primarily serves as the user facing coordination and reporting layer, integrating outputs from deterministic qualification, unsupervised anomaly detection, Kalman forecasting and physics-based simulation into interpretable reports for experimentalists. Importantly, all numerical computations and operational decisions remain governed by their respective computational modules, ensuring that language generation cannot override authoritative equipment state decisions or safety constraints.

However, the present reporting functionality represents an initial stage of a broader research direction involving interconnected scientific agents for autonomous semiconductor fabrication process development. As part of this ongoing work, we are developing dedicated LLM-coordinated characterization agents for X-ray photoelectron spectroscopy (XPS), Raman spectroscopy, photoluminescence (PL) and ellipsometry. These agents use domain-specific scientific workflows and validated numerical algorithms to automate measurement data processing, calibration, spectral interpretation and extraction of physically meaningful material properties.

A natural extension of ALDTwin is to integrate its process monitoring and physics-based simulation capabilities with these autonomous characterization agents. Such integration would enable process conditions and equipment-state information to be connected directly with experimentally measured film thickness, chemical composition, oxidation states, interfacial properties and other relevant material characteristics. An overarching orchestration agent could combine these complementary sources of information to recommend subsequent process adjustments based on measured outcomes, thereby progressing toward an autonomous, characterization-guided closed-loop fabrication process recipe-development framework.

We have included a few more details in our paper in order to strengthen the Related Work section (Section 2) to establish these distinctions more clearly and incorporate relevant developments in LLM-assisted industrial monitoring and scientific orchestration.

Reviewer_zQkjOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

A physical and data-driven fusion digital twin model for near-real-time post-run monitoring of a shared, label-sparse lab-scale ALD. It integrates deterministic run qualification, recipe-specific raw-sensor and metric-based LSTM autoencoders, robust Kalman forecasting of tool drift, experimentally calibrated physics-based simulation, and LLM-agent-coordinated reporting with human oversight.

Critical strengths

Focused on a real and practical problem. Provide human verifiable report. Integrated multiple data/modeling domain.

Critical weaknesses
  1. Very limited generalization investigation, only two receipe.
  2. Very limited original contribution, below NIPS average.
  3. Rough analysis of failure rate,
Questions for authors

What are the systematic contriubtion especailly for the theoratical contribution? More elaboration of the application of proposed method to other processing equipments.

5 comments
AuthorsOpenReview ↗

Response to the Reviewer

We thank the reviewer for recognizing the practical relevance of the problem, the human-verifiable reporting capabilities and the integration of multiple modeling domains. We appreciate the concerns regarding methodological novelty, generalization and quantitative failure analysis and address them below as follows:

AuthorsOpenReview ↗

Systematic and Theoretical Contributions

We agree that the individual components employed in ALDTwin, including LSTM autoencoders, Kalman filtering and physics-based ALD simulation, are established methods but all isolated from one another in existing literature. Our intended contribution is not a new standalone neural architecture or filtering algorithm but rather a systematic methodology for integrating complementary computational models into a trustworthy, stateful equipment-monitoring digital twin under weakly labeled operating conditions.

A central contribution of ALDTwin is the unified treatment of three fundamental and interconnected aspects of semiconductor equipment health monitoring, as illustrated in Figure 1 of our paper:

  • Unsupervised learning from historical equipment data: In the absence of reliable fault and maintenance labels, recipe-specific LSTM autoencoders learn characteristic equipment behavior directly from historical sensor measurements. The complementary raw sensor and metric-based representations capture within-run abnormalities and longer term changes in equipment state without requiring supervised fault annotations.
  • Near-real time process monitoring and operational decision support: Each newly completed process run undergoes sensor synchronization, process metric extraction and authoritative deterministic qualification against adaptive recipe-specific operating baselines. Qualified runs proceed to calibrated physics-based deposition simulation while abnormal runs trigger engineer inspection. Frozen LSTM models provide complementary anomaly evidence without overriding safety-critical decisions.
  • Time series forecasting for preventive maintenance decision support: Finalized metric-based LSTM drift scores are incorporated into a robust Kalman forecasting framework to estimate the near-term evolution of equipment behavior. This provides forward looking information about changes in the learned operating baseline, supporting proactive equipment health assessment without claiming to predict specific maintenance events or remaining useful life.

The methodological contribution lies in connecting these three functions through a unified, recipe-aware and auditable architecture, rather than treating historical learning, incoming run monitoring and future state prediction as independent tasks. This integration is further strengthened through three important design principles:

  • Safety-constrained qualification and simulation: An authoritative, recipe-specific deterministic monitoring layer governs simulation eligibility and equipment-state transitions, while learned anomaly scores and LLM-generated interpretations remain supporting evidence.
  • Dual timescale, temporally consistent monitoring: Raw sensor reconstruction scores characterize individual run behavior while overlapping metric-based reconstruction windows characterize longer term equipment state changes. Metric-based scores become immutable only after sufficient subsequent run context is available and only finalized scores enter the forecasting model.
  • Physically grounded and auditable integration: Process-relevant sensor measurements are connected to a calibrated surface-kinetic simulation, while the agent coordinates numerical modules, maintains traceability and provides interpretable reports without overriding authoritative operational decisions.

We have already clarified these contributions and their mathematical formulation in the manuscript. We also acknowledge that the work primarily contributes a methodological and systems-level physics-based framework rather than a new theoretical result concerning the underlying learning algorithms.

AuthorsOpenReview ↗

Generalization beyond the two evaluated recipes

We agree that the present empirical evaluation is limited to two thermal ALD recipes on a single equipment platform. The two recipes nevertheless provide distinct chemistry-dependent process responses, allowing our framework to demonstrate recipe-conditioned monitoring without treating expected differences between process chemistries as equipment abnormalities.

The broader applicability of ALDTwin follows from its modular structure. Sensor acquisition, preprocessing, recipe-specific feature extraction, deterministic qualification, unsupervised monitoring, forecasting, physics-based prediction and agent-coordinated reporting are separated into replaceable functional modules.

For other semiconductor processes, such as reactive ion etching (RIE), physical vapor deposition (PVD) or chemical vapor deposition (CVD), the monitoring and orchestration architecture can be retained, while process-dependent sensors, physical descriptors, simulation models and calibration parameters must be adapted and validated.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.