Machine Learning Enhances Plasma Plume Analysis in PLD

In an article published in the journal Nature, researchers focused on the use of autonomous experimentation in materials synthesis, specifically through pulsed laser deposition (PLD). They investigated how machine learning (ML) techniques can derive valuable insights from in situ diagnostics, specifically using intensified-charge coupled device (ICCD) image sequences of the plasma plume produced during PLD.

Pulsed laser deposition (PLD) growth for each sample is carried out by selecting a chamber pressure, substrate temperature, and laser energy on each PLD target (P, T, E1, E2). During the growth, an ICCD image sequence and in situ laser reflectivity curve are collected. The image sequence is fed into a (2 + 1)D convolutional neural network (CNN) to extract deep features from the plume dynamics. For experimental anomaly detection, P, E1, E2 are predicted from the ICCD features alone. The growth kinetics parameters (s0, s1, J) derived from reflectivity measurements are predicted by either a multilayer perceptron (MLP) using the growth parameters, the (2 + 1)D CNN, or a by combining the features from both in a final MLP. Image Credit: https://www.nature.com/articles/s41524-024-01275-w
Pulsed laser deposition (PLD) growth for each sample is carried out by selecting a chamber pressure, substrate temperature, and laser energy on each PLD target (P, T, E1, E2). During the growth, an ICCD image sequence and in situ laser reflectivity curve are collected. The image sequence is fed into a (2 + 1)D convolutional neural network (CNN) to extract deep features from the plume dynamics. For experimental anomaly detection, P, E1, E2 are predicted from the ICCD features alone. The growth kinetics parameters (s0, s1, J) derived from reflectivity measurements are predicted by either a multilayer perceptron (MLP) using the growth parameters, the (2 + 1)D CNN, or a by combining the features from both in a final MLP. Image Credit: https://www.nature.com/articles/s41524-024-01275-w

The authors developed multi-output two-plus-one-dimensional ((2 + 1)D) convolutional neural network (CNN) regression models to analyze plume dynamics, which correlate with measured chamber pressure and laser energy and predict thin film growth kinetics. The findings suggested that ML can optimize deposition conditions and provide predictive insights into film growth, facilitating rapid pre-screening and materials optimization for non-experts.

Background

The emergence of autonomous synthesis platforms, which feature in situ or automated diagnostics and characterization techniques, requires the creation of ML models to harness this multimodal synthesis data. In situ diagnostics are essential in autonomous materials synthesis for deriving optimization metrics, correlating growth kinetics with experimental controls, predicting material properties prior to synthesis, and identifying anomalies during long-term unsupervised experiments. PLD, an effective physical vapor deposition method, is highly compatible with autonomous synthesis because it supports a variety of diagnostic measurements, such as optical, electrical, and electron-based techniques.

Previous work has demonstrated the use of ML with in situ diagnostics, particularly with reflection high-energy electron diffraction (RHEED) patterns, to predict film properties and forecast RHEED intensity during growth. Other studies have explored ML algorithms to understand RHEED image sequences and construct structural phase maps. However, the application of ML to in situ diagnostics data in PLD remains limited, primarily due to undisciplined or sparse datasets.

This paper aimed to fill these gaps by utilizing an autonomous PLD platform to generate large, disciplined synthesis datasets. The focus is on ICCD imaging of plasma plumes during PLD, which can potentially increase the reproducibility of experiments and predict film characteristics such as growth rate and crystallinity. Despite the promise, analysis of plume image sequences has been a time-consuming offline effort, and correlations between plume dynamics and material properties are not well understood. This study employed a (2 + 1)D CNN to extract features from ICCD images, correlating plume dynamics with film growth kinetics and PLD conditions, thus proving the viability of deep learning for real-time monitoring and prediction in PLD synthesis.

Methodology for Data Generation and ML Integration

The researchers utilized ICCD image sequences in combination with growth parameters and growth kinetics data obtained from an autonomous PLD synthesis campaign. Specifically, 127 tungsten diselenide (WSe2) films were grown on silicon oxide (SiO2)/silicon (Si) substrates by co-ablating WSe2 and selenium (Se) targets. Each film growth involved capturing a sequence of 50 ICCD images, varying chamber pressure (P), substrate temperature (T), and laser energies on the targets. Laser reflectivity measured sub-monolayer nucleation and growth, providing a basis for the growth kinetics model with parameters.

ICCD images, initially 1024×1024 pixels, were processed by Gaussian filtering, resizing to 40×40 pixels, and log-transformed for intensity normalization. This resulted in sequences with dimensions 50×40×40. Laser reflectivity curves were fitted to an auto-catalytic growth model to derive the growth kinetics parameters.

ML models included a (2+1)D CNN for extracting features from ICCD images and a multilayer perceptron (MLP) for predicting growth parameters. The (2+1)D CNN used sequential three-dimensional (3D) convolutions for spatial and temporal data, reducing parameters and increasing nonlinearity. This model, combined with a simple MLP, aimed to predict experimental conditions and growth kinetics parameters. The final mixed input model incorporated both ICCD features and growth parameters into a comprehensive MLP.

Despite a small dataset of 127 samples, data augmentation techniques such as random transformations and Gaussian noise improved model generalization. The dataset was split 70/30 into training and validation sets. Training employed batch gradient descent with the Adam optimizer and mean square error loss function. Hyperparameter tuning, conducted using Ray Tune and Optuna, optimized the model for better performance.

Findings and Analysis

  • Anomaly Detection: In PLD, reproducibility issues often arose due to factors like material buildup on the chamber’s laser window or changes in the PLD target condition. These variations impacted plume dynamics, making deep learning a valuable tool for capturing essential plume features and providing feedback for future experiments. By comparing ICCD image sequences to current conditions, models could detect anomalies in plume dynamics, indicating unreplicated conditions. For rapid deposition sequences, model predictions could help mitigate issues like laser window coating or gas flow control problems. A (2 + 1)D CNN was trained to predict P and laser energies using ICCD image sequences. The model performed well, with P prediction achieving an r-squared (r²) of 0.963. In contrast, a single-image model showed significantly lower accuracy. These results highlighted the effectiveness of incorporating both spatial and temporal plume dynamics in anomaly detection.
  • Prediction of Growth Kinetics: Predicting growth kinetics from plume diagnostics could expedite material discovery and optimization by reducing time and material costs. Various species in a PLD plume affected film stoichiometry, influencing film properties. ICCD imaging could provide valuable information for ML models that guide experimental sequences. Models were trained to predict growth kinetics parameters from laser reflectivity data. The ICCD model showed good predictive power with an r² of 0.815. A growth parameter-only model performed slightly better but was less stable. The combined model, using both ICCD and growth parameters, achieved the best performance with an r² of 0.847.
  • Feature Map Analysis: Visualizing feature maps from the (2 + 1)D CNN revealed that deeper layers encode complex spatiotemporal dynamics of plume expansion and interaction. Saliency maps indicated that the model focused on plume features rather than background noise, reinforcing the importance of plume dynamics in predictions.

Conclusion

In conclusion, the researchers demonstrated the efficacy of using deep learning with ICCD image sequences for real-time monitoring and prediction in PLD synthesis. By employing a (2 + 1)D CNN, researchers successfully captured complex spatiotemporal plume dynamics, correlating them with experimental conditions and growth kinetics.

This approach enabled effective anomaly detection and accurate predictions of film growth parameters, proving valuable for optimizing deposition conditions. The integration of ICCD imaging with ML models offered a promising path for enhancing reproducibility and accelerating material discovery in autonomous PLD systems.

Journal reference:
Soham Nandi

Written by

Soham Nandi

Soham Nandi is a technical writer based in Memari, India. His academic background is in Computer Science Engineering, specializing in Artificial Intelligence and Machine learning. He has extensive experience in Data Analytics, Machine Learning, and Python. He has worked on group projects that required the implementation of Computer Vision, Image Classification, and App Development.

Citations

Please use one of the following formats to cite this article in your essay, paper or report:

  • APA

    Nandi, Soham. (2024, May 21). Machine Learning Enhances Plasma Plume Analysis in PLD. AZoAi. Retrieved on October 08, 2024 from https://www.azoai.com/news/20240521/Machine-Learning-Enhances-Plasma-Plume-Analysis-in-PLD.aspx.

  • MLA

    Nandi, Soham. "Machine Learning Enhances Plasma Plume Analysis in PLD". AZoAi. 08 October 2024. <https://www.azoai.com/news/20240521/Machine-Learning-Enhances-Plasma-Plume-Analysis-in-PLD.aspx>.

  • Chicago

    Nandi, Soham. "Machine Learning Enhances Plasma Plume Analysis in PLD". AZoAi. https://www.azoai.com/news/20240521/Machine-Learning-Enhances-Plasma-Plume-Analysis-in-PLD.aspx. (accessed October 08, 2024).

  • Harvard

    Nandi, Soham. 2024. Machine Learning Enhances Plasma Plume Analysis in PLD. AZoAi, viewed 08 October 2024, https://www.azoai.com/news/20240521/Machine-Learning-Enhances-Plasma-Plume-Analysis-in-PLD.aspx.

Comments

The opinions expressed here are the views of the writer and do not necessarily reflect the views and opinions of AZoAi.
Post a new comment
Post

While we only use edited and approved content for Azthena answers, it may on occasions provide incorrect responses. Please confirm any data provided with the related suppliers or authors. We do not provide medical advice, if you search for medical information you must always consult a medical professional before acting on any information provided.

Your questions, but not your email details will be shared with OpenAI and retained for 30 days in accordance with their privacy principles.

Please do not ask questions that use sensitive or confidential information.

Read the full Terms & Conditions.

You might also like...
Machine Learning Advances Radiation Modeling