Utilization of AI in Semiconductor Manufacturing Processes

Introduction

Semiconductor manufacturing is one of the fundamental industries supporting the global economy, and its processes require extremely precise and complex control. As device miniaturization and multilayer structures continue to advance, multiple processes and pieces of equipment are involved even for a single wafer, and small process variations can directly affect yield and quality. In recent years, to address this increasing complexity, efforts toward “smart manufacturing” using artificial intelligence (AI) based on large volumes of data obtained from manufacturing processes and equipment have been accelerating.

One important source of information in managing semiconductor manufacturing processes is image data. For example, systems have been reported that detect process abnormalities at an early stage by classifying images acquired during manufacturing into predefined defect types.¹ In contrast, this article focuses primarily on one-dimensional signals obtained from sensors.

In semiconductor manufacturing, a wide variety of data—such as temperature, pressure, flow rate, and rotational speed—is recorded as time series data at high sampling rates. Such data often exhibits complex structures characterized by high dimensionality, nonlinearity, and noise. As a result, it can be difficult to understand the overall behavior using conventional statistical methods alone, and AI has attracted attention as an effective means to automatically extract latent patterns and correlations in the data (Figure 1).

Overview of AI Technologies (Machine Learning and Deep Learning)

AI refers to mechanisms that learn from data to make decisions and predictions. One of its core technologies is machine learning (Figure 2). Machine learning is a methodology for automatically learning patterns within data and constructing models that can make appropriate decisions when faced with new data.² It can be broadly classified into supervised learning, which predicts outputs for unknown inputs, and unsupervised learning, which extracts structures and patterns inherent in the data. Supervised learning can be further divided into classification, where the output is a discrete category, and regression, where the output is a continuous value.

Model performance often improves when more representative training data is available. Figure 3 illustrates representative machine learning algorithms. In many cases, there is a trade-off between predictive performance and model interpretability.³ Historically, creating appropriate models required extensive expertise in statistics and related fields; however, easy-to-use, low-code applications have recently emerged.

Application of AI to Sensor Data

This section presents three methods for applying AI to sensor data in semiconductor manufacturing: etching-profile estimation, unsupervised anomaly detection, and process optimization.

Estimation of Etching Profiles

Semiconductor manufacturing processes can be broadly divided into “front-end processes,” in which circuits are formed on silicon wafers, and “back-end processes,” in which fabricated IC chips are cut out and packaged. Between these processes, cleaning steps are inserted to influence device yield.⁴ These steps remove unwanted materials from the wafer surface generated during front-end processing through chemical etching. In recent years, single-wafer cleaning processes, in which chemicals are supplied while rotating individual wafers, have become mainstream.

In such processes, estimating the radial distribution of the etching rate over a wafer (etching profile) is critical for process optimization and quality control (Figure 4). Etching rates depend heavily on recipe variables, including chemical properties, temperature, liquid flow rate, and wafer rotational speed. Because machine learning can learn relationships in data and make predictions for new inputs, the estimation of etching profiles can be framed as a regression problem by using wafer radius and recipe variables as inputs and etching profiles as outputs (Figure 5).

Figure 6 (top and middle) shows the training and estimation results of etching profiles using the MATLAB low-code Regression Learner app. Four profiles with multiple recipe variables were used as training data, while different profiles were used as test data to verify model generalization. The results indicate that the model was appropriately trained and there are no obvious signs of overfitting. However, model accuracy strongly depends on the choice of algorithms and hyperparameters. For example, optimizing parameters such as minimum leaf size, kernel function, and number of layers can further improve predictive performance. Bayesian optimization is effective for hyperparameter tuning because it can identify optimal values more efficiently than conventional grid search or random search methods. Figure 6 (bottom) shows a tool that automates this process, enabling high-performance models to be obtained with fewer trials.

Unsupervised Anomaly Detection Using Vibration Data

Stable operation of semiconductor manufacturing equipment is essential for maintaining process quality, making equipment condition monitoring indispensable. For example, equipment with rotating components—such as cleaning or etching tools—may develop mechanical abnormalities in motors or bearings that adversely affect productivity and quality. Anomaly detection techniques that use time series signals such as vibration data are effective because they can capture subtle changes embedded in the data that may not be detected by conventional thresholds or statistical process control.⁵

When abnormal events are rare, labeled failure data is difficult to collect. In these cases, unsupervised learning becomes especially valuable because models can be trained using only normal-operation data.

Figure 7 shows three-axis vibration data acquired from an industrial machine before and after maintenance.⁶ In this example, it is assumed that the data before maintenance represents an abnormal state, while the data after maintenance represents a normal state. This setup implies that only normal-state data is used for model training and mixed normal and abnormal data is used for model validation. However, deploying AI does not eliminate the need for data preparation. In general, AI applications require data cleansing (preprocessing) and feature extraction.

Features are numerical representations of certain characteristics of individual data samples⁷ and may be derived from time-domain, frequency-domain, or time-frequency analyses. Selecting appropriate features is challenging even for AI experts. Consequently, exploratory analysis—such as plotting histograms or computing basic statistics such as mean and variance—is often a good starting point.⁸⁻⁹

In Figure 8, 12 features are calculated from pre- and post-maintenance vibration data, two-sample t-tests are applied to evaluate statistically significant differences between them, and then feature importance is ranked. Once highly important features are identified, training a model using those features can be expected to improve accuracy. A representative unsupervised learning method is the autoencoder, which leverages the structure of neural networks.¹⁰ In autoencoders, the reconstruction error between inputs and outputs is used as an anomaly score.

Twelve-dimensional features are extracted from the data, which is randomly split into training and test sets at a 9:1 ratio. The autoencoder is trained using only features from normal data (Figure 9). When the test features are input into the trained model, the reconstruction error (anomaly score) clearly differs between pre- and post-maintenance conditions (Figure 10). That is, although the model was trained only on normal post-maintenance data, it can detect pre-maintenance abnormalities by applying an appropriate threshold.

Reference 11 introduces an alternative unsupervised anomaly detection approach using a one-class support vector machine. However, care must be taken because hyperparameter settings significantly influence performance. Figure 11 illustrates how the kernel scale parameter affects the anomaly region (decision boundary), showing that smaller kernel scales increase nonlinearity.¹³

Finally, beyond anomaly detection, prognostics uses sequential observation data to model the current and future health state of equipment. Once an application-specific failure threshold is defined, the remaining useful life can be estimated (Figure 12). In this case, you can use Bayesian estimation to update model parameters sequentially.¹⁴⁻¹⁵ Interested readers are referred to Reference 16.

Process Optimization

Deep learning can be expressed as an optimization problem in which parameters (weights and biases) of functions constructed from layered combinations of linear transformations and nonlinear activation functions are optimized by minimizing a loss function.¹⁷ Accordingly, this section focuses on the optimization problem.

In semiconductor manufacturing processes, it is often necessary to optimize multiple objectives simultaneously, such as deposition rate and film thickness uniformity.¹⁸ In such cases, optimal input variables must be identified for multiple objective functions; however, an optimal solution for one objective is generally not optimal for another. This competitive relationship implies that improving one objective often worsens another; hence, the goal is to identify the compromise solutions—Pareto-optimal solutions, which form a Pareto front.

A Pareto-optimal solution is defined as one in which improvement in any one objective function necessarily degrades at least one other objective. Figure 13 (inset) illustrates a two-objective optimization problem in which both \(f_₁\) and \(f_₂\) are to be minimized but are in a trade-off relationship. Genetic algorithms are a well-known method for identifying Pareto-optimal solutions.¹⁹ Alternatively, a multiobjective optimization problem can be converted into a single-objective problem by assigning weights to each objective:

\(F(x) = \alpha f_1(x) + (1 - \alpha) f_2(x) \)

By changing the weight \( \alpha \), the trade-off ratio between \(f_₁\) and \(f_₂\) is altered, yielding multiple Pareto solutions. The figure shows that Pareto fronts obtained via genetic algorithms and Bayesian optimization are in good agreement.

Although a two-dimensional multiobjective optimization example with one variable is presented here, the same procedure can be extended to multidimensional objective functions with multiple variables²⁰:

$$ \begin{aligned} \min_{x_1,x_2}\quad & \{f_1(x_1,x_2),\,f_2(x_1,x_2)\} \\[1em] \text{where}\quad & f_1(x_1,x_2)=x_1^{4}+x_2^{4}+x_1x_2-(x_1x_2)^2-10x_1^{2} \\ & f_2(x_1,x_2)=x_1^{4}+x_2^{4}+x_1x_2-(x_1x_2)^2 \\[1em] \text{subject to}\quad & 0\le x_1\le5,\qquad -5\le x_2\le0 \end{aligned} $$

Figure 14 shows how multiobjective optimization problems can be constructed using low-code methods. The GUI-based application called Live Editor Task allows users to try various solvers through mouse operations, while the configured conditions are automatically converted into code, ensuring model transparency.

System Deployment

Finally, consider system integration. The deployment location of the developed algorithms—edge devices, servers, or cloud platforms—depends on data transfer volume, computational requirements, and operational constraints. If minimizing data transfer is critical, necessary preprocessing should be performed on the target equipment, with only essential data transmitted to edge devices for prediction so that you can then forward results to a server and share them with end users.

It is also important to consider how prediction results will be used in operations, such as whether results are displayed on desktop PCs or portable tablets, and whether users possess sufficient domain knowledge. For a detailed deployment case study, refer to Reference 21.

Conclusion

AI is increasingly used in semiconductor manufacturing to detect subtle process variation and equipment anomalies that are difficult to identify using conventional methods. The examples in this article demonstrate how sensor time series data can be used to develop models that support quality improvement and more reliable operations.

Machine learning and deep learning applied to time series data can be used for tasks such as etching-profile estimation and vibration-based monitoring. To get reliable results, focus on collecting representative data, performing careful preprocessing and feature extraction, and using domain expertise to define validation and thresholds. When labeling is difficult, unsupervised methods provide a practical alternative. Continued growth in data availability and AI capabilities is expected to further increase autonomy in process optimization.

This study employed MATLAB, Statistics and Machine Learning Toolbox™, Predictive Maintenance Toolbox™, Deep Learning Toolbox™, Optimization Toolbox™, and Global Optimization Toolbox™. The data used can be downloaded from the MathWorks website (see References 6 and 20).