language-icon Old Web
English
Sign In

The self-driving photomask

2018 
Machine learning (ML) has become increasingly powerful and several recent works have demonstrated the capa- bility of neural networks to achieve performance gains for lithography applications. Much of the general literature on neural networks involves image classification. Application of neural networks to lithography requires increased scrutiny. How far can such a system be trusted, and how should we respond if the system fails? Neural net- works can appear inscrutable and we lack understanding of why these systems generalize so well. On the other hand, the benefits neural networks appear to offer, in terms of reduced runtime or more accurate models, are compelling. This work will illustrate how two techniques, the Information Bottleneck (IB) and t-Distributed Stochastic Nearest Neighbors (t-SNE), that can improve our understanding of how neural networks work. We will use a multilayer perceptron for a simple resist model implemented with neural networks. We will then discuss how visualiztion methods can help assess the readiness of a neural network for a task, or help diagnose potential causes of failure.
    • Correction
    • Source
    • Cite
    • Save
    • Machine Reading By IdeaReader
    14
    References
    1
    Citations
    NaN
    KQI
    []