LAMP

Motivation

‍Extreme Ultraviolet (EUV) lithography is the current technology for semiconductor device manufacturing, and it allows to uphold Moore’s law in upcoming years. The use of EUV wavelength (13.5 nm) allows a tremendous increase in the resolution, but at the same time, it also increases the cost of the infrastructure for the lithography process. PSI is pioneering a lensless imaging approach - ptychography or coherent diffraction imaging (CDI) - for EUV mask inspection. A dedicated imaging tool, RESCAN, has been built and developed in recent years. With RESCAN, the main challenge is that a very large computational infrastructure is needed to achieve the desired reconstruction speed, which is highly impractical. The objective of the project is to explore alternative data processing approaches to achieve the required reconstruction speed with an affordable hardware infrastructure. The development of a cost-effective framework for mask inspection would have an enormous impact on future technology development in semiconductor device manufacturing.

Proposed Approach / Solution

‍We propose to optimize computational and data storage bottlenecks by designing efficient algorithms for solving the phase retrieval problem in ptychography. Ptychography is an imaging technique that allows the observation of samples at the nanometer resolution by successively probing small overlapping regions in the sample (see Fig. 1). In this project we are focusing on the inspection of semiconductor device-related samples like EUV photomasks. Usually, these samples are very well characterized and we can leverage the abundant amount of prior information in our possession to drive the image reconstruction process.

In this project, we employ deep generative models such as Generative Adversarial Networks (GANs), Denoising Diffusion Probabilistic Models (DDPMs), and Latent Diffusion Models (LDMs) to capture the data distribution of complex objects (photomasks in this case) in order to solve the phase retrieval problem. We consider image-to-image translation approaches, as well as inverse problem formulations coupled with conditional data generation from the generative models. Fig. 3 illustrates the qualitative phase reconstruction achieved by our proposed conditional latent diffusion model (LDM). It is organized into three rows for clear comparison. In the top row, we present the condition images obtained through a ptychography algorithm running for a minimal number of iterations, resulting in very noisy photomask images. The middle row displays the samples generated by our LDM model, showcasing its ability to produce high-quality photomasks. The bottom row contains the ground truth photomasks for reference. It is noteworthy that despite the high level of noise in the conditioning images, our LDM model is capable of generating photomasks of good quality. This demonstrates the efficacy of the LDM in handling noisy input data and accurately reconstructing the desired phase information.

Since deep generative models like GANs, DDPMs, and LDMs require large amounts of data for training, we have developed a synthetic photomask generation pipeline for data agumentation. This pipeline can generate a large number of synthetic photomasks (with or without defects) within a few hours on an HPC cluster (see Fig. 2).

Impact

‍The impact of this work is significant in advancing semiconductor manufacturing technologies. Firstly, by implementing deep generative models for phase retrieval in ptychography, this approach has the potential to enhance the accuracy and resolution of defect detection on EUV photomasks. This improvement directly supports the industry's ability to enable the production of smaller, more powerful semiconductor devices. Secondly, the use of such advanced models reduces the computational load and associated costs compared to traditional methods, making the process more economically viable and potentially lowering the barrier to adoption for manufacturers. Finally, the application of machine learning and computer vision techniques, such as deep generative modeling and image-to-image translation, provides a novel pathway to overcome the challenges of high data volumes and dynamic range in lithography, thus improving the efficiency and speed of the imaging process.

‍