SBIR Phase I: Fuzzy CMAC for Real-Time Enhancement of Scanning Probe Microscope Images

SBIR第一阶段项目将利用一种新型神经网络来提高扫描探针显微镜（SPMs）捕获图像的准确性。扫描探针显微镜可以观察到原子尺度的材料，这种能力在许多行业中变得越来越重要。原子力显微镜成像误差的主要来源是由探针的几何形状引起的。目前的“探针校准”系统带来了一些好处，但还有更多的事情可以做。探针校准系统使用数学模型来计算基于已知样品的测量探针的几何形状。如果显微镜的输出显示了扫描样品时探针表面与已知标准样品之间的接触，那么就有可能根据测量样品的几何形状计算出探针的形状。然而，在实践中，这并不奏效。原子力显微镜测得的力不仅取决于探针表面与样品之间的相互作用。所涉及的所有力都非线性地依赖于探头的形状、大小、磁化强度、静电荷和其他特性。第一阶段的工作将验证基于模糊CMAC的原子力显微镜误差补偿方法的可行性。使用神经网络方法，特别是模糊CMAC的美妙之处在于，它不需要在标准样品的测量和探针几何模型之间建立封闭的形式关系。所需要做的就是扫描标准样本，提供足够的学习集来训练网络。同样，也不需要建立封闭的形式关系来利用计算误差模型来校正测量数据。模糊CMAC的输出将根据学习到的理想数据和实测数据之间的关系实时生成校正值。扫描探针显微镜的销量在全球范围内以每年约40%至60%的速度增长，并且它们正在越来越多的行业中使用。这项技术将提高这些机器的精度，并增加用户观察它们所生产的任何东西的原子结构的能力。该公司与一家世界领先的spm供应商签订了联合开发协议，从而为早期商业化提供了一条自然的道路。

This SBIR Phase 1 project will exploit a new type of neural net to improve the accuracy of images captured using scanning probe microscopes (SPMs). Scanning probe microscopes view materials down to atomic dimensions, and this ability is becoming increasingly important in many industries. A primary source of error in the images obtained from atomic force microscopes is caused by probe geometry. Current `probe calibration` systems yield some benefits but much more can be done. Probe calibration systems use mathematical models to compute the geometry of the probe based on measurements of known samples. If the microscope output indicated the contact between the surface of the probe and the known standard sample when a sample is scanned, then it would be possible to compute the probe shape from the measured sample geometry. In practice, however, this does not work well. The force sensed by the atomic force microscope does not depend only on the interaction between the surface of the probe and the sample. All of the forces involved depend non-linearly on the shape, size, magnetization, electrostatic charge, and on other characteristics of the probe. The Phase 1 effort will validate the feasibility of the fuzzy CMAC based error compensation approach for atomic force microscopes. The beauty of using a neural-net approach, and particularly a Fuzzy CMAC, is that it is not necessary to develop a closed form relationship between the measurements made on the standard sample and a model of the probe geometry. All that is necessary is to scan the standard sample providing enough learning sets to train the net. Likewise it is not necessary to develop closed form relations to exploit a computed error model to correct measured data. The output of the Fuzzy CMAC will generate the correction values in real-time based on the learned relationship between ideal and measured data. Scanning probe microscope sales are increasing worldwide at approximately 40 to 60% per year, and they are being used in an increasing number of industries. This technology will improve the accuracy of these machines and increase a user's ability to observe the atomic structure of whatever they are producing. This firm has a joint development agreement in place with one of the world's leading suppliers of SPMs, thereby providing a natural path to early commercialization.