Enrichment of limited training sets in machine-learning-based analog/RF test

Abstract

This paper discusses the generation of informationrich, arbitrarily-large synthetic data sets which can be used to (a) efficiently learn tests that correlate a set of low-cost measurements to a set of device performances and (b) grade such tests with parts per million (PPM) accuracy. This is achieved by sampling a non-parametric estimate of the joint probability density function of measurements and performances. Our case study is an ultra-high frequency receiver front-end and the focus of the paper is to learn the mapping between a lowcost test measurement pattern and a single pass/fail test decision which reflects compliance to all performances. The small fraction of devices for which such a test decision is prone to error are identified and retested through standard specification-based test. The mapping can be set to explore thoroughly the tradeoff between test escapes, yield loss, and percentage of retested devices.

Integrated optimization of semiconductor manufacturing: A machine learning approach

“…To do this, we fit the joint probability density function of the instances from the first wafer. We sample the empirical probability density function to generate a much larger, information-rich training set that has a more balanced population of good, faulty, and critical devices across the decision boundary in a similar fashion to the approach taken in [16]. 7) Summary and Results: Assembling the preceding steps, we arrive at the complete analysis approach shown in Figure 7.…”

On proving the efficiency of alternative RF tests

“…We sample the empirical probability density function to generate an information-rich training set that has a more balanced population of good, faulty, and critical devices across the decision boundary in a similar fashion to the approach taken in [10].…”

“…As will be described in greater detail in section VI-E, the training phase employs an information-rich synthetic set of device instances that is generated through statistical simulation. This set comprises marginal instances whose footprints in the ORBiT space cover the areas around the true separation boundary [10] and, thereby, they allow a good approximation of this boundary.…”

Realistic Synthetic Data Generation: The ATEN Framework

“…For example, the heart rate should be a numerical value that falls within healthy resting (60 -100), exercising (100 -160) or disease state (40 -60 or 160+) ranges. Some projects require increasingly more complicated datasets where not only the values of single attributes must be valid, but all values and interrelationships must be indistinguishable from observed data [28,29]. This is where the problem of realism becomes imperative, yet it remains unexplored in current SDG literature [30].…”