Yield Test Structures Accelerate Learning

Laura Peters, Senior Editor -- Semiconductor International, 11/1/1999

Part 10 of Series

By employing large-area yield test structures early in the process development cycle, engineers rapidly identify the most significant yield detractors. Nick Atchison and Ron Ross of Silicon Systems (Santa Cruz, Calif.) developed new test structures they included in the reticle sets in the first few products using a 0.8 µm BICMOS process. The yield structures identified two main problems -- poly-poly leakage caused by poly stringers and metal 2-metal 2 leakage caused by TiW stringers. The corrected problems accounted for 80% of the total yield improvement attained from first silicon testing to full production. Optimum yield was reached in six months.

Though yield and parametric test structures occupy about 40% of the reticle field (Figure), their use results in rapid payback from accelerated yield improvement. The structures include serpentine patterns to test for opens, comb patterns to check for shorts and leakage, contact and via chains for all combinations, and arrays of transistors for each major transfer type. To distinguish between random and systematic defects, each structure should have three areas that can be tested individually -- a large-area structure (containing >20% of the product die area), medium structure (3-5X smaller) and small structure (3-5X smaller). Contact chains, for instance, might contain 100,000, 25,000 and 5,000 contacts. Structures indicating random yield problems have losses approximately proportional to the area. For example, if the area ratios are 16:4:1 and the smaller area structure fails 2% of the time, the medium structure should fail 8% of the time and the large area 32%. With systematic problems, the measured parameter value is proportional to the area. In measuring emitter-to-base current leakage with a 16:4:1 transistor ratio, the medium structure should have 4X the leakage of the smaller structure.

Using both design rule and sub-design-rule (0.8X or 0.9X) structures helps scale systematic problems to predict the effect on product yield. 'Worst-case topology' should also be reflected -- for example, designing metal 1 combs with underlying poly on thick oxide alternating with areas with no poly and no thick oxide. The engineers designed poly combs and serpentine structures with poly probe pads so they can perform probe testing after poly doping, patterning and activation for earlier results. The structures also must be positioned to represent a significant portion of wafer area.

Texas Instruments specifies that at least two reticle sets be run prior to the first product -- the 'devices' test chip with all transistor types, resistors, capacitors for characterization and modeling; and the 'circuits' chip containing optimized devices that may be modified after characterization. With both reticle sets containing yield test structures, initially 100% of wafer lots are run with structures. As yield problems are detected, analyzed and solved, the proportion of lots containing structures is reduced by 100%, 20%, 5% and then none.

Calculating product yield limits from test structure data requires calculation of the critical area for the critical mask layers (see SI, June 1999, p.66) and defect density for each type of defect (from electrical testing of yield test structures), by size (from visual inspection). Then the defect limited yield for each defect type is calculated using the negative binomial equation:

where a = cluster factor (' 2.5 for Santa Cruz) and

l_i = A ¦ P_i(X)D_i(X)dx

where A = die area, P_i(X) = probability of failure as a function of defect size (X) for defect type i and D_i(X) = defect density of defect type i as a function of size. The engineers generate the probability data using computer simulation of large numbers of defects randomly dispersed on each critical layer. Then they count the number of defects causing shorts, for instance, and calculate the probability of failure by simply dividing the number causing shorts by the total number distributed. Finally, l_i is inserted into the negative binomial equation to calculate the yield limit.