January AMP_Digital

A D V A N C E D M A T E R I A L S & P R O C E S S E S | J A N U A R Y 2 0 1 9 2 0 quantified using a computer algorithm. For example, an algorithm can simulate a point count at every point on an im- age and apply logic to subclassify ma- terial phases, inclusions, and porosity, and generate relevant measurements, all in a matter of seconds. This method provides high speed, high accuracy, and correctable bias. Technological advancements in microscopes and cameras open the possibility to automate many industry standards for micrograph analysis. Fur- ther, advancements in computer algo- rithms and the introduction of machine learning to this field increase problem solving capabilities. Mipar is working with industry to automate microstruc- ture characterization following both in- ternal and industry standards. Because variability exists in sample preparation, image capture method, and sample characteristics, algorithms are tailored to the user to accommodate inherent variability, but calibration relies on in- dustry standards such as the ASTMA247 nodularity standard test method [11] . ASTM A247 is an example of how a computer automated algorithm can reduce error and highlight shortcom- ings of ambiguous wall chart analy- sis. Section 10 of the standard states, “Nodularity is expressed by count- ing the nodular particles and report- ing the results as a percentage of the total amount of graphite present in the microstructure.” Some might find it unclear whether the nodular graph- ite should be counted and reported as a count fraction, or if it should be point counted and reported as an area fraction. This ambiguity is resolved by cal- ibrating the computer algorithm to the standards chart at three points using both count fraction and area fraction (Fig. 4 and Table 3). Root-mean-square error analy- sis shows that area fraction interpre- tation results in lower overall error (7.22%) than the count fraction method (9.23%), as shown in Table 4. The error analysis also points to inherent error in labeling the standards chart. While re- viewing the area fraction results, chart micrographs at 20%, 30%, and 40% nodularity have the greatest absolute error from the algorithm. There are no obvious mistakes in the classification of the nodular graphite (Fig. 5), which highlights a subjective inaccuracy in the chart standard. The most likely source of error is a combination of mislabeled chart values (as it is unlikely that the values are exactly at 10% intervals) and the TABLE 3 – ALGORITHM CALIBRATION TO ASTM A247 NODULAR GRAPHITE STANDARDS CHART Chart Count fraction, % Area fraction, % 0 1.06 1.41 50 50.00 51.24 100 98.51 100.00 Root mean square error (RMSE), % 1.06 1.08 TABLE 4 – ALGORITHM ERROR ANALYSIS Chart Count fraction, % Area fraction, % 0 1.06 1.41 10 20.00 17.02 20 23.64 32.40 30 42.34 39.55 40 38.89 53.60 50 50.00 51.21 60 48.45 55.04 70 54.74 64.38 80 75.90 82.00 90 73.13 84.53 100 98.51 100.00 Root mean square error (RMSE), % 9.23 7.22 (a) (b) (c) Fig. 4 — Count fraction algorithm calibration results showing ambiguity in ASTM standard reference micrographs for estimating graphite nodularity in ductile iron (see Table 3): (a) 0% nodularity; (b) 50% nodularity; and (c) 100% nodularity (blue = non-nodular graphite and green = nodular graphite).