Оценка размера зерна по GB/T 6394: методы сравнения и пересечений
Метод сравнения быстр, но точность ±0,5 балла; метод пересечений даёт ±0,25 балла и обязателен при арбитраже.

Definition of the grade number
GB/T 6394 follows the same system as ASTM E112: the grain size number G relates logarithmically to the mean number of grains per square millimetre at 100x, based on the definition m=2^(G+3) grains per square millimetre at 1x. Higher G means finer grains. Annealed structural steels typically rate 4 to 8, fine-grained steels 8 to 12. The comparison method matches the field against standard rating charts at 100x; it is fast but carries an operator uncertainty of about ±0.5 grade.
Intercept method essentials
The intercept (Heyn) method is the arbitration method. Lay test lines totaling at least 500 mm, straight or circular, at random over the polished section, and count grain-boundary intersections P. The mean intercept is l̄=L_T/(P·M), where M is the magnification. Convert to grade by table or by G=(-6.64·lg l̄)-3.29 with l̄ in millimetres. The key statistical criterion: at least 400 intersections give a 95 percent confidence interval of about ±0.25 grade. Fields must be chosen at random; picking attractive areas is prohibited.
Planimetric method and mixed structures
The planimetric (Jeffries) method counts grains inside a known area, a 5000 mm² circle being standard, counting whole grains as 1 and boundary grains as one half; it suits coarse as-cast structures below G=0. A bimodal distribution cannot be described by one G value: report both grain populations with volume fractions, for example 70 percent grade 8 plus 30 percent grade 4, as explicitly required for mixed structures in GB/T 6394-2017.
Preparation and image analysis
Rating presupposes fully revealed boundaries: 4 percent nital for carbon steels; prior-austenite boundaries need overheating quench or carburizing treatments. Image analysis identifies boundaries automatically and computes intercepts five times faster than manual counting, but an operator must verify a boundary recognition rate of at least 90 percent, and twin boundaries must be excluded in the software settings since they do not count.



