One common approach is to look for the elbow in the curve <metric> vs K (number of clusters). This is essentially finding the number of clusters after which the rate of information gained/variance explained/<metric> slows. I believe it's possible to binary search for this point if you can assume the curve is convex.