You could just use data, machine learning and optimal control to make better estimates. It has aleady worked for fighter jets, self driving cars, and automated factories. With enough data it will work for anything.
What kind of features would we use? Which prior projects will help predict how long it takes a team of 3 developers (one of them senior with a history of underestimating time, two junior with little track history) to produce a novel photo sharing app where the requirements might or might not change every day?
The problem I see is that there isn't enough data, and the problem space (the number of possible projects and teams) is too large to draw meaningfully on past data.
Unless we're looking to predict specific CRUD apps where the requirements are known up front and a team cranks these out consistently.
Humans have already been flying fighter jets, driving cars, and doing work in factories, so we know it's a learnable skill. We have not yet learned how to estimate arbitrary software production.
