Knowledge of actual physics. Explosions can "easily" be simulated from first principles. Easily in scare quotes because it takes quite a bit of computing power. This was actually my wife's first job back in 2003, simulating missile strikes for the Naval Research Lab. A thorough simulation took a few days back then, but given that was almost 20 years ago, I'm sure it's a lot faster now.
In contrast, think of what you'd need to do this via machine learning. You'd need to gather data from actual missile strikes first and learn approximation functions from that. While it's certainly doable, this is inherently less accurate, thanks to both approximation error and measurement error. It's not like pixels -> cat where the true function isn't known.
Details are obviously scarce but computers are pretty much the number 1 reason weapons testing isn't particularly necessary any more.
That's true, but it really applies only if you have access to historical nuclear test data. The computer simulations are reportedly parameterized using empirical factors that aren't openly published and aren't available to additional nuclear aspirants unless they run tests of their own. The alternative, if you don't have test data and don't want to test, is to build very simple, conservative bomb designs like the US did with the Little Boy gun-type uranium bomb. It wasn't tested before it was dropped on Hiroshima. That's also the kind of design South Africa used for its secret nuclear weapons program:
Knowledge of actual physics. Explosions can "easily" be simulated from first principles. Easily in scare quotes because it takes quite a bit of computing power. This was actually my wife's first job back in 2003, simulating missile strikes for the Naval Research Lab. A thorough simulation took a few days back then, but given that was almost 20 years ago, I'm sure it's a lot faster now.
In contrast, think of what you'd need to do this via machine learning. You'd need to gather data from actual missile strikes first and learn approximation functions from that. While it's certainly doable, this is inherently less accurate, thanks to both approximation error and measurement error. It's not like pixels -> cat where the true function isn't known.
Some finite element analysis packages come with a label saying "Please pinky promise you won't use this for developing nuclear weapons".
https://wci.llnl.gov/simulation/computer-codes
https://www.worldscientific.com/doi/10.1142/9789812707130_00... (google the author)
https://www.lanl.gov/orgs/adtsc/publications/nw_highlights_2...
https://erdc-library.erdc.dren.mil/jspui/bitstream/11681/678...
Details are obviously scarce but computers are pretty much the number 1 reason weapons testing isn't particularly necessary any more.
Details are obviously scarce but computers are pretty much the number 1 reason weapons testing isn't particularly necessary any more.
That's true, but it really applies only if you have access to historical nuclear test data. The computer simulations are reportedly parameterized using empirical factors that aren't openly published and aren't available to additional nuclear aspirants unless they run tests of their own. The alternative, if you don't have test data and don't want to test, is to build very simple, conservative bomb designs like the US did with the Little Boy gun-type uranium bomb. It wasn't tested before it was dropped on Hiroshima. That's also the kind of design South Africa used for its secret nuclear weapons program:
https://en.wikipedia.org/wiki/South_Africa_and_weapons_of_ma...