First, it's highly unclear a priori which scientific discoveries will pay off. The discoverer of Green Fluorescent Protein was denied funding, with others eventually winning the Nobel Prize for it. Same for mRNA vaccines, most recently featured in COVID-19 vaccine, which also recently won a Nobel Prize.

https://en.wikipedia.org/wiki/Green_fluorescent_protein https://en.wikipedia.org/wiki/Katalin_Karik%C3%B3

Second, while there are always improvements to be made, the system as is (or was) worked pretty well in practice without knowing what the expected ROI was. The PageRank algorithm which led to Google was funded in part by an NSF grant on Digital Libraries. The ROI on that single invention just from taxes, jobs, and increased productivity likely exceeds NSF's annual budget. DataBricks and Duolingo are also based in part on NSF research.

Yeah, the system is imperfect, as all human-oriented systems are, but for the most part it works pretty well in practice and has been a linchpin in the US economic growth and national security.

If we're going to count the COVID-19 vaccine as a benefit of federal research funding, surely we need to also count COVID-19 itself as a cost, given the strong evidence that the virus was a product of US-funded gain of function research.