...yet at the same time biology and medicine will probably be the hardest fields for computers to "make discoveries in" for quite some time, except low level molecular biology things like low scale protein interaction or selecting from dug-candidate molecules. There's so much data so "badly formatted" and full of ambiguity that I can't even imagine when an AI will be able to read medical research papers and make sense of them - it's hard even for humans. Otoh there will be tons of data where computers will be able to search for patterns, but the actual knowledge necessary to transform "patterns" into "discoveries" is in the icky ambiguous format of papers and books that are impossible to make sense of for any less-than-human-level-AI.
>Running complex simulations can't truly provide new information, just validation.
This is entirely and absolutely false.
I think you may want to read up on simulation (FEM, molecular dynamics, etc.). It's my research field. For instance, Quantum Monte Carlo has been used to calculate energy levels to a ridiculous level of agreement with experimentally measured values. There's plenty that can be learned about reality that can't currently be studied through experimentation because it's either 1) too expensive or 2) impossible to study with current technology.
Also, I wouldn't say science is "about" experimentation. It's about prediction. Any theory that can more accurately predict a system's future state than another theory is the better science. And simulations can do this better in many cases than experiment.
What if the current model is wrong?
An experiment in the real/natural world is the only way to find out if a model (or its simulation) is wrong. You have to ask nature.
What I mean by this is that the whole nature of chemistry and biology is already rooted on a very secure foundation. So as simulation capabilities increase, you can be sure that these have a very high probability of modeling the "higher-order" phenomena correctly. Even superconductors and superfluids can be modeled correctly. The only things that may not be correct would be the high energy limit of quantum gravity models. But simulation in this area is almost (and may be?) nonexistent.
Now... having stated that we have an almost perfect description of the underlying reality for any practical phenomena we would be interested in, it is true that this model is currently mostly unamenable to calculation. The amount of computation required to reach chemical accuracy is very high and we may not have that capability for quite a while. So we make concessions and approximate things. For example, wavefunctions are modeled as Slater determinants which satisfy the wavefunction's antisymmetry but do not take into effect all the correlative effects (like electron-electron interactions).
It is a scientific and mathematical challenge to characterize exactly how much these models are off from the "true" solution. Most of the time, you have a good idea if the approximation you are using will work for your domain. If I want to fold proteins, I don't need anything more accurate than CHARMM (describes pairwise interactions; it's a non-reactive potential). But if I want to study Cooper Pairs (in superconductors) I'm going to need to make many less approximations, and it will be a much more expensive model computationally.
Want to simulate a cell? Even with hardware 10^100 times faster you can't use QM, it's all models of models. None of which are 100% accurate, but that's no reason you can't Kidd yourself.
Even at the timescales of atomic explosions and the best hardware / software around those models are vary simplified.
PS: That's not to say you can't make useful simulations, but trusting them to be accurate without testing is not science.
This is the reason for computational science. Imagine trying to render a photorealistic scene analytically. There's just no way you could do that. Exactly integrate the rendering equation over complex geometries? It's unfathomable. There's a reason the Schrodinger equation has only been solved analytically for systems of just a few particles (not to be confused with exactly, which is possible with numerical methods).
Eventually science experiments will be conducted by setting up your initial system, plugging in the relatively simple laws of physics, and letting it go (well... if we solve the n-body problem).
I don't think biology will ever be "understood" in the same way that derivatives or diffusion or gravity is understood. There's so many levels of nested complexity that it is totally mind-boggling, and I'm surprised more biologists aren't impressed by the sheer fact that any of it works at all.